From aea2aa3fb5e63ccd9614c24a465f12fde4563d43 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sun, 23 Mar 2025 15:55:16 +0700 Subject: [PATCH 01/27] GTAO depth prefilter pipeline. --- CMakeLists.txt | 3 + assets/maps/sampleScene.willmap | 47 +++++-- assets/models/sponza2/Sponza.willmodel | 8 ++ .../ambient_occlusion/ground_truth/gtao.comp | 92 ++++++++++++++ .../ground_truth/gtaodepthprefilter.comp | 5 + shaders/deferredResolve.comp | 5 +- src/core/engine.cpp | 1 - src/renderer/environment/environment.cpp | 8 +- src/renderer/imgui_wrapper.cpp | 2 +- .../ambient_occlusion_types.h | 34 +++++ .../ground_truth_ambient_occlusion.cpp | 119 ++++++++++++++++++ .../ground_truth_ambient_occlusion.h | 56 +++++++++ .../lighting/shadows/cascaded_shadow_map.cpp | 10 +- src/renderer/resource_manager.cpp | 29 ++++- src/renderer/resource_manager.h | 2 + 15 files changed, 405 insertions(+), 16 deletions(-) create mode 100644 assets/models/sponza2/Sponza.willmodel create mode 100644 shaders/ambient_occlusion/ground_truth/gtao.comp create mode 100644 shaders/ambient_occlusion/ground_truth/gtaodepthprefilter.comp create mode 100644 src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h create mode 100644 src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp create mode 100644 src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h diff --git a/CMakeLists.txt b/CMakeLists.txt index 17e4f71e..01c37f27 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -247,6 +247,9 @@ set(TEMP_SOURCES src/renderer/assets/texture/texture_resource.cpp src/renderer/assets/texture/texture_resource.h src/renderer/assets/texture/texture_types.h + src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp + src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h + src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h ) diff --git a/assets/maps/sampleScene.willmap b/assets/maps/sampleScene.willmap index 6995a167..7cf48529 100644 --- a/assets/maps/sampleScene.willmap +++ b/assets/maps/sampleScene.willmap @@ -6,7 +6,7 @@ }, "metadata": { "name": "sampleScene", - "created": "2025-03-21 18:18:15", + "created": "2025-03-22 17:23:20", "formatVersion": 1 }, "rootComponents": { @@ -280,15 +280,15 @@ "name": "Cube\u0000ing Orb", "transform": { "position": { - "x": 13.52876091003418, - "y": 21.364248275756836, - "z": 37.19429397583008 + "x": 13.52898120880127, + "y": 21.364240646362305, + "z": 37.19432067871094 }, "rotation": { - "x": 0.19281113147735596, - "y": 0.20267827808856964, - "z": 0.6591735482215881, - "w": 0.6980226635932922 + "x": 0.19284461438655853, + "y": 0.2026415914297104, + "z": 0.6591105461120605, + "w": 0.6980835795402527 }, "scale": { "x": 1.0, @@ -319,6 +319,37 @@ "componentName": "New RigidBodyComponent" } } + }, + { + "id": 4, + "name": "Sponza\u0000ect_4", + "transform": { + "position": { + "x": 102.19999694824219, + "y": 77.0, + "z": -49.099998474121094 + }, + "rotation": { + "x": 5.711166295441217e-08, + "y": -0.3826834559440613, + "z": 5.711166295441217e-08, + "w": 0.9238795042037964 + }, + "scale": { + "x": 0.05000000074505806, + "y": 0.05000000074505806, + "z": 0.05000000074505806 + } + }, + "components": { + "MeshRendererComponent": { + "renderReference": 195023067, + "renderMeshIndex": 0, + "renderIsVisible": true, + "renderIsShadowCaster": true, + "componentName": "MeshRendererComponent" + } + } } ] } diff --git a/assets/models/sponza2/Sponza.willmodel b/assets/models/sponza2/Sponza.willmodel new file mode 100644 index 00000000..d945e982 --- /dev/null +++ b/assets/models/sponza2/Sponza.willmodel @@ -0,0 +1,8 @@ +{ + "renderObject": { + "gltfPath": "assets\\models\\sponza2\\Sponza.gltf", + "id": 195023067, + "name": "Sponza" + }, + "version": 1 +} diff --git a/shaders/ambient_occlusion/ground_truth/gtao.comp b/shaders/ambient_occlusion/ground_truth/gtao.comp new file mode 100644 index 00000000..c48a172d --- /dev/null +++ b/shaders/ambient_occlusion/ground_truth/gtao.comp @@ -0,0 +1,92 @@ +#version 460 + +void main() { + +} + +void XeGTAO_PrefilterDepths16x16( +uvec2 dispatchThreadID, +uvec2 groupThreadID, +const GTAOConstants consts, +sampler2D sourceNDCDepth, +out writeonly image2D outDepth0, +out writeonly image2D outDepth1, +out writeonly image2D outDepth2, +out writeonly image2D outDepth3, +out writeonly image2D outDepth4 +) { + // MIP 0 + const uvec2 baseCoord = dispatchThreadID; + // 2x because ao image is downsampled + const uvec2 pixCoord = baseCoord * 2u; + + + // todo: get width and height from sceneData + vec2 uvCoord = vec2(pixCoord) * consts.ViewportPixelSize; + + vec4 depths 4; + depths4.w = texture(sourceNDCDepth, uvCoord + vec2(0.0, 0.0) * consts.ViewportPixelSize).r; + depths4.z = texture(sourceNDCDepth, uvCoord + vec2(1.0, 0.0) * consts.ViewportPixelSize).r; + depths4.x = texture(sourceNDCDepth, uvCoord + vec2(0.0, 1.0) * consts.ViewportPixelSize).r; + depths4.y = texture(sourceNDCDepth, uvCoord + vec2(1.0, 1.0) * consts.ViewportPixelSize).r; + + float depth0 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.w, consts)); + float depth1 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.z, consts)); + float depth2 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.x, consts)); + float depth3 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.y, consts)); + + imageStore(outDepth0, ivec2(pixCoord + uvec2(0, 0)), vec4(depth0, 0.0, 0.0, 0.0)); + imageStore(outDepth0, ivec2(pixCoord + uvec2(1, 0)), vec4(depth1, 0.0, 0.0, 0.0)); + imageStore(outDepth0, ivec2(pixCoord + uvec2(0, 1)), vec4(depth2, 0.0, 0.0, 0.0)); + imageStore(outDepth0, ivec2(pixCoord + uvec2(1, 1)), vec4(depth3, 0.0, 0.0, 0.0)); + + // MIP 1 + float dm1 = XeGTAO_DepthMIPFilter(depth0, depth1, depth2, depth3, consts); + imageStore(outDepth1, ivec2(baseCoord), vec4(dm1, 0.0, 0.0, 0.0)); + g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; + + memoryBarrierShared(); + barrier(); + + // MIP 2 + if (all(equal(groupThreadID.xy % 2u, uvec2(0u, 0u)))) { + float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; + float inTR = g_scratchDepths[groupThreadID.x+1u][groupThreadID.y+0u]; + float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+1u]; + float inBR = g_scratchDepths[groupThreadID.x+1u][groupThreadID.y+1u]; + + float dm2 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); + imageStore(outDepth2, ivec2(baseCoord / 2u), vec4(dm2, 0.0, 0.0, 0.0)); + g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm2; + } + + memoryBarrierShared(); + barrier(); + + // MIP 3 + if (all(equal(groupThreadID.xy % 4u, uvec2(0u, 0u)))) { + float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; + float inTR = g_scratchDepths[groupThreadID.x+2u][groupThreadID.y+0u]; + float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+2u]; + float inBR = g_scratchDepths[groupThreadID.x+2u][groupThreadID.y+2u]; + + float dm3 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); + imageStore(outDepth3, ivec2(baseCoord / 4u), vec4(dm3, 0.0, 0.0, 0.0)); + g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm3; + } + + memoryBarrierShared(); + barrier(); + + // MIP 4 + if (all(equal(groupThreadID.xy % 8u, uvec2(0u, 0u)))) { + float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; + float inTR = g_scratchDepths[groupThreadID.x+4u][groupThreadID.y+0u]; + float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+4u]; + float inBR = g_scratchDepths[groupThreadID.x+4u][groupThreadID.y+4u]; + + float dm4 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); + imageStore(outDepth4, ivec2(baseCoord / 8u), vec4(dm4, 0.0, 0.0, 0.0)); + // g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; // commented out as in original + } +} \ No newline at end of file diff --git a/shaders/ambient_occlusion/ground_truth/gtaodepthprefilter.comp b/shaders/ambient_occlusion/ground_truth/gtaodepthprefilter.comp new file mode 100644 index 00000000..747d43d5 --- /dev/null +++ b/shaders/ambient_occlusion/ground_truth/gtaodepthprefilter.comp @@ -0,0 +1,5 @@ +#version 460 + +void main() { + +} diff --git a/shaders/deferredResolve.comp b/shaders/deferredResolve.comp index 6b5837aa..811c53ed 100644 --- a/shaders/deferredResolve.comp +++ b/shaders/deferredResolve.comp @@ -130,7 +130,7 @@ void main() { shadowFactor = 1.0f; } - float indirectAttenuation = mix(0.2, 1.0, shadowFactor); + float indirectAttenuation = mix(0.5, 1.0, shadowFactor); // IBL REFLECTIONS vec3 irradiance = DiffuseIrradiance(environmentDiffuseAndSpecular, N); @@ -186,5 +186,8 @@ void main() { case 8: imageStore(outputImage, screenPos, vec4(vec3(dot(N, L)), 1.0)); break; + case 9: + // imageStore(outputImage, screenPos, vec4(vec3(ao), 1.0f)); + break; } } \ No newline at end of file diff --git a/src/core/engine.cpp b/src/core/engine.cpp index 5a8d7cbb..4100a18b 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -490,7 +490,6 @@ void Engine::draw(float deltaTime) environmentMap->getCubemapDescriptorBuffer().getDescriptorBufferBindingInfo(), environmentMap->getCubemapDescriptorBuffer().getDescriptorBufferSize() * environmentMapIndex, }; - // todo: make environment pipeline draw to the render targets rather than directly to the draw image!!! environmentPipeline->draw(cmd, environmentPipelineDrawInfo); diff --git a/src/renderer/environment/environment.cpp b/src/renderer/environment/environment.cpp index c0345fb9..8942f4f8 100644 --- a/src/renderer/environment/environment.cpp +++ b/src/renderer/environment/environment.cpp @@ -215,7 +215,13 @@ will_engine::environment::Environment::Environment(ResourceManager& resourceMana { lutDescriptorBuffer = resourceManager.createDescriptorBufferSampler(lutLayout, 1); - lutImage = resourceManager.createImage(LUT_IMAGE_EXTENT, VK_FORMAT_R32G32_SFLOAT, VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_STORAGE_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT); + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; + + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(VK_FORMAT_R32G32_SFLOAT, usage, LUT_IMAGE_EXTENT); + lutImage = resourceManager.createImage(imgInfo); VkDescriptorImageInfo lutDescriptorInfo{}; lutDescriptorInfo.sampler = nullptr; // not sampled (storage) diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index 5d9376e6..e70f3f3a 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -642,7 +642,7 @@ void ImguiWrapper::imguiInterface(Engine* engine) gltfPath = IGFD::FileDialog::Instance()->GetFilePathName(); gltfPath = file::getRelativePath(gltfPath); - willmodelPath = std::filesystem::current_path() / "assets" / "willmodels" / gltfPath.filename().string(); + willmodelPath = gltfPath.parent_path() / gltfPath.filename().string(); willmodelPath = file::getRelativePath(willmodelPath); willmodelPath.replace_extension(".willmodel"); } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h new file mode 100644 index 00000000..a65394df --- /dev/null +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -0,0 +1,34 @@ +// +// Created by William on 2025-03-23. +// + +#ifndef AMBIENT_OCCLUSION_TYPES_H +#define AMBIENT_OCCLUSION_TYPES_H +#include + +namespace will_engine::ambient_occlusion +{ +static constexpr int32_t DEPTH_PREFILTER_MIP_COUNT = 5; + +struct GTAOPushConstants +{ + glm::vec2 viewportSize; + glm::vec2 viewportPixelSize; + + // AO parameters + float radius; + float falloff; + float strength; + float radiusMultiplier; + + // Sampling parameters + uint32_t numDirections; + uint32_t numSteps; + + // Temporal/filter parameters + float temporalWeight; + float spatialFilterRadius; +}; +} + +#endif //AMBIENT_OCCLUSION_TYPES_H diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp new file mode 100644 index 00000000..580d1a91 --- /dev/null +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -0,0 +1,119 @@ +// +// Created by William on 2025-03-23. +// + +#include "ground_truth_ambient_occlusion.h" + +#include "src/renderer/renderer_constants.h" +#include "src/renderer/vk_descriptors.h" +#include "src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h" + +will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruthAmbientOcclusionPipeline(ResourceManager& resourceManager) : resourceManager(resourceManager) +{ + // Depth Pre-filtering + { + DescriptorLayoutBuilder layoutBuilder; + layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // depth image + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 0 + layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 1 + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 2 + layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 3 + layoutBuilder.addBinding(5, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 4 + + depthPrefilterSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + + VkPushConstantRange pushConstants{}; + pushConstants.offset = 0; + pushConstants.size = sizeof(GTAOPushConstants); + pushConstants.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT; + + VkDescriptorSetLayout setLayouts[2]; + setLayouts[0] = resourceManager.getSceneDataLayout(); + setLayouts[1] = depthPrefilterSetLayout; + + VkPipelineLayoutCreateInfo layoutInfo{}; + layoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO; + layoutInfo.pNext = nullptr; + layoutInfo.pSetLayouts = setLayouts; + layoutInfo.setLayoutCount = 2; + layoutInfo.pPushConstantRanges = &pushConstants; + layoutInfo.pushConstantRangeCount = 1; + + depthPrefilterPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); + + createDepthPrefilterPipeline(); + + depthPrefilterDescriptorBuffer = resourceManager.createDescriptorBufferSampler(depthPrefilterSetLayout, 1); + + + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(depthPrefilterFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + + // 5 mips, suggested by Intel's implementation + // https://github.com/GameTechDev/XeGTAO + imgInfo.mipLevels = DEPTH_PREFILTER_MIP_COUNT; + + depthPrefilterImage = resourceManager.createImage(imgInfo); + + VkImageViewCreateInfo viewInfo = vk_helpers::imageviewCreateInfo(depthPrefilterFormat, depthPrefilterImage.image, VK_IMAGE_ASPECT_COLOR_BIT); + + for (int32_t i = 0; i < DEPTH_PREFILTER_MIP_COUNT; ++i) { + viewInfo.subresourceRange.baseMipLevel = i; + depthPrefilterImageViews[i] = resourceManager.createImageView(viewInfo); + } + } +} + +will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTruthAmbientOcclusionPipeline() +{} + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(VkSampler depthImageSampler, VkImageView depthImageView) +{ + std::vector imageDescriptors{}; + imageDescriptors.reserve(1 + DEPTH_PREFILTER_MIP_COUNT); + + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + {depthImageSampler, depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + false + } + ); + + for (int32_t i = 0; i < DEPTH_PREFILTER_MIP_COUNT; ++i) { + DescriptorImageData imageData{}; + imageData.type = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE; + imageData.imageInfo = {VK_NULL_HANDLE, depthPrefilterImageViews[i], VK_IMAGE_LAYOUT_GENERAL}; + imageData.bIsPadding = false; + + imageDescriptors.push_back(imageData); + } + + resourceManager.setupDescriptorBufferSampler(depthPrefilterDescriptorBuffer, imageDescriptors, 0); +} + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createDepthPrefilterPipeline() +{ + resourceManager.destroyPipeline(depthPrefilterPipeline); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/gtaodepthprefilter.comp"); + + VkPipelineShaderStageCreateInfo stageInfo{}; + stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; + stageInfo.pNext = nullptr; + stageInfo.stage = VK_SHADER_STAGE_COMPUTE_BIT; + stageInfo.module = computeShader; + stageInfo.pName = "main"; + + VkComputePipelineCreateInfo pipelineInfo{}; + pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO; + pipelineInfo.pNext = nullptr; + pipelineInfo.layout = depthPrefilterPipelineLayout; + pipelineInfo.stage = stageInfo; + pipelineInfo.flags = VK_PIPELINE_CREATE_DESCRIPTOR_BUFFER_BIT_EXT; + + depthPrefilterPipeline = resourceManager.createComputePipeline(pipelineInfo); + resourceManager.destroyShaderModule(computeShader); +} diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h new file mode 100644 index 00000000..25a78d3b --- /dev/null +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -0,0 +1,56 @@ +// +// Created by William on 2025-03-23. +// + +#ifndef GROUND_TRUTH_AMBIENT_OCCLUSION_H +#define GROUND_TRUTH_AMBIENT_OCCLUSION_H + +#include +#include + +#include "src/renderer/resource_manager.h" +#include "src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h" + + +namespace will_engine::ambient_occlusion +{ +class GroundTruthAmbientOcclusionPipeline +{ +public: + GroundTruthAmbientOcclusionPipeline(ResourceManager& resourceManager); + + ~GroundTruthAmbientOcclusionPipeline(); + + void setupDepthPrefilterDescriptorBuffer(VkSampler depthImageSampler, VkImageView depthImageView); + +private: + void createDepthPrefilterPipeline(); + +private: + VkDescriptorSetLayout depthPrefilterSetLayout{VK_NULL_HANDLE}; + VkPipelineLayout depthPrefilterPipelineLayout{VK_NULL_HANDLE}; + VkPipeline depthPrefilterPipeline{VK_NULL_HANDLE}; + + // 16 vs 32. look at cost later. + VkFormat depthPrefilterFormat{VK_FORMAT_R16_SFLOAT}; + AllocatedImage depthPrefilterImage{VK_NULL_HANDLE}; + std::array depthPrefilterImageViews{}; + + DescriptorBufferSampler depthPrefilterDescriptorBuffer; + + VkPipelineLayout ambientOcclusionPipelineLayout{VK_NULL_HANDLE}; + VkPipeline ambientOcclusionPipeline{VK_NULL_HANDLE}; + + VkPipelineLayout spatialFilteringPipelineLayout{VK_NULL_HANDLE}; + VkPipeline spatialFilteringPipeline{VK_NULL_HANDLE}; + + VkPipelineLayout temporalAccumulationPipelineLayout{VK_NULL_HANDLE}; + VkPipeline temporalAccumulationPipeline{VK_NULL_HANDLE}; + +private: + ResourceManager& resourceManager; +}; +} + + +#endif //GROUND_TRUTH_AMBIENT_OCCLUSION_H diff --git a/src/renderer/lighting/shadows/cascaded_shadow_map.cpp b/src/renderer/lighting/shadows/cascaded_shadow_map.cpp index 0627dd41..3bd9b091 100644 --- a/src/renderer/lighting/shadows/cascaded_shadow_map.cpp +++ b/src/renderer/lighting/shadows/cascaded_shadow_map.cpp @@ -159,8 +159,14 @@ will_engine::cascaded_shadows::CascadedShadowMap::CascadedShadowMap(ResourceMana sampler = resourceManager.createSampler(samplerCreateInfo); for (CascadeShadowMapData& cascadeShadowMapData : shadowMaps) { - cascadeShadowMapData.depthShadowMap = resourceManager.createImage({shadows::CASCADE_WIDTH, shadows::CASCADE_HEIGHT, 1}, shadows::CASCADE_DEPTH_FORMAT, - VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT | VK_IMAGE_USAGE_TRANSFER_SRC_BIT); + + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; + + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(shadows::CASCADE_DEPTH_FORMAT, usage, {shadows::CASCADE_WIDTH, shadows::CASCADE_HEIGHT, 1}); + cascadeShadowMapData.depthShadowMap = resourceManager.createImage(imgInfo); } // diff --git a/src/renderer/resource_manager.cpp b/src/renderer/resource_manager.cpp index 3943a1e0..7767e8dc 100644 --- a/src/renderer/resource_manager.cpp +++ b/src/renderer/resource_manager.cpp @@ -308,6 +308,31 @@ VkSampler will_engine::ResourceManager::createSampler(const VkSamplerCreateInfo& return newSampler; } + +AllocatedImage will_engine::ResourceManager::createImage(const VkImageCreateInfo& createInfo) const +{ + AllocatedImage newImage{}; + newImage.imageFormat = createInfo.format; + newImage.imageExtent = createInfo.extent; + + VmaAllocationCreateInfo allocInfo = {}; + allocInfo.usage = VMA_MEMORY_USAGE_GPU_ONLY; + allocInfo.requiredFlags = static_cast(VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT); + + // allocate and create the image + VK_CHECK(vmaCreateImage(context.allocator, &createInfo, &allocInfo, &newImage.image, &newImage.allocation, nullptr)); + + const VkImageAspectFlags aspectFlag = createInfo.format == VK_FORMAT_D32_SFLOAT ? VK_IMAGE_ASPECT_DEPTH_BIT : VK_IMAGE_ASPECT_COLOR_BIT; + + // build an image-view for the image + VkImageViewCreateInfo view_info = vk_helpers::imageviewCreateInfo(createInfo.format, newImage.image, aspectFlag); + view_info.subresourceRange.levelCount = createInfo.mipLevels; + + VK_CHECK(vkCreateImageView(context.device, &view_info, nullptr, &newImage.imageView)); + + return newImage; +} + AllocatedImage will_engine::ResourceManager::createImage(const VkExtent3D size, const VkFormat format, const VkImageUsageFlags usage, const bool mipmapped) const { AllocatedImage newImage{}; @@ -416,7 +441,7 @@ will_engine::DescriptorBufferSampler will_engine::ResourceManager::createDescrip return DescriptorBufferSampler(context, layout, maxObjectCount); } -int32_t will_engine::ResourceManager::setupDescriptorBufferSampler(DescriptorBufferSampler& descriptorBuffer, const std::vector& imageBuffers, const int index) const +int32_t will_engine::ResourceManager::setupDescriptorBufferSampler(DescriptorBufferSampler& descriptorBuffer, const std::vector& imageBuffers, const int index) const { return descriptorBuffer.setupData(context.device, imageBuffers, index); } @@ -426,7 +451,7 @@ will_engine::DescriptorBufferUniform will_engine::ResourceManager::createDescrip return DescriptorBufferUniform(context, layout, maxObjectCount); } -int32_t will_engine::ResourceManager::setupDescriptorBufferUniform(DescriptorBufferUniform& descriptorBuffer, const std::vector& uniformBuffers, const int index) const +int32_t will_engine::ResourceManager::setupDescriptorBufferUniform(DescriptorBufferUniform& descriptorBuffer, const std::vector& uniformBuffers, const int index) const { return descriptorBuffer.setupData(context.device, uniformBuffers, index); } diff --git a/src/renderer/resource_manager.h b/src/renderer/resource_manager.h index 705011c0..d26134af 100644 --- a/src/renderer/resource_manager.h +++ b/src/renderer/resource_manager.h @@ -58,6 +58,8 @@ class ResourceManager void destroySampler(const VkSampler& sampler) const; + [[nodiscard]] AllocatedImage createImage(const VkImageCreateInfo& createInfo) const; + [[nodiscard]] AllocatedImage createImage(VkExtent3D size, VkFormat format, VkImageUsageFlags usage, bool mipmapped = false) const; [[nodiscard]] AllocatedImage createImage(const void* data, size_t dataSize, VkExtent3D size, VkFormat format, VkImageUsageFlags usage, bool mipmapped = false) const; From 103eb9c012fb5c18f70c42141beed8cfe2b301f5 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sun, 23 Mar 2025 16:15:47 +0700 Subject: [PATCH 02/27] Improved scene data buffer. Minor improvements --- assets/maps/sampleScene.willmap | 28 +++++++++---------- shaders/include/scene.glsl | 2 ++ shaders/taa.comp | 23 +++++++-------- src/core/engine.cpp | 1 + src/renderer/imgui_wrapper.cpp | 20 ++++++------- .../temporal_antialiasing_pipeline.cpp | 3 -- .../temporal_antialiasing_pipeline.h | 3 -- src/renderer/resource_manager.cpp | 7 ++++- src/renderer/resource_manager.h | 2 +- src/renderer/vk_types.h | 1 + 10 files changed, 44 insertions(+), 46 deletions(-) diff --git a/assets/maps/sampleScene.willmap b/assets/maps/sampleScene.willmap index 7cf48529..cb7e5575 100644 --- a/assets/maps/sampleScene.willmap +++ b/assets/maps/sampleScene.willmap @@ -6,7 +6,7 @@ }, "metadata": { "name": "sampleScene", - "created": "2025-03-22 17:23:20", + "created": "2025-03-23 16:02:52", "formatVersion": 1 }, "rootComponents": { @@ -34,9 +34,9 @@ "y": 50.0 }, "baseColor": { - "x": 0.5343137383460999, - "y": 0.5343137383460999, - "z": 0.5343137383460999, + "x": 0.8725489974021912, + "y": 0.8725489974021912, + "z": 0.8725489974021912, "w": 1.0 } }, @@ -109,9 +109,9 @@ "y": 50.0 }, "baseColor": { - "x": 0.5343137383460999, - "y": 0.5343137383460999, - "z": 0.5343137383460999, + "x": 0.8725489974021912, + "y": 0.8725489974021912, + "z": 0.8725489974021912, "w": 1.0 } }, @@ -280,15 +280,15 @@ "name": "Cube\u0000ing Orb", "transform": { "position": { - "x": 13.52898120880127, - "y": 21.364240646362305, - "z": 37.19432067871094 + "x": 13.530956268310547, + "y": 21.363996505737305, + "z": 37.1947135925293 }, "rotation": { - "x": 0.19284461438655853, - "y": 0.2026415914297104, - "z": 0.6591105461120605, - "w": 0.6980835795402527 + "x": 0.19317233562469482, + "y": 0.2023322433233261, + "z": 0.6584249138832092, + "w": 0.698729395866394 }, "scale": { "x": 1.0, diff --git a/shaders/include/scene.glsl b/shaders/include/scene.glsl index f511e143..9da8db70 100644 --- a/shaders/include/scene.glsl +++ b/shaders/include/scene.glsl @@ -25,5 +25,7 @@ layout (std140, set = 0, binding = 0) uniform SceneData { vec4 jitter; vec2 renderTargetSize; + // equal to 1 / renderTargetSize + vec2 texelSize; float deltaTime; } sceneData; \ No newline at end of file diff --git a/shaders/taa.comp b/shaders/taa.comp index ea5c6b25..f0027bdf 100644 --- a/shaders/taa.comp +++ b/shaders/taa.comp @@ -15,9 +15,6 @@ layout (set = 1, binding = 3) uniform sampler2D velocityBuffer; layout (rgba16f, set = 1, binding = 4) uniform image2D outputImage; layout (push_constant) uniform PushConstants { - vec2 texelSize; - int width; - int height; float blendValue; int debug; } push; @@ -42,8 +39,8 @@ vec3 clipAABB(vec3 cMin, vec3 cMax, vec3 cAvg, vec3 color) { vec3 findClosestFragment3x3(vec2 uv) { - vec2 du = vec2(push.texelSize.x); - vec2 dv = vec2(push.texelSize.y); + vec2 du = vec2(sceneData.texelSize.x); + vec2 dv = vec2(sceneData.texelSize.y); vec3 topL = vec3(-1, -1, texture(depthImage, uv - dv - du).x); vec3 topM = vec3(0, -1, texture(depthImage, uv - dv).x); @@ -68,15 +65,15 @@ vec3 findClosestFragment3x3(vec2 uv) if (botM.z > dMin.z) dMin = botM; if (botR.z > dMin.z) dMin = botR; - return vec3(uv + push.texelSize * dMin.xy, dMin.z); + return vec3(uv + sceneData.texelSize * dMin.xy, dMin.z); } vec3 temporalReprojection(vec2 uv, vec2 velocity) { vec3 currentSample = RGBToYCoCg(texture(drawImage, uv).rgb); vec3 historySample = RGBToYCoCg(texture(drawHistory, uv - velocity).rgb); - vec2 du = vec2(push.texelSize.x); - vec2 dv = vec2(push.texelSize.y); + vec2 du = vec2(sceneData.texelSize.x); + vec2 dv = vec2(sceneData.texelSize.y); // MINMAX_3X3_ROUNDED vec3 topL = RGBToYCoCg(texture(drawImage, uv - dv - du).rgb); @@ -138,7 +135,7 @@ vec3 varianceClipping(vec3 color, vec2 uv) { for (int y = -1; y <= 1; y++) { for (int x = -1; x <= 1; x++) { - vec2 offset = vec2(x, y) * push.texelSize; + vec2 offset = vec2(x, y) * sceneData.texelSize; vec3 sampleRGB = texture(drawImage, uv + offset).rgb; vec3 sampleYCoCg = RGBToYCoCg(sampleRGB); @@ -166,8 +163,8 @@ vec3 varianceClipping(vec3 color, vec2 uv) { } vec3 neighborhoodClamping(vec3 historyColor, vec2 uv) { - vec2 du = vec2(push.texelSize.x); - vec2 dv = vec2(push.texelSize.y); + vec2 du = vec2(sceneData.texelSize.x); + vec2 dv = vec2(sceneData.texelSize.y); // Sample current frame neighborhood vec3 topL = texture(drawImage, uv - dv - du).rgb; @@ -192,10 +189,10 @@ vec3 neighborhoodClamping(vec3 historyColor, vec2 uv) { void main() { ivec2 pixel = ivec2(gl_GlobalInvocationID.xy); - if (pixel.x >= push.width || pixel.y >= push.height) { + if (pixel.x >= sceneData.renderTargetSize.x || pixel.y >= sceneData.renderTargetSize.y) { return; } - vec2 uv = (vec2(pixel) + 0.5) * push.texelSize; + vec2 uv = (vec2(pixel) + 0.5) * sceneData.texelSize; if (push.debug == 1) { // disabled diff --git a/src/core/engine.cpp b/src/core/engine.cpp index 4100a18b..de85f276 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -389,6 +389,7 @@ void Engine::updateRender(const float deltaTime, const int32_t currentFrameOverl pSceneData->renderTargetSize = {RENDER_EXTENT_WIDTH, RENDER_EXTENT_HEIGHT}; + pSceneData->texelSize = {1.0f / RENDER_EXTENT_WIDTH, 1.0f / RENDER_EXTENT_HEIGHT}; pSceneData->deltaTime = deltaTime; diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index e70f3f3a..4178d9f1 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -933,19 +933,17 @@ void ImguiWrapper::drawSceneGraph(Engine* engine) if (ImGui::BeginTabItem("Terrain")) { ImGui::Checkbox("Draw Vertex Lines Only", &engine->bDrawTerrainLines); - if (ImGui::BeginTabBar("Terrain Tab Bar")) { - auto currentTerrainComponent = selectedMap->getComponent(); + const auto currentTerrainComponent = selectedMap->getComponent(); + ImGui::BeginDisabled(!currentTerrainComponent); + if (ImGui::Button("Save Terrain as HeightMap")) { + const std::vector heightmapData = currentTerrainComponent->getHeightMapData(); + const std::filesystem::path path = file::imagesSavePath / "TerrainHeightMap.png"; + vk_helpers::saveHeightmap(heightmapData, NOISE_MAP_DIMENSIONS, NOISE_MAP_DIMENSIONS, path); + } + ImGui::EndDisabled(); + if (ImGui::BeginTabBar("Terrain Tab Bar")) { if (ImGui::BeginTabItem("Terrain Generation")) { - ImGui::BeginDisabled(!currentTerrainComponent); - if (ImGui::Button("Save Terrain as HeightMap")) { - std::vector heightmapData = currentTerrainComponent->getHeightMapData(); - std::filesystem::path path = file::imagesSavePath / "TerrainHeightMap.png"; - vk_helpers::saveHeightmap(heightmapData, NOISE_MAP_DIMENSIONS, NOISE_MAP_DIMENSIONS, path); - } - ImGui::EndDisabled(); - - ImGui::Separator(); if (ImGui::CollapsingHeader("Noise Settings", ImGuiTreeNodeFlags_DefaultOpen)) { diff --git a/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.cpp b/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.cpp index e682da01..3b188961 100644 --- a/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.cpp +++ b/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.cpp @@ -105,9 +105,6 @@ void will_engine::temporal_antialiasing_pipeline::TemporalAntialiasingPipeline:: vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipeline); TemporalAntialiasingPushConstants properties{};; - properties.width = RENDER_EXTENT_WIDTH; - properties.height = RENDER_EXTENT_HEIGHT; - properties.texelSize = {1.0f / properties.width, 1.0f / properties.height}; properties.blendValue = drawInfo.blendValue; properties.taaDebug = drawInfo.debugMode; diff --git a/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.h b/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.h index 0f8a7400..e3541bbc 100644 --- a/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.h +++ b/src/renderer/pipelines/temporal_antialiasing_pipeline/temporal_antialiasing_pipeline.h @@ -11,9 +11,6 @@ namespace will_engine::temporal_antialiasing_pipeline { struct TemporalAntialiasingPushConstants { - glm::vec2 texelSize; - int32_t width; - int32_t height; float blendValue; int32_t taaDebug; }; diff --git a/src/renderer/resource_manager.cpp b/src/renderer/resource_manager.cpp index 7767e8dc..14f2b7a5 100644 --- a/src/renderer/resource_manager.cpp +++ b/src/renderer/resource_manager.cpp @@ -425,10 +425,15 @@ AllocatedImage will_engine::ResourceManager::createCubemap(const VkExtent3D size return newImage; } -void will_engine::ResourceManager::destroyImage(const AllocatedImage& img) const +void will_engine::ResourceManager::destroyImage(AllocatedImage& img) const { vkDestroyImageView(context.device, img.imageView, nullptr); vmaDestroyImage(context.allocator, img.image, img.allocation); + img.image = VK_NULL_HANDLE; + img.imageView = VK_NULL_HANDLE; + img.allocation = VK_NULL_HANDLE; + img.imageExtent = {}; + img.imageFormat = {}; } void will_engine::ResourceManager::destroySampler(const VkSampler& sampler) const diff --git a/src/renderer/resource_manager.h b/src/renderer/resource_manager.h index d26134af..99f392cd 100644 --- a/src/renderer/resource_manager.h +++ b/src/renderer/resource_manager.h @@ -66,7 +66,7 @@ class ResourceManager [[nodiscard]] AllocatedImage createCubemap(VkExtent3D size, VkFormat format, VkImageUsageFlags usage, bool mipmapped = false) const; - void destroyImage(const AllocatedImage& img) const; + void destroyImage(AllocatedImage& img) const; [[nodiscard]] DescriptorBufferSampler createDescriptorBufferSampler(VkDescriptorSetLayout layout, int32_t maxObjectCount) const; diff --git a/src/renderer/vk_types.h b/src/renderer/vk_types.h index 7e82567f..634e270a 100644 --- a/src/renderer/vk_types.h +++ b/src/renderer/vk_types.h @@ -71,6 +71,7 @@ struct SceneData glm::vec4 jitter{0.0f}; glm::vec2 renderTargetSize{}; + glm::vec2 texelSize{}; float deltaTime{}; }; From 14bedb80cf528d73af1ff42057b02c30e15363d6 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sun, 23 Mar 2025 17:37:12 +0700 Subject: [PATCH 03/27] GTAO main ambient occlusion calculation pipeline. --- ...efilter.comp => gtao_depth_prefilter.comp} | 0 .../ambient_occlusion_types.h | 7 + .../ground_truth_ambient_occlusion.cpp | 241 +++++++++++++++++- .../ground_truth_ambient_occlusion.h | 23 +- src/renderer/resource_manager.cpp | 3 +- src/renderer/resource_manager.h | 2 +- 6 files changed, 264 insertions(+), 12 deletions(-) rename shaders/ambient_occlusion/ground_truth/{gtaodepthprefilter.comp => gtao_depth_prefilter.comp} (100%) diff --git a/shaders/ambient_occlusion/ground_truth/gtaodepthprefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp similarity index 100% rename from shaders/ambient_occlusion/ground_truth/gtaodepthprefilter.comp rename to shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index a65394df..f9d7f0cd 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -29,6 +29,13 @@ struct GTAOPushConstants float temporalWeight; float spatialFilterRadius; }; + +struct GTAODrawInfo +{ + GTAOPushConstants pushConstants{}; + VkDescriptorBufferBindingInfoEXT sceneDataBinding{}; + VkDeviceSize sceneDataOffset{0}; +}; } #endif //AMBIENT_OCCLUSION_TYPES_H diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 580d1a91..be502b76 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -42,6 +42,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth depthPrefilterPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); createDepthPrefilterPipeline(); + createAmbientOcclusionPipeline(); depthPrefilterDescriptorBuffer = resourceManager.createDescriptorBufferSampler(depthPrefilterSetLayout, 1); @@ -51,26 +52,138 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth usage |= VK_IMAGE_USAGE_SAMPLED_BIT; VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(depthPrefilterFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); - // 5 mips, suggested by Intel's implementation // https://github.com/GameTechDev/XeGTAO imgInfo.mipLevels = DEPTH_PREFILTER_MIP_COUNT; - depthPrefilterImage = resourceManager.createImage(imgInfo); - VkImageViewCreateInfo viewInfo = vk_helpers::imageviewCreateInfo(depthPrefilterFormat, depthPrefilterImage.image, VK_IMAGE_ASPECT_COLOR_BIT); for (int32_t i = 0; i < DEPTH_PREFILTER_MIP_COUNT; ++i) { viewInfo.subresourceRange.baseMipLevel = i; depthPrefilterImageViews[i] = resourceManager.createImageView(viewInfo); } + + VkSamplerCreateInfo samplerInfo = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO}; + samplerInfo.magFilter = VK_FILTER_NEAREST; + samplerInfo.minFilter = VK_FILTER_NEAREST; + samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST; + samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.anisotropyEnable = VK_FALSE; + samplerInfo.maxAnisotropy = 1.0f; + samplerInfo.compareEnable = VK_FALSE; + samplerInfo.minLod = 0.0f; + samplerInfo.maxLod = 0.0f; + + depthPrefilterSampler = resourceManager.createSampler(samplerInfo); + } + + // AO Calculation + { + DescriptorLayoutBuilder layoutBuilder; + layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // pre-filtered depth + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // normal buffer + layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao output + + ambientOcclusionSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + + VkPushConstantRange pushConstants{}; + pushConstants.offset = 0; + pushConstants.size = sizeof(GTAOPushConstants); + pushConstants.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT; + + VkDescriptorSetLayout setLayouts[2]; + setLayouts[0] = resourceManager.getSceneDataLayout(); + setLayouts[1] = ambientOcclusionSetLayout; + + VkPipelineLayoutCreateInfo layoutInfo{}; + layoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO; + layoutInfo.pNext = nullptr; + layoutInfo.pSetLayouts = setLayouts; + layoutInfo.setLayoutCount = 2; + layoutInfo.pPushConstantRanges = &pushConstants; + layoutInfo.pushConstantRangeCount = 1; + + ambientOcclusionPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); + createAmbientOcclusionPipeline(); + + ambientOcclusionDescriptorBuffer = resourceManager.createDescriptorBufferSampler(ambientOcclusionSetLayout, 1); + + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + ambientOcclusionImage = resourceManager.createImage(imgInfo); + + // Depth Mip sampler + { + VkSamplerCreateInfo samplerInfo = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO}; + samplerInfo.magFilter = VK_FILTER_NEAREST; + samplerInfo.minFilter = VK_FILTER_NEAREST; + samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST; + samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.anisotropyEnable = VK_FALSE; + samplerInfo.maxAnisotropy = 1.0f; + samplerInfo.compareEnable = VK_FALSE; + samplerInfo.minLod = 0.0f; + samplerInfo.maxLod = DEPTH_PREFILTER_MIP_COUNT - 1; + + ambientOcclusionDepthSampler = resourceManager.createSampler(samplerInfo); + } + + // Normals sampler + { + VkSamplerCreateInfo samplerInfo = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO}; + samplerInfo.magFilter = VK_FILTER_NEAREST; + samplerInfo.minFilter = VK_FILTER_NEAREST; + samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST; + samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.anisotropyEnable = VK_FALSE; + samplerInfo.maxAnisotropy = 1.0f; + samplerInfo.compareEnable = VK_FALSE; + samplerInfo.minLod = 0.0f; + samplerInfo.maxLod = 0.0f; + + ambientOcclusionNormalsSampler = resourceManager.createSampler(samplerInfo); + } } } will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTruthAmbientOcclusionPipeline() -{} +{ + // Depth Prefilter Resources + resourceManager.destroyDescriptorSetLayout(depthPrefilterSetLayout); + resourceManager.destroyPipelineLayout(depthPrefilterPipelineLayout); + resourceManager.destroyPipeline(depthPrefilterPipeline); + + for (int32_t i = 0; i < DEPTH_PREFILTER_MIP_COUNT; ++i) { + resourceManager.destroyImageView(depthPrefilterImageViews[i]); + } + + resourceManager.destroyImage(depthPrefilterImage); + resourceManager.destroySampler(depthPrefilterSampler); + + resourceManager.destroyDescriptorBuffer(depthPrefilterDescriptorBuffer); + + // AO Resources + resourceManager.destroyDescriptorSetLayout(ambientOcclusionSetLayout); + resourceManager.destroyPipelineLayout(ambientOcclusionPipelineLayout); + resourceManager.destroyPipeline(ambientOcclusionPipeline); -void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(VkSampler depthImageSampler, VkImageView depthImageView) + resourceManager.destroySampler(ambientOcclusionDepthSampler); + resourceManager.destroySampler(ambientOcclusionNormalsSampler); + resourceManager.destroyImage(ambientOcclusionImage); + + resourceManager.destroyDescriptorBuffer(ambientOcclusionDescriptorBuffer); +} + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(const VkImageView depthImageView) { std::vector imageDescriptors{}; imageDescriptors.reserve(1 + DEPTH_PREFILTER_MIP_COUNT); @@ -78,7 +191,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupD imageDescriptors.push_back( { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, - {depthImageSampler, depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + {depthPrefilterSampler, depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, false } ); @@ -95,10 +208,101 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupD resourceManager.setupDescriptorBufferSampler(depthPrefilterDescriptorBuffer, imageDescriptors, 0); } +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView) +{ + std::vector imageDescriptors{}; + imageDescriptors.reserve(2); + + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + {ambientOcclusionDepthSampler, depthPrefilterImage.imageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + false + }); + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + {ambientOcclusionNormalsSampler, normalsImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + false + }); + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, + {VK_NULL_HANDLE, ambientOcclusionImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + false + }); +} + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) +{ + VkDebugUtilsLabelEXT label{}; + label.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT; + label.pLabelName = "GT Ambient Occlusion"; + vkCmdBeginDebugUtilsLabelEXT(cmd, &label); + + vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + + // Depth Prefilter + { + vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipeline); + vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); + + VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; + bindingInfos[0] = drawInfo.sceneDataBinding; + bindingInfos[1] = depthPrefilterDescriptorBuffer.getDescriptorBufferBindingInfo(); + vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); + + constexpr VkDeviceSize zeroOffset{0}; + constexpr uint32_t sceneDataIndex{0}; + constexpr uint32_t descriptorIndex{1}; + + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 1, &sceneDataIndex, &drawInfo.sceneDataOffset); + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 1, 1, &descriptorIndex, &zeroOffset); + + const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + vkCmdDispatch(cmd, x, y, 1); + vkCmdEndRendering(cmd); + } + + vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + + // Ambient Occlusion + { + vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); + vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); + + VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; + bindingInfos[0] = drawInfo.sceneDataBinding; + bindingInfos[1] = ambientOcclusionDescriptorBuffer.getDescriptorBufferBindingInfo(); + vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); + + constexpr VkDeviceSize zeroOffset{0}; + constexpr uint32_t sceneDataIndex{0}; + constexpr uint32_t descriptorIndex{1}; + + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 0, 1, &sceneDataIndex, &drawInfo.sceneDataOffset); + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 1, 1, &descriptorIndex, &zeroOffset); + + const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + vkCmdDispatch(cmd, x, y, 1); + vkCmdEndRendering(cmd); + } + + + vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + + + + vkCmdEndDebugUtilsLabelEXT(cmd); +} + void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createDepthPrefilterPipeline() { resourceManager.destroyPipeline(depthPrefilterPipeline); - VkShaderModule computeShader = resourceManager.createShaderModule("shaders/gtaodepthprefilter.comp"); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/gtao_depth_prefilter.comp"); VkPipelineShaderStageCreateInfo stageInfo{}; stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; @@ -117,3 +321,26 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::create depthPrefilterPipeline = resourceManager.createComputePipeline(pipelineInfo); resourceManager.destroyShaderModule(computeShader); } + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createAmbientOcclusionPipeline() +{ + resourceManager.destroyPipeline(ambientOcclusionPipeline); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/gtao_ambient_occlusion.comp"); + + VkPipelineShaderStageCreateInfo stageInfo{}; + stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; + stageInfo.pNext = nullptr; + stageInfo.stage = VK_SHADER_STAGE_COMPUTE_BIT; + stageInfo.module = computeShader; + stageInfo.pName = "main"; + + VkComputePipelineCreateInfo pipelineInfo{}; + pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO; + pipelineInfo.pNext = nullptr; + pipelineInfo.layout = ambientOcclusionPipelineLayout; + pipelineInfo.stage = stageInfo; + pipelineInfo.flags = VK_PIPELINE_CREATE_DESCRIPTOR_BUFFER_BIT_EXT; + + ambientOcclusionPipeline = resourceManager.createComputePipeline(pipelineInfo); + resourceManager.destroyShaderModule(computeShader); +} diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index 25a78d3b..4abba7b3 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -17,20 +17,26 @@ namespace will_engine::ambient_occlusion class GroundTruthAmbientOcclusionPipeline { public: - GroundTruthAmbientOcclusionPipeline(ResourceManager& resourceManager); + explicit GroundTruthAmbientOcclusionPipeline(ResourceManager& resourceManager); ~GroundTruthAmbientOcclusionPipeline(); - void setupDepthPrefilterDescriptorBuffer(VkSampler depthImageSampler, VkImageView depthImageView); + void setupDepthPrefilterDescriptorBuffer(VkImageView depthImageView); + + void setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView); + + void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo); private: void createDepthPrefilterPipeline(); + void createAmbientOcclusionPipeline(); -private: +private: // Depth prefilter VkDescriptorSetLayout depthPrefilterSetLayout{VK_NULL_HANDLE}; VkPipelineLayout depthPrefilterPipelineLayout{VK_NULL_HANDLE}; VkPipeline depthPrefilterPipeline{VK_NULL_HANDLE}; + VkSampler depthPrefilterSampler{VK_NULL_HANDLE}; // 16 vs 32. look at cost later. VkFormat depthPrefilterFormat{VK_FORMAT_R16_SFLOAT}; AllocatedImage depthPrefilterImage{VK_NULL_HANDLE}; @@ -38,9 +44,20 @@ class GroundTruthAmbientOcclusionPipeline DescriptorBufferSampler depthPrefilterDescriptorBuffer; +private: // ao + VkDescriptorSetLayout ambientOcclusionSetLayout{VK_NULL_HANDLE}; VkPipelineLayout ambientOcclusionPipelineLayout{VK_NULL_HANDLE}; VkPipeline ambientOcclusionPipeline{VK_NULL_HANDLE}; + VkSampler ambientOcclusionDepthSampler{VK_NULL_HANDLE}; + VkSampler ambientOcclusionNormalsSampler{VK_NULL_HANDLE}; + VkFormat ambientOcclusionFormat{VK_FORMAT_R8_UNORM}; + AllocatedImage ambientOcclusionImage{VK_NULL_HANDLE}; + + DescriptorBufferSampler ambientOcclusionDescriptorBuffer; + +private: // + VkPipelineLayout spatialFilteringPipelineLayout{VK_NULL_HANDLE}; VkPipeline spatialFilteringPipeline{VK_NULL_HANDLE}; diff --git a/src/renderer/resource_manager.cpp b/src/renderer/resource_manager.cpp index 14f2b7a5..de3c8bfc 100644 --- a/src/renderer/resource_manager.cpp +++ b/src/renderer/resource_manager.cpp @@ -436,9 +436,10 @@ void will_engine::ResourceManager::destroyImage(AllocatedImage& img) const img.imageFormat = {}; } -void will_engine::ResourceManager::destroySampler(const VkSampler& sampler) const +void will_engine::ResourceManager::destroySampler(VkSampler& sampler) const { vkDestroySampler(context.device, sampler, nullptr); + sampler = VK_NULL_HANDLE; } will_engine::DescriptorBufferSampler will_engine::ResourceManager::createDescriptorBufferSampler(VkDescriptorSetLayout layout, int32_t maxObjectCount) const diff --git a/src/renderer/resource_manager.h b/src/renderer/resource_manager.h index 99f392cd..c767589a 100644 --- a/src/renderer/resource_manager.h +++ b/src/renderer/resource_manager.h @@ -55,7 +55,7 @@ class ResourceManager [[nodiscard]] VkSampler createSampler(const VkSamplerCreateInfo& createInfo) const; - void destroySampler(const VkSampler& sampler) const; + void destroySampler(VkSampler& sampler) const; [[nodiscard]] AllocatedImage createImage(const VkImageCreateInfo& createInfo) const; From 3731e2302aadeded370e0a90c029c0d211c73752 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 24 Mar 2025 21:18:32 +0700 Subject: [PATCH 04/27] Spatial Prefilter and Temporal Accumulation. --- .../ground_truth/gtao_depth_prefilter.comp | 11 ++ .../ground_truth/gtao_main_pass.comp | 13 ++ .../ground_truth/gtao_spatial_filter.comp | 14 ++ .../gtao_temporal_accumulation.comp | 15 ++ src/core/engine.cpp | 6 + src/core/engine.h | 8 + .../assets/render_object/render_object.cpp | 2 +- .../ground_truth_ambient_occlusion.cpp | 171 +++++++++++++++++- .../ground_truth_ambient_occlusion.h | 29 ++- 9 files changed, 254 insertions(+), 15 deletions(-) create mode 100644 shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp create mode 100644 shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp create mode 100644 shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 747d43d5..e92bdd91 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -1,5 +1,16 @@ #version 460 +#include "scene.glsl" + +// layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl + +layout (set = 1, binding = 0) uniform sampler2D depthImage; +layout (r16f, set = 1, binding = 1) uniform image2D depthMip0; +layout (r16f, set = 1, binding = 2) uniform image2D depthMip1; +layout (r16f, set = 1, binding = 3) uniform image2D depthMip2; +layout (r16f, set = 1, binding = 4) uniform image2D depthMip3; +layout (r16f, set = 1, binding = 5) uniform image2D depthMip4; + void main() { } diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp new file mode 100644 index 00000000..6f9db5be --- /dev/null +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -0,0 +1,13 @@ +#version 460 + +#include "scene.glsl" + +// layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl + +layout (set = 1, binding = 0) uniform sampler2D prefilteredDepth; +layout (set = 1, binding = 1) uniform sampler2D normalBuffer; +layout (r8, set = 1, binding = 2) uniform image2D aoOutput; + +void main() { + +} diff --git a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp new file mode 100644 index 00000000..ee6acbe2 --- /dev/null +++ b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp @@ -0,0 +1,14 @@ +#version 460 + +#include "scene.glsl" + +// layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl + +layout (set = 1, binding = 0) uniform sampler2D rawAO; +layout (set = 1, binding = 1) uniform sampler2D depthBuffer; +layout (set = 1, binding = 2) uniform sampler2D normalBuffer; +layout (r8, set = 1, binding = 3) uniform image2D filteredAO; + +void main() { + +} diff --git a/shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp b/shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp new file mode 100644 index 00000000..bf32d9ad --- /dev/null +++ b/shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp @@ -0,0 +1,15 @@ +#version 460 + +#include "scene.glsl" + +// layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl + +layout (set = 1, binding = 0) uniform sampler2D filteredAO; +layout (set = 1, binding = 1) uniform sampler2D historyOutputAO; +layout (set = 1, binding = 2) uniform sampler2D depthBuffer; +layout (set = 1, binding = 2) uniform sampler2D velocityBuffer; +layout (r8, set = 1, binding = 2) uniform image2D outputAO; + +void main() { + +} diff --git a/src/core/engine.cpp b/src/core/engine.cpp index de85f276..6e5b420c 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -27,6 +27,7 @@ #include "src/renderer/assets/render_object/render_object.h" #include "src/renderer/descriptor_buffer/descriptor_buffer_uniform.h" #include "src/renderer/environment/environment.h" +#include "src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h" #include "src/renderer/lighting/shadows/cascaded_shadow_map.h" #include "src/renderer/pipelines/deferred_mrt/deferred_mrt.h" #include "src/renderer/pipelines/deferred_resolve/deferred_resolve.h" @@ -199,11 +200,15 @@ void Engine::initRenderer() environmentPipeline = new environment_pipeline::EnvironmentPipeline(*resourceManager, environmentMap->getCubemapDescriptorSetLayout()); terrainPipeline = new terrain::TerrainPipeline(*resourceManager); deferredMrtPipeline = new deferred_mrt::DeferredMrtPipeline(*resourceManager); + ambientOcclusionPipeline = new ambient_occlusion::GroundTruthAmbientOcclusionPipeline(*resourceManager); deferredResolvePipeline = new deferred_resolve::DeferredResolvePipeline(*resourceManager, environmentMap->getDiffSpecMapDescriptorSetlayout(), cascadedShadowMap->getCascadedShadowMapUniformLayout(), cascadedShadowMap->getCascadedShadowMapSamplerLayout()); temporalAntialiasingPipeline = new temporal_antialiasing_pipeline::TemporalAntialiasingPipeline(*resourceManager); postProcessPipeline = new post_process_pipeline::PostProcessPipeline(*resourceManager); + ambientOcclusionPipeline->setupDepthPrefilterDescriptorBuffer(depthImage.imageView); + ambientOcclusionPipeline->setupAmbientOcclusionDescriptorBuffer(normalRenderTarget.imageView); + const deferred_resolve::DeferredResolveDescriptor deferredResolveDescriptor{ normalRenderTarget.imageView, albedoRenderTarget.imageView, @@ -654,6 +659,7 @@ void Engine::cleanup() delete environmentPipeline; delete terrainPipeline; delete deferredMrtPipeline; + delete ambientOcclusionPipeline; delete deferredResolvePipeline; delete temporalAntialiasingPipeline; delete postProcessPipeline; diff --git a/src/core/engine.h b/src/core/engine.h index a5abaa7f..547d5b5c 100644 --- a/src/core/engine.h +++ b/src/core/engine.h @@ -26,6 +26,11 @@ class VulkanContext; namespace will_engine { +namespace ambient_occlusion +{ + class GroundTruthAmbientOcclusionPipeline; +} + namespace terrain { class TerrainChunk; @@ -142,8 +147,10 @@ class Engine ImmediateSubmitter* immediate = nullptr; ResourceManager* resourceManager = nullptr; identifier::IdentifierManager* identifierManager = nullptr; + environment::Environment* environmentMap{nullptr}; cascaded_shadows::CascadedShadowMap* cascadedShadowMap{nullptr}; + terrain::TerrainManager* terrainManager{nullptr}; ImguiWrapper* imguiWrapper = nullptr; @@ -195,6 +202,7 @@ class Engine terrain::TerrainPipeline* terrainPipeline{nullptr}; deferred_mrt::DeferredMrtPipeline* deferredMrtPipeline{nullptr}; deferred_resolve::DeferredResolvePipeline* deferredResolvePipeline{nullptr}; + ambient_occlusion::GroundTruthAmbientOcclusionPipeline* ambientOcclusionPipeline{nullptr}; temporal_antialiasing_pipeline::TemporalAntialiasingPipeline* temporalAntialiasingPipeline{nullptr}; post_process_pipeline::PostProcessPipeline* postProcessPipeline{nullptr}; diff --git a/src/renderer/assets/render_object/render_object.cpp b/src/renderer/assets/render_object/render_object.cpp index 9041b37c..9de1b356 100644 --- a/src/renderer/assets/render_object/render_object.cpp +++ b/src/renderer/assets/render_object/render_object.cpp @@ -599,7 +599,7 @@ void RenderObject::unload() resourceManager.destroyImage(image); } - for (const auto& sampler : samplers) { + for (auto& sampler : samplers) { if (sampler == resourceManager.getDefaultSamplerNearest() || sampler == resourceManager.getDefaultSamplerLinear()) { //dont destroy the default samplers continue; diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index be502b76..14231bda 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -4,6 +4,8 @@ #include "ground_truth_ambient_occlusion.h" +#include + #include "src/renderer/renderer_constants.h" #include "src/renderer/vk_descriptors.h" #include "src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h" @@ -13,7 +15,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth // Depth Pre-filtering { DescriptorLayoutBuilder layoutBuilder; - layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // depth image + layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT depth buffer layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 0 layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 1 layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 2 @@ -40,9 +42,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth layoutInfo.pushConstantRangeCount = 1; depthPrefilterPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); - createDepthPrefilterPipeline(); - createAmbientOcclusionPipeline(); depthPrefilterDescriptorBuffer = resourceManager.createDescriptorBufferSampler(depthPrefilterSetLayout, 1); @@ -83,7 +83,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth { DescriptorLayoutBuilder layoutBuilder; layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // pre-filtered depth - layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // normal buffer + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT normal buffer layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao output ambientOcclusionSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); @@ -153,6 +153,94 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth ambientOcclusionNormalsSampler = resourceManager.createSampler(samplerInfo); } } + + // Spatial Filtering + { + DescriptorLayoutBuilder layoutBuilder; + layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // raw ao + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT depth buffer + layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT normal buffer + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // filtered ao + + spatialFilteringSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + + VkPushConstantRange pushConstants{}; + pushConstants.offset = 0; + pushConstants.size = sizeof(GTAOPushConstants); + pushConstants.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT; + + VkDescriptorSetLayout setLayouts[2]; + setLayouts[0] = resourceManager.getSceneDataLayout(); + setLayouts[1] = spatialFilteringSetLayout; + + VkPipelineLayoutCreateInfo layoutInfo{}; + layoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO; + layoutInfo.pNext = nullptr; + layoutInfo.pSetLayouts = setLayouts; + layoutInfo.setLayoutCount = 2; + layoutInfo.pPushConstantRanges = &pushConstants; + layoutInfo.pushConstantRangeCount = 1; + + spatialFilteringPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); + createSpatialFilteringPipeline(); + + spatialFilteringDescriptorBuffer = resourceManager.createDescriptorBufferSampler(spatialFilteringSetLayout, 1); + + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + spatialFilteringImage = resourceManager.createImage(imgInfo); + } + + // Temporal Accumulation + { + DescriptorLayoutBuilder layoutBuilder; + layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // filtered ao + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // final output history + layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT velocity buffer + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT depth buffer + layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // final output + + temporalAccumulationSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + + VkPushConstantRange pushConstants{}; + pushConstants.offset = 0; + pushConstants.size = sizeof(GTAOPushConstants); + pushConstants.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT; + + VkDescriptorSetLayout setLayouts[2]; + setLayouts[0] = resourceManager.getSceneDataLayout(); + setLayouts[1] = temporalAccumulationSetLayout; + + VkPipelineLayoutCreateInfo layoutInfo{}; + layoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO; + layoutInfo.pNext = nullptr; + layoutInfo.pSetLayouts = setLayouts; + layoutInfo.setLayoutCount = 2; + layoutInfo.pPushConstantRanges = &pushConstants; + layoutInfo.pushConstantRangeCount = 1; + + temporalAccumulationPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); + createTemporalAccumulationPipeline(); + + temporalAccumulationDescriptorBuffer = resourceManager.createDescriptorBufferSampler(temporalAccumulationSetLayout, 1); + + + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + historyOutputImage = resourceManager.createImage(imgInfo); + + usage = {}; + usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + ambientOcclusionOutputImage = resourceManager.createImage(imgInfo); + } } will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTruthAmbientOcclusionPipeline() @@ -181,6 +269,26 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTrut resourceManager.destroyImage(ambientOcclusionImage); resourceManager.destroyDescriptorBuffer(ambientOcclusionDescriptorBuffer); + + // Spatial Filtering Resources + resourceManager.destroyDescriptorSetLayout(spatialFilteringSetLayout); + resourceManager.destroyPipelineLayout(spatialFilteringPipelineLayout); + resourceManager.destroyPipeline(spatialFilteringPipeline); + + resourceManager.destroyImage(spatialFilteringImage); + + resourceManager.destroyDescriptorBuffer(spatialFilteringDescriptorBuffer); + + + // Temporal Accumulation Resources + resourceManager.destroyDescriptorSetLayout(temporalAccumulationSetLayout); + resourceManager.destroyPipelineLayout(temporalAccumulationPipelineLayout); + resourceManager.destroyPipeline(temporalAccumulationPipeline); + + resourceManager.destroyImage(historyOutputImage); + resourceManager.destroyImage(ambientOcclusionOutputImage); + + resourceManager.destroyDescriptorBuffer(temporalAccumulationDescriptorBuffer); } void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(const VkImageView depthImageView) @@ -231,9 +339,11 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupA {VK_NULL_HANDLE, ambientOcclusionImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, false }); + + resourceManager.setupDescriptorBufferSampler(ambientOcclusionDescriptorBuffer, imageDescriptors, 0); } -void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const { VkDebugUtilsLabelEXT label{}; label.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT; @@ -295,14 +405,13 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vkCmdEndDebugUtilsLabelEXT(cmd); } void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createDepthPrefilterPipeline() { resourceManager.destroyPipeline(depthPrefilterPipeline); - VkShaderModule computeShader = resourceManager.createShaderModule("shaders/gtao_depth_prefilter.comp"); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp"); VkPipelineShaderStageCreateInfo stageInfo{}; stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; @@ -325,7 +434,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::create void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createAmbientOcclusionPipeline() { resourceManager.destroyPipeline(ambientOcclusionPipeline); - VkShaderModule computeShader = resourceManager.createShaderModule("shaders/gtao_ambient_occlusion.comp"); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp"); VkPipelineShaderStageCreateInfo stageInfo{}; stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; @@ -344,3 +453,49 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::create ambientOcclusionPipeline = resourceManager.createComputePipeline(pipelineInfo); resourceManager.destroyShaderModule(computeShader); } + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createSpatialFilteringPipeline() +{ + resourceManager.destroyPipeline(spatialFilteringPipeline); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp"); + + VkPipelineShaderStageCreateInfo stageInfo{}; + stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; + stageInfo.pNext = nullptr; + stageInfo.stage = VK_SHADER_STAGE_COMPUTE_BIT; + stageInfo.module = computeShader; + stageInfo.pName = "main"; + + VkComputePipelineCreateInfo pipelineInfo{}; + pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO; + pipelineInfo.pNext = nullptr; + pipelineInfo.layout = spatialFilteringPipelineLayout; + pipelineInfo.stage = stageInfo; + pipelineInfo.flags = VK_PIPELINE_CREATE_DESCRIPTOR_BUFFER_BIT_EXT; + + spatialFilteringPipeline = resourceManager.createComputePipeline(pipelineInfo); + resourceManager.destroyShaderModule(computeShader); +} + +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createTemporalAccumulationPipeline() +{ + resourceManager.destroyPipeline(temporalAccumulationPipeline); + VkShaderModule computeShader = resourceManager.createShaderModule("shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp"); + + VkPipelineShaderStageCreateInfo stageInfo{}; + stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; + stageInfo.pNext = nullptr; + stageInfo.stage = VK_SHADER_STAGE_COMPUTE_BIT; + stageInfo.module = computeShader; + stageInfo.pName = "main"; + + VkComputePipelineCreateInfo pipelineInfo{}; + pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO; + pipelineInfo.pNext = nullptr; + pipelineInfo.layout = temporalAccumulationPipelineLayout; + pipelineInfo.stage = stageInfo; + pipelineInfo.flags = VK_PIPELINE_CREATE_DESCRIPTOR_BUFFER_BIT_EXT; + + temporalAccumulationPipeline = resourceManager.createComputePipeline(pipelineInfo); + resourceManager.destroyShaderModule(computeShader); +} diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index 4abba7b3..48fa8840 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -6,7 +6,6 @@ #define GROUND_TRUTH_AMBIENT_OCCLUSION_H #include -#include #include "src/renderer/resource_manager.h" #include "src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h" @@ -25,18 +24,21 @@ class GroundTruthAmbientOcclusionPipeline void setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView); - void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo); + void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const; private: void createDepthPrefilterPipeline(); void createAmbientOcclusionPipeline(); + void createSpatialFilteringPipeline(); + void createTemporalAccumulationPipeline(); -private: // Depth prefilter +private: // Depth Pre-filter VkDescriptorSetLayout depthPrefilterSetLayout{VK_NULL_HANDLE}; VkPipelineLayout depthPrefilterPipelineLayout{VK_NULL_HANDLE}; VkPipeline depthPrefilterPipeline{VK_NULL_HANDLE}; VkSampler depthPrefilterSampler{VK_NULL_HANDLE}; + // 16 vs 32. look at cost later. VkFormat depthPrefilterFormat{VK_FORMAT_R16_SFLOAT}; AllocatedImage depthPrefilterImage{VK_NULL_HANDLE}; @@ -44,26 +46,41 @@ class GroundTruthAmbientOcclusionPipeline DescriptorBufferSampler depthPrefilterDescriptorBuffer; -private: // ao +private: // Ambient Occlusion VkDescriptorSetLayout ambientOcclusionSetLayout{VK_NULL_HANDLE}; VkPipelineLayout ambientOcclusionPipelineLayout{VK_NULL_HANDLE}; VkPipeline ambientOcclusionPipeline{VK_NULL_HANDLE}; VkSampler ambientOcclusionDepthSampler{VK_NULL_HANDLE}; VkSampler ambientOcclusionNormalsSampler{VK_NULL_HANDLE}; + + // 8 is supposedly enough? VkFormat ambientOcclusionFormat{VK_FORMAT_R8_UNORM}; AllocatedImage ambientOcclusionImage{VK_NULL_HANDLE}; DescriptorBufferSampler ambientOcclusionDescriptorBuffer; -private: // - +private: // Spatial Filtering + VkDescriptorSetLayout spatialFilteringSetLayout{VK_NULL_HANDLE}; VkPipelineLayout spatialFilteringPipelineLayout{VK_NULL_HANDLE}; VkPipeline spatialFilteringPipeline{VK_NULL_HANDLE}; + AllocatedImage spatialFilteringImage{VK_NULL_HANDLE}; + + DescriptorBufferSampler spatialFilteringDescriptorBuffer; + +private: // Temporal Accumulation + VkDescriptorSetLayout temporalAccumulationSetLayout{VK_NULL_HANDLE}; VkPipelineLayout temporalAccumulationPipelineLayout{VK_NULL_HANDLE}; VkPipeline temporalAccumulationPipeline{VK_NULL_HANDLE}; + AllocatedImage historyOutputImage{VK_NULL_HANDLE}; + + DescriptorBufferSampler temporalAccumulationDescriptorBuffer; + +private: // Output + AllocatedImage ambientOcclusionOutputImage{VK_NULL_HANDLE}; + private: ResourceManager& resourceManager; }; From 7d73ffbc4bfd2abec65d7bd7acb78a5cf2dd4f4a Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 24 Mar 2025 23:09:34 +0700 Subject: [PATCH 05/27] Depth prefilter shader progress. --- .../ground_truth/gtao_depth_prefilter.comp | 66 ++++++++++++++-- shaders/include/scene.glsl | 1 + src/core/engine.cpp | 2 +- .../ambient_occlusion_types.h | 11 ++- .../ground_truth_ambient_occlusion.cpp | 77 ++++++++++--------- src/renderer/vk_types.h | 2 + 6 files changed, 115 insertions(+), 44 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index e92bdd91..7b04259e 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -5,12 +5,66 @@ // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl layout (set = 1, binding = 0) uniform sampler2D depthImage; -layout (r16f, set = 1, binding = 1) uniform image2D depthMip0; -layout (r16f, set = 1, binding = 2) uniform image2D depthMip1; -layout (r16f, set = 1, binding = 3) uniform image2D depthMip2; -layout (r16f, set = 1, binding = 4) uniform image2D depthMip3; -layout (r16f, set = 1, binding = 5) uniform image2D depthMip4; +layout (r16f, set = 1, binding = 1) uniform image2D outDepth0; +layout (r16f, set = 1, binding = 2) uniform image2D outDepth1; +layout (r16f, set = 1, binding = 3) uniform image2D outDepth2; +layout (r16f, set = 1, binding = 4) uniform image2D outDepth3; +layout (r16f, set = 1, binding = 5) uniform image2D outDepth4; -void main() { +layout (push_constant) uniform PushConstants { + float depthLinearizeMult; + float depthLinearizeAdd; + + vec2 ndcToViewMult; + vec2 ndcToViewAdd; + + float radius; + float faloff; + float strength; + float radiusMultiplier; + + int numDirections; + int numSteps; + + float spatialFilterRadius; + + float temporalWeight; +} pushConstants; + +shared float g_scratchDepths[16][16]; + +layout(local_size_x = 16, local_size_y = 16) in; + +//float screenSpaceToViewSpaceDepth(float screenDepth) { +// vec4 clipSpacePos = vec4(0.0, 0.0, screenDepth, 1.0); +// vec4 viewSpacePos = sceneData.invProjection * clipSpacePos; +// return -viewSpacePos.z / viewSpacePos.w; +//} +// +//float clampDepth(){ +// // using half float precision +// return clamp(depth, 0.0, 65504.0); +//} + +void main() { +// vec2 dispatchID = gl_GlobalInvocationID.xy; +// uvec2 groupThreadID = gl_LocalInvocationID.xy; +// +// // mip 0 +// const uvec2 baseCoord = dispatchThreadID; +// const uvec2 pixCoord = baseCoord * 2;// We process 2x2 pixels in MIP 0 +// +// vec2 uvCoord = vec2(pixCoord) * sceneData.texelSize; +// +// vec4 depths4; +// depths4.w = texture(depthImage, uvCoord + vec2(0.0, 0.0) * sceneData.texelSize).r; +// depths4.z = texture(depthImage, uvCoord + vec2(1.0, 0.0) * sceneData.texelSize).r; +// depths4.x = texture(depthImage, uvCoord + vec2(0.0, 1.0) * sceneData.texelSize).r; +// depths4.y = texture(depthImage, uvCoord + vec2(1.0, 1.0) * sceneData.texelSize).r; +// +// float depth0 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.w)); +// float depth1 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.z)); +// float depth2 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.x)); +// float depth3 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.y)); } diff --git a/shaders/include/scene.glsl b/shaders/include/scene.glsl index 9da8db70..04176d1c 100644 --- a/shaders/include/scene.glsl +++ b/shaders/include/scene.glsl @@ -27,5 +27,6 @@ layout (std140, set = 0, binding = 0) uniform SceneData { vec2 renderTargetSize; // equal to 1 / renderTargetSize vec2 texelSize; + vec2 cameraPlanes; float deltaTime; } sceneData; \ No newline at end of file diff --git a/src/core/engine.cpp b/src/core/engine.cpp index 6e5b420c..6c567e87 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -395,6 +395,7 @@ void Engine::updateRender(const float deltaTime, const int32_t currentFrameOverl pSceneData->renderTargetSize = {RENDER_EXTENT_WIDTH, RENDER_EXTENT_HEIGHT}; pSceneData->texelSize = {1.0f / RENDER_EXTENT_WIDTH, 1.0f / RENDER_EXTENT_HEIGHT}; + pSceneData->cameraPlanes = {camera->getNearPlane(), camera->getFarPlane()}; pSceneData->deltaTime = deltaTime; @@ -415,7 +416,6 @@ void Engine::updateRender(const float deltaTime, const int32_t currentFrameOverl pDebugSceneData->prevCameraWorldPos = glm::vec4(0.0f); pDebugSceneData->cameraWorldPos = glm::vec4(0.0f); - pDebugSceneData->renderTargetSize = {RENDER_EXTENT_WIDTH, RENDER_EXTENT_HEIGHT}; pDebugSceneData->deltaTime = deltaTime; } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index f9d7f0cd..918fbd39 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -6,14 +6,20 @@ #define AMBIENT_OCCLUSION_TYPES_H #include +#include "src/core/camera/camera.h" + namespace will_engine::ambient_occlusion { static constexpr int32_t DEPTH_PREFILTER_MIP_COUNT = 5; struct GTAOPushConstants { - glm::vec2 viewportSize; - glm::vec2 viewportPixelSize; + // Depth prefilter parameters + float depthLinearizeMult; + float depthLinearizeAdd; + + glm::vec2 ndcToViewMult; + glm::vec2 ndcToViewAdd; // AO parameters float radius; @@ -32,6 +38,7 @@ struct GTAOPushConstants struct GTAODrawInfo { + Camera* camera{nullptr}; GTAOPushConstants pushConstants{}; VkDescriptorBufferBindingInfoEXT sceneDataBinding{}; VkDeviceSize sceneDataOffset{0}; diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 14231bda..bddb0fd7 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -350,6 +350,12 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V label.pLabelName = "GT Ambient Occlusion"; vkCmdBeginDebugUtilsLabelEXT(cmd, &label); + GTAOPushConstants push = drawInfo.pushConstants; + glm::mat4 projMatrix = drawInfo.camera->getProjMatrix(); + push.depthLinearizeMult = -projMatrix[2][3]; + push.depthLinearizeAdd = projMatrix[2][2]; + + vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); // Depth Prefilter @@ -362,48 +368,49 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V bindingInfos[1] = depthPrefilterDescriptorBuffer.getDescriptorBufferBindingInfo(); vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - constexpr VkDeviceSize zeroOffset{0}; - constexpr uint32_t sceneDataIndex{0}; - constexpr uint32_t descriptorIndex{1}; + constexpr std::array indices{0, 1}; + const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 1, &sceneDataIndex, &drawInfo.sceneDataOffset); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 1, 1, &descriptorIndex, &zeroOffset); + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); - const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); - const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + // divided by 2 because depth prepass operates on 2x2 (still input4 -> output4) + x /= 2; + y /= 2; vkCmdDispatch(cmd, x, y, 1); vkCmdEndRendering(cmd); } vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - - // Ambient Occlusion - { - vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); - vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); - - VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; - bindingInfos[0] = drawInfo.sceneDataBinding; - bindingInfos[1] = ambientOcclusionDescriptorBuffer.getDescriptorBufferBindingInfo(); - vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - - constexpr VkDeviceSize zeroOffset{0}; - constexpr uint32_t sceneDataIndex{0}; - constexpr uint32_t descriptorIndex{1}; - - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 0, 1, &sceneDataIndex, &drawInfo.sceneDataOffset); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 1, 1, &descriptorIndex, &zeroOffset); - - const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); - const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); - vkCmdDispatch(cmd, x, y, 1); - vkCmdEndRendering(cmd); - } - - - vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - + // vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + // + // // Ambient Occlusion + // { + // vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); + // vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); + // + // VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; + // bindingInfos[0] = drawInfo.sceneDataBinding; + // bindingInfos[1] = ambientOcclusionDescriptorBuffer.getDescriptorBufferBindingInfo(); + // vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); + // + // constexpr VkDeviceSize zeroOffset{0}; + // constexpr uint32_t sceneDataIndex{0}; + // constexpr uint32_t descriptorIndex{1}; + // + // vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 0, 1, &sceneDataIndex, &drawInfo.sceneDataOffset); + // vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 1, 1, &descriptorIndex, &zeroOffset); + // + // const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + // const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + // vkCmdDispatch(cmd, x, y, 1); + // vkCmdEndRendering(cmd); + // } + // + // + // vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + // vkCmdEndDebugUtilsLabelEXT(cmd); } diff --git a/src/renderer/vk_types.h b/src/renderer/vk_types.h index 634e270a..daafdd3a 100644 --- a/src/renderer/vk_types.h +++ b/src/renderer/vk_types.h @@ -72,6 +72,8 @@ struct SceneData glm::vec2 renderTargetSize{}; glm::vec2 texelSize{}; + + glm::vec2 cameraPlanes{1000.0f, 0.1f}; float deltaTime{}; }; From 41663b05406621cce3b4a05f959393aa3ce93508 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Tue, 25 Mar 2025 20:19:25 +0700 Subject: [PATCH 06/27] GTAO prefilter helper functions. --- .../ground_truth/gtao_depth_prefilter.comp | 18 ++++++++++++++---- 1 file changed, 14 insertions(+), 4 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 7b04259e..557bf783 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -36,16 +36,26 @@ shared float g_scratchDepths[16][16]; layout(local_size_x = 16, local_size_y = 16) in; +// Using this technique because +// depthLinearizeMul / (depthLinearizeAdd - screenDepth) +// doesn't seem to be stable, maybe because of the use of a reversed depth buffer? +float screenToViewSpaceDepth(float projDepth, float clipNear, float clipFar) { + return -clipNear / (clipFar - projDepth * (clipFar - clipNear)) * clipFar; +} + +float clampDepth(){ + // using half float precision + return clamp(depth, 0.0, 65504.0); +} + + //float screenSpaceToViewSpaceDepth(float screenDepth) { // vec4 clipSpacePos = vec4(0.0, 0.0, screenDepth, 1.0); // vec4 viewSpacePos = sceneData.invProjection * clipSpacePos; // return -viewSpacePos.z / viewSpacePos.w; //} // -//float clampDepth(){ -// // using half float precision -// return clamp(depth, 0.0, 65504.0); -//} + void main() { // vec2 dispatchID = gl_GlobalInvocationID.xy; From 2176da22840cc3aa3e4bba4afe9d3eb5b5dda47b Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Tue, 25 Mar 2025 21:03:18 +0700 Subject: [PATCH 07/27] Revert "GTAO prefilter helper functions." This reverts commit 41663b05406621cce3b4a05f959393aa3ce93508. --- .../ground_truth/gtao_depth_prefilter.comp | 18 ++++-------------- 1 file changed, 4 insertions(+), 14 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 557bf783..7b04259e 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -36,26 +36,16 @@ shared float g_scratchDepths[16][16]; layout(local_size_x = 16, local_size_y = 16) in; -// Using this technique because -// depthLinearizeMul / (depthLinearizeAdd - screenDepth) -// doesn't seem to be stable, maybe because of the use of a reversed depth buffer? -float screenToViewSpaceDepth(float projDepth, float clipNear, float clipFar) { - return -clipNear / (clipFar - projDepth * (clipFar - clipNear)) * clipFar; -} - -float clampDepth(){ - // using half float precision - return clamp(depth, 0.0, 65504.0); -} - - //float screenSpaceToViewSpaceDepth(float screenDepth) { // vec4 clipSpacePos = vec4(0.0, 0.0, screenDepth, 1.0); // vec4 viewSpacePos = sceneData.invProjection * clipSpacePos; // return -viewSpacePos.z / viewSpacePos.w; //} // - +//float clampDepth(){ +// // using half float precision +// return clamp(depth, 0.0, 65504.0); +//} void main() { // vec2 dispatchID = gl_GlobalInvocationID.xy; From 5f770136111ec389222865a40e65bad44875fee5 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Tue, 25 Mar 2025 22:25:32 +0700 Subject: [PATCH 08/27] GTAO depth prefilter mip 0. --- .../ambient_occlusion/ground_truth/gtao.comp | 174 +++++++++--------- .../ground_truth/gtao_depth_prefilter.comp | 65 ++++--- src/core/engine.cpp | 8 + .../ground_truth_ambient_occlusion.cpp | 5 +- 4 files changed, 133 insertions(+), 119 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao.comp b/shaders/ambient_occlusion/ground_truth/gtao.comp index c48a172d..a7cc35fb 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao.comp @@ -3,90 +3,90 @@ void main() { } - -void XeGTAO_PrefilterDepths16x16( -uvec2 dispatchThreadID, -uvec2 groupThreadID, -const GTAOConstants consts, -sampler2D sourceNDCDepth, -out writeonly image2D outDepth0, -out writeonly image2D outDepth1, -out writeonly image2D outDepth2, -out writeonly image2D outDepth3, -out writeonly image2D outDepth4 -) { - // MIP 0 - const uvec2 baseCoord = dispatchThreadID; - // 2x because ao image is downsampled - const uvec2 pixCoord = baseCoord * 2u; - - - // todo: get width and height from sceneData - vec2 uvCoord = vec2(pixCoord) * consts.ViewportPixelSize; - - vec4 depths 4; - depths4.w = texture(sourceNDCDepth, uvCoord + vec2(0.0, 0.0) * consts.ViewportPixelSize).r; - depths4.z = texture(sourceNDCDepth, uvCoord + vec2(1.0, 0.0) * consts.ViewportPixelSize).r; - depths4.x = texture(sourceNDCDepth, uvCoord + vec2(0.0, 1.0) * consts.ViewportPixelSize).r; - depths4.y = texture(sourceNDCDepth, uvCoord + vec2(1.0, 1.0) * consts.ViewportPixelSize).r; - - float depth0 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.w, consts)); - float depth1 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.z, consts)); - float depth2 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.x, consts)); - float depth3 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.y, consts)); - - imageStore(outDepth0, ivec2(pixCoord + uvec2(0, 0)), vec4(depth0, 0.0, 0.0, 0.0)); - imageStore(outDepth0, ivec2(pixCoord + uvec2(1, 0)), vec4(depth1, 0.0, 0.0, 0.0)); - imageStore(outDepth0, ivec2(pixCoord + uvec2(0, 1)), vec4(depth2, 0.0, 0.0, 0.0)); - imageStore(outDepth0, ivec2(pixCoord + uvec2(1, 1)), vec4(depth3, 0.0, 0.0, 0.0)); - - // MIP 1 - float dm1 = XeGTAO_DepthMIPFilter(depth0, depth1, depth2, depth3, consts); - imageStore(outDepth1, ivec2(baseCoord), vec4(dm1, 0.0, 0.0, 0.0)); - g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; - - memoryBarrierShared(); - barrier(); - - // MIP 2 - if (all(equal(groupThreadID.xy % 2u, uvec2(0u, 0u)))) { - float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; - float inTR = g_scratchDepths[groupThreadID.x+1u][groupThreadID.y+0u]; - float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+1u]; - float inBR = g_scratchDepths[groupThreadID.x+1u][groupThreadID.y+1u]; - - float dm2 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); - imageStore(outDepth2, ivec2(baseCoord / 2u), vec4(dm2, 0.0, 0.0, 0.0)); - g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm2; - } - - memoryBarrierShared(); - barrier(); - - // MIP 3 - if (all(equal(groupThreadID.xy % 4u, uvec2(0u, 0u)))) { - float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; - float inTR = g_scratchDepths[groupThreadID.x+2u][groupThreadID.y+0u]; - float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+2u]; - float inBR = g_scratchDepths[groupThreadID.x+2u][groupThreadID.y+2u]; - - float dm3 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); - imageStore(outDepth3, ivec2(baseCoord / 4u), vec4(dm3, 0.0, 0.0, 0.0)); - g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm3; - } - - memoryBarrierShared(); - barrier(); - - // MIP 4 - if (all(equal(groupThreadID.xy % 8u, uvec2(0u, 0u)))) { - float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; - float inTR = g_scratchDepths[groupThreadID.x+4u][groupThreadID.y+0u]; - float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+4u]; - float inBR = g_scratchDepths[groupThreadID.x+4u][groupThreadID.y+4u]; - - float dm4 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); - imageStore(outDepth4, ivec2(baseCoord / 8u), vec4(dm4, 0.0, 0.0, 0.0)); - // g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; // commented out as in original - } -} \ No newline at end of file +// +//void XeGTAO_PrefilterDepths16x16( +//uvec2 dispatchThreadID, +//uvec2 groupThreadID, +//const GTAOConstants consts, +//sampler2D sourceNDCDepth, +//out writeonly image2D outDepth0, +//out writeonly image2D outDepth1, +//out writeonly image2D outDepth2, +//out writeonly image2D outDepth3, +//out writeonly image2D outDepth4 +//) { +// // MIP 0 +// const uvec2 baseCoord = dispatchThreadID; +// // 2x because ao image is downsampled +// const uvec2 pixCoord = baseCoord * 2u; +// +// +// // todo: get width and height from sceneData +// vec2 uvCoord = vec2(pixCoord) * consts.ViewportPixelSize; +// +// vec4 depths 4; +// depths4.w = texture(sourceNDCDepth, uvCoord + vec2(0.0, 0.0) * consts.ViewportPixelSize).r; +// depths4.z = texture(sourceNDCDepth, uvCoord + vec2(1.0, 0.0) * consts.ViewportPixelSize).r; +// depths4.x = texture(sourceNDCDepth, uvCoord + vec2(0.0, 1.0) * consts.ViewportPixelSize).r; +// depths4.y = texture(sourceNDCDepth, uvCoord + vec2(1.0, 1.0) * consts.ViewportPixelSize).r; +// +// float depth0 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.w, consts)); +// float depth1 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.z, consts)); +// float depth2 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.x, consts)); +// float depth3 = XeGTAO_ClampDepth(XeGTAO_ScreenSpaceToViewSpaceDepth(depths4.y, consts)); +// +// imageStore(outDepth0, ivec2(pixCoord + uvec2(0, 0)), vec4(depth0, 0.0, 0.0, 0.0)); +// imageStore(outDepth0, ivec2(pixCoord + uvec2(1, 0)), vec4(depth1, 0.0, 0.0, 0.0)); +// imageStore(outDepth0, ivec2(pixCoord + uvec2(0, 1)), vec4(depth2, 0.0, 0.0, 0.0)); +// imageStore(outDepth0, ivec2(pixCoord + uvec2(1, 1)), vec4(depth3, 0.0, 0.0, 0.0)); +// +// // MIP 1 +// float dm1 = XeGTAO_DepthMIPFilter(depth0, depth1, depth2, depth3, consts); +// imageStore(outDepth1, ivec2(baseCoord), vec4(dm1, 0.0, 0.0, 0.0)); +// g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; +// +// memoryBarrierShared(); +// barrier(); +// +// // MIP 2 +// if (all(equal(groupThreadID.xy % 2u, uvec2(0u, 0u)))) { +// float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; +// float inTR = g_scratchDepths[groupThreadID.x+1u][groupThreadID.y+0u]; +// float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+1u]; +// float inBR = g_scratchDepths[groupThreadID.x+1u][groupThreadID.y+1u]; +// +// float dm2 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); +// imageStore(outDepth2, ivec2(baseCoord / 2u), vec4(dm2, 0.0, 0.0, 0.0)); +// g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm2; +// } +// +// memoryBarrierShared(); +// barrier(); +// +// // MIP 3 +// if (all(equal(groupThreadID.xy % 4u, uvec2(0u, 0u)))) { +// float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; +// float inTR = g_scratchDepths[groupThreadID.x+2u][groupThreadID.y+0u]; +// float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+2u]; +// float inBR = g_scratchDepths[groupThreadID.x+2u][groupThreadID.y+2u]; +// +// float dm3 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); +// imageStore(outDepth3, ivec2(baseCoord / 4u), vec4(dm3, 0.0, 0.0, 0.0)); +// g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm3; +// } +// +// memoryBarrierShared(); +// barrier(); +// +// // MIP 4 +// if (all(equal(groupThreadID.xy % 8u, uvec2(0u, 0u)))) { +// float inTL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+0u]; +// float inTR = g_scratchDepths[groupThreadID.x+4u][groupThreadID.y+0u]; +// float inBL = g_scratchDepths[groupThreadID.x+0u][groupThreadID.y+4u]; +// float inBR = g_scratchDepths[groupThreadID.x+4u][groupThreadID.y+4u]; +// +// float dm4 = XeGTAO_DepthMIPFilter(inTL, inTR, inBL, inBR, consts); +// imageStore(outDepth4, ivec2(baseCoord / 8u), vec4(dm4, 0.0, 0.0, 0.0)); +// // g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; // commented out as in original +// } +//} \ No newline at end of file diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 7b04259e..ebe3c57e 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -36,35 +36,42 @@ shared float g_scratchDepths[16][16]; layout(local_size_x = 16, local_size_y = 16) in; -//float screenSpaceToViewSpaceDepth(float screenDepth) { -// vec4 clipSpacePos = vec4(0.0, 0.0, screenDepth, 1.0); -// vec4 viewSpacePos = sceneData.invProjection * clipSpacePos; -// return -viewSpacePos.z / viewSpacePos.w; -//} -// -//float clampDepth(){ -// // using half float precision -// return clamp(depth, 0.0, 65504.0); -//} +float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float depthLinearizeAdd) { + // Optimization by XeGTAO + // https://github.com/GameTechDev/XeGTAO/blob/a5b1686c7ea37788eeb3576b5be47f7c03db532c/Source/Rendering/Shaders/XeGTAO.hlsli#L112 + return depthLinearizeMul / (depthLinearizeAdd - screenDepth); +} + +float clampDepth(float depth){ + // using half float precision + return clamp(depth, 0.0, 65504.0); +} void main() { -// vec2 dispatchID = gl_GlobalInvocationID.xy; -// uvec2 groupThreadID = gl_LocalInvocationID.xy; -// -// // mip 0 -// const uvec2 baseCoord = dispatchThreadID; -// const uvec2 pixCoord = baseCoord * 2;// We process 2x2 pixels in MIP 0 -// -// vec2 uvCoord = vec2(pixCoord) * sceneData.texelSize; -// -// vec4 depths4; -// depths4.w = texture(depthImage, uvCoord + vec2(0.0, 0.0) * sceneData.texelSize).r; -// depths4.z = texture(depthImage, uvCoord + vec2(1.0, 0.0) * sceneData.texelSize).r; -// depths4.x = texture(depthImage, uvCoord + vec2(0.0, 1.0) * sceneData.texelSize).r; -// depths4.y = texture(depthImage, uvCoord + vec2(1.0, 1.0) * sceneData.texelSize).r; -// -// float depth0 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.w)); -// float depth1 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.z)); -// float depth2 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.x)); -// float depth3 = clampDepth(ScreenSpaceToViewSpaceDepth(depths4.y)); + ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + + ivec2 dispatchID = ivec2(gl_GlobalInvocationID.xy); + ivec2 groupThreadID = ivec2(gl_LocalInvocationID.xy); + + // MIP 0 + const ivec2 baseCoord = dispatchID; + const ivec2 pixCoord = baseCoord * 2;// We process 2x2 pixels in MIP 0 + + vec2 uvCoord = vec2(pixCoord) * sceneData.texelSize; + + vec4 depths4; + depths4.x = texture(depthImage, uvCoord + vec2(0.0, 0.0) * sceneData.texelSize).r; + depths4.y = texture(depthImage, uvCoord + vec2(1.0, 0.0) * sceneData.texelSize).r; + depths4.z = texture(depthImage, uvCoord + vec2(0.0, 1.0) * sceneData.texelSize).r; + depths4.w = texture(depthImage, uvCoord + vec2(1.0, 1.0) * sceneData.texelSize).r; + + float depth0 = clampDepth(screenToViewSpaceDepth(depths4.x, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float depth1 = clampDepth(screenToViewSpaceDepth(depths4.y, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float depth2 = clampDepth(screenToViewSpaceDepth(depths4.z, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float depth3 = clampDepth(screenToViewSpaceDepth(depths4.w, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + + imageStore(outDepth0, pixCoord + ivec2(0, 0), vec4(depth0, 0.0f, 0.0f, 0.0f)); + imageStore(outDepth0, pixCoord + ivec2(1, 0), vec4(depth1, 0.0f, 0.0f, 0.0f)); + imageStore(outDepth0, pixCoord + ivec2(0, 1), vec4(depth2, 0.0f, 0.0f, 0.0f)); + imageStore(outDepth0, pixCoord + ivec2(1, 1), vec4(depth3, 0.0f, 0.0f, 0.0f)); } diff --git a/src/core/engine.cpp b/src/core/engine.cpp index 6c567e87..d58a8c6e 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -557,6 +557,14 @@ void Engine::draw(float deltaTime) deferredMrtPipeline->draw(cmd, debugDeferredMrtDrawInfo); } + ambient_occlusion::GTAODrawInfo gtaoDrawInfo{ + camera, + {}, + sceneDataDescriptorBuffer.getDescriptorBufferBindingInfo(), + sceneDataDescriptorBuffer.getDescriptorBufferSize() * FRAME_OVERLAP + }; + ambientOcclusionPipeline->draw(cmd, gtaoDrawInfo); + vk_helpers::transitionImage(cmd, normalRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, albedoRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, pbrRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index bddb0fd7..6d64d039 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -352,14 +352,14 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V GTAOPushConstants push = drawInfo.pushConstants; glm::mat4 projMatrix = drawInfo.camera->getProjMatrix(); - push.depthLinearizeMult = -projMatrix[2][3]; + push.depthLinearizeMult = -projMatrix[3][2]; push.depthLinearizeAdd = projMatrix[2][2]; vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); // Depth Prefilter - { + {; vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipeline); vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); @@ -379,7 +379,6 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V x /= 2; y /= 2; vkCmdDispatch(cmd, x, y, 1); - vkCmdEndRendering(cmd); } vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); From 3a3444a7d731d988d1029dbadcfbf39c7fe20080 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Tue, 25 Mar 2025 22:51:10 +0700 Subject: [PATCH 09/27] GTAO depth prefilter mip 1. --- .../ground_truth/gtao_depth_prefilter.comp | 39 +++++++++++++++++-- .../ambient_occlusion_types.h | 9 +++-- 2 files changed, 42 insertions(+), 6 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index ebe3c57e..2257837f 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -15,13 +15,14 @@ layout (push_constant) uniform PushConstants { float depthLinearizeMult; float depthLinearizeAdd; + float radius; + float radiusMultiplier; + float falloff; + vec2 ndcToViewMult; vec2 ndcToViewAdd; - float radius; - float faloff; float strength; - float radiusMultiplier; int numDirections; int numSteps; @@ -47,6 +48,29 @@ float clampDepth(float depth){ return clamp(depth, 0.0, 65504.0); } +float depthMipFilter(float depth0, float depth1, float depth2, float depth3, float effectRadius, float radiusMultiplier, float falloffRange){ + float maxDepth = max(max(depth0, depth1), max(depth2, depth3)); + + // https://github.com/GameTechDev/XeGTAO/blob/e7698f874e90f2516fca26c696ec3cd2c70e505a/Source/Rendering/Shaders/XeGTAO.hlsli#L583C13-L583C14 + const float depthRangeScaleFactor = 0.75f;// found empirically :) + + + const float _effectRadius = depthRangeScaleFactor * effectRadius * radiusMultiplier; + const float _falloffRange = falloffRange * _effectRadius; + const float falloffFrom = _effectRadius * (1 - falloffRange); + + const float falloffMul = -1.0 / _falloffRange; + const float falloffAdd = falloffFrom / (_falloffRange) + 1.0; + + float weight0 = clamp((maxDepth-depth0) * falloffMul + falloffAdd, 0.0f, 1.0f); + float weight1 = clamp((maxDepth-depth1) * falloffMul + falloffAdd, 0.0f, 1.0f); + float weight2 = clamp((maxDepth-depth2) * falloffMul + falloffAdd, 0.0f, 1.0f); + float weight3 = clamp((maxDepth-depth3) * falloffMul + falloffAdd, 0.0f, 1.0f); + + float weightSum = weight0 + weight1 + weight2 + weight3; + return (weight0 * depth0 + weight1 * depth1 + weight2 * depth2 + weight3 * depth3) / weightSum; +} + void main() { ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); @@ -74,4 +98,13 @@ void main() { imageStore(outDepth0, pixCoord + ivec2(1, 0), vec4(depth1, 0.0f, 0.0f, 0.0f)); imageStore(outDepth0, pixCoord + ivec2(0, 1), vec4(depth2, 0.0f, 0.0f, 0.0f)); imageStore(outDepth0, pixCoord + ivec2(1, 1), vec4(depth3, 0.0f, 0.0f, 0.0f)); + + // MIP 1 + float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + imageStore(outDepth1, ivec2(baseCoord), vec4(dm1)); + + g_scratchDepths[gl_LocalInvocationID.x][gl_LocalInvocationID.y] = dm1; + + memoryBarrierShared(); + barrier(); } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 918fbd39..26589ab1 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -18,14 +18,17 @@ struct GTAOPushConstants float depthLinearizeMult; float depthLinearizeAdd; + // Defaults follow Intel's implementation + float radius = 0.5f; + float falloff = 0.615f; + float radiusMultiplier = 1.457f; + glm::vec2 ndcToViewMult; glm::vec2 ndcToViewAdd; // AO parameters - float radius; - float falloff; float strength; - float radiusMultiplier; + // Sampling parameters uint32_t numDirections; From d068040a3912280d56a33a9ef35f660faeac7dfb Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Wed, 26 Mar 2025 23:21:34 +0700 Subject: [PATCH 10/27] GTAO depth prefilter shader final. --- .../ground_truth/gtao_depth_prefilter.comp | 47 +++++++++++++++++-- .../ground_truth_ambient_occlusion.cpp | 4 +- 2 files changed, 46 insertions(+), 5 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 2257837f..2551c2e7 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -33,9 +33,9 @@ layout (push_constant) uniform PushConstants { } pushConstants; -shared float g_scratchDepths[16][16]; +shared float g_scratchDepths[8][8]; -layout(local_size_x = 16, local_size_y = 16) in; +layout(local_size_x = 8, local_size_y = 8) in; float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float depthLinearizeAdd) { // Optimization by XeGTAO @@ -102,9 +102,50 @@ void main() { // MIP 1 float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); imageStore(outDepth1, ivec2(baseCoord), vec4(dm1)); + g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; - g_scratchDepths[gl_LocalInvocationID.x][gl_LocalInvocationID.y] = dm1; + memoryBarrierShared(); + barrier(); + + // MIP 2 + if (all(equal(groupThreadID.xy % 2u, uvec2(0u)))){ + float inTL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+0]; + float inTR = g_scratchDepths[groupThreadID.x+1][groupThreadID.y+0]; + float inBL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+1]; + float inBR = g_scratchDepths[groupThreadID.x+1][groupThreadID.y+1]; + + float dm2 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + imageStore(outDepth2, ivec2(baseCoord/2u), vec4(dm2)); + g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm2; + } + + memoryBarrierShared(); + barrier(); + + // MIP 3 + if (all(equal(groupThreadID.xy % 4u, uvec2(0u)))){ + float inTL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+0]; + float inTR = g_scratchDepths[groupThreadID.x+2][groupThreadID.y+0]; + float inBL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+2]; + float inBR = g_scratchDepths[groupThreadID.x+2][groupThreadID.y+2]; + + float dm3 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + imageStore(outDepth3, ivec2(baseCoord/4u), vec4(dm3)); + g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm3; + } memoryBarrierShared(); barrier(); + + // MIP 4 + if (all(equal(groupThreadID.xy % 8u, uvec2(0u)))){ + float inTL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+0]; + float inTR = g_scratchDepths[groupThreadID.x+4][groupThreadID.y+0]; + float inBL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+4]; + float inBR = g_scratchDepths[groupThreadID.x+4][groupThreadID.y+4]; + + float dm4 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + imageStore(outDepth4, ivec2(baseCoord/8u), vec4(dm4)); + //g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; + } } diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 6d64d039..6d111e6f 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -373,8 +373,8 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); - auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); - auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 8.0f)); + auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 8.0f)); // divided by 2 because depth prepass operates on 2x2 (still input4 -> output4) x /= 2; y /= 2; From f559a9ad199ccae39854525479f3406aa640de6f Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Thu, 27 Mar 2025 22:27:19 +0700 Subject: [PATCH 11/27] Debugging depth prefilter for GTAO --- .../ground_truth/gtao_depth_prefilter.comp | 49 ++++++++++--------- shaders/deferredResolve.comp | 9 ++-- src/core/engine.cpp | 16 +++--- src/renderer/imgui_wrapper.cpp | 27 ++++++++++ .../ambient_occlusion_types.h | 7 ++- .../ground_truth_ambient_occlusion.cpp | 32 +++++------- .../ground_truth_ambient_occlusion.h | 5 +- .../pipelines/deferred_mrt/deferred_mrt.cpp | 1 - src/renderer/vk_helpers.cpp | 47 ++++++++++++++++++ src/renderer/vk_helpers.h | 4 ++ 10 files changed, 136 insertions(+), 61 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 2551c2e7..ee53f7bf 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -5,31 +5,30 @@ // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl layout (set = 1, binding = 0) uniform sampler2D depthImage; -layout (r16f, set = 1, binding = 1) uniform image2D outDepth0; -layout (r16f, set = 1, binding = 2) uniform image2D outDepth1; -layout (r16f, set = 1, binding = 3) uniform image2D outDepth2; -layout (r16f, set = 1, binding = 4) uniform image2D outDepth3; -layout (r16f, set = 1, binding = 5) uniform image2D outDepth4; +layout (r32f, set = 1, binding = 1) uniform image2D outDepth0; +layout (r32f, set = 1, binding = 2) uniform image2D outDepth1; +layout (r32f, set = 1, binding = 3) uniform image2D outDepth2; +layout (r32f, set = 1, binding = 4) uniform image2D outDepth3; +layout (r32f, set = 1, binding = 5) uniform image2D outDepth4; layout (push_constant) uniform PushConstants { + vec2 ndcToViewMult; + vec2 ndcToViewAdd; + float depthLinearizeMult; float depthLinearizeAdd; float radius; - float radiusMultiplier; float falloff; - - vec2 ndcToViewMult; - vec2 ndcToViewAdd; + float radiusMultiplier; float strength; int numDirections; int numSteps; - float spatialFilterRadius; - float temporalWeight; + float spatialFilterRadius; } pushConstants; @@ -72,17 +71,12 @@ float depthMipFilter(float depth0, float depth1, float depth2, float depth3, flo } void main() { - ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); - - ivec2 dispatchID = ivec2(gl_GlobalInvocationID.xy); ivec2 groupThreadID = ivec2(gl_LocalInvocationID.xy); // MIP 0 - const ivec2 baseCoord = dispatchID; - const ivec2 pixCoord = baseCoord * 2;// We process 2x2 pixels in MIP 0 - - vec2 uvCoord = vec2(pixCoord) * sceneData.texelSize; + const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy) * 2;// We process 2x2 pixels in MIP 0 + vec2 uvCoord = vec2(screenPos) * sceneData.texelSize; vec4 depths4; depths4.x = texture(depthImage, uvCoord + vec2(0.0, 0.0) * sceneData.texelSize).r; depths4.y = texture(depthImage, uvCoord + vec2(1.0, 0.0) * sceneData.texelSize).r; @@ -94,12 +88,19 @@ void main() { float depth2 = clampDepth(screenToViewSpaceDepth(depths4.z, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); float depth3 = clampDepth(screenToViewSpaceDepth(depths4.w, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); - imageStore(outDepth0, pixCoord + ivec2(0, 0), vec4(depth0, 0.0f, 0.0f, 0.0f)); - imageStore(outDepth0, pixCoord + ivec2(1, 0), vec4(depth1, 0.0f, 0.0f, 0.0f)); - imageStore(outDepth0, pixCoord + ivec2(0, 1), vec4(depth2, 0.0f, 0.0f, 0.0f)); - imageStore(outDepth0, pixCoord + ivec2(1, 1), vec4(depth3, 0.0f, 0.0f, 0.0f)); +// imageStore(outDepth0, pixCoord + ivec2(0, 0), vec4(depth0)); +// imageStore(outDepth0, pixCoord + ivec2(1, 0), vec4(depth1)); +// imageStore(outDepth0, pixCoord + ivec2(0, 1), vec4(depth2)); +// imageStore(outDepth0, pixCoord + ivec2(1, 1), vec4(depth3)); + + float test = texture(depthImage, uvCoord).r; + imageStore(outDepth0, screenPos + ivec2(0, 0), vec4(test)); + imageStore(outDepth0, screenPos + ivec2(1, 0), vec4(depths4.y)); + imageStore(outDepth0, screenPos + ivec2(0, 1), vec4(depths4.z)); + imageStore(outDepth0, screenPos + ivec2(1, 1), vec4(depths4.w)); - // MIP 1 + +/* // MIP 1 float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); imageStore(outDepth1, ivec2(baseCoord), vec4(dm1)); g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; @@ -147,5 +148,5 @@ void main() { float dm4 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); imageStore(outDepth4, ivec2(baseCoord/8u), vec4(dm4)); //g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; - } + }*/ } diff --git a/shaders/deferredResolve.comp b/shaders/deferredResolve.comp index 811c53ed..a6de0812 100644 --- a/shaders/deferredResolve.comp +++ b/shaders/deferredResolve.comp @@ -50,10 +50,9 @@ layout (push_constant) uniform PushConstants { float farPlane; } pushConstants; -vec3 reconstructPosition(ivec2 texCoord, float depth) { +vec3 reconstructPosition(ivec2 texCoord, vec2 texelSize, float depth) { // Get normalized device coordinates - vec2 texSize = vec2(pushConstants.width, pushConstants.height); - vec2 ndc = (vec2(texCoord) + 0.5) / texSize * 2.0 - 1.0; + vec2 ndc = (vec2(texCoord) + 0.5) * texelSize * 2.0 - 1.0; // Reconstruct view-space position vec4 positionVS = sceneData.invProjection * vec4(ndc, depth, 1.0); @@ -72,7 +71,7 @@ void main() { return; } - vec2 uv = (vec2(screenPos) + 0.5) / vec2(pushConstants.width, pushConstants.height); + vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; vec4 albedo = texture(albedoRenderTarget, uv); if (albedo.w != 1) { if (pushConstants.debug == 2) { @@ -88,7 +87,7 @@ void main() { float depth = texture(depthBuffer, uv).r; vec3 normal = texture(normalRenderTarget, uv).rgb; vec4 pbrData = texture(pbrRenderTarget, uv); - vec3 position = reconstructPosition(screenPos, depth); + vec3 position = reconstructPosition(screenPos, sceneData.texelSize, depth); float roughness = pbrData.g; float metallic = pbrData.r; diff --git a/src/core/engine.cpp b/src/core/engine.cpp index d58a8c6e..65f61d9a 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -417,6 +417,7 @@ void Engine::updateRender(const float deltaTime, const int32_t currentFrameOverl pDebugSceneData->cameraWorldPos = glm::vec4(0.0f); pDebugSceneData->renderTargetSize = {RENDER_EXTENT_WIDTH, RENDER_EXTENT_HEIGHT}; + pDebugSceneData->texelSize = {1.0f / RENDER_EXTENT_WIDTH, 1.0f / RENDER_EXTENT_HEIGHT}; pDebugSceneData->deltaTime = deltaTime; } @@ -557,20 +558,21 @@ void Engine::draw(float deltaTime) deferredMrtPipeline->draw(cmd, debugDeferredMrtDrawInfo); } + vk_helpers::transitionImage(cmd, normalRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, albedoRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, pbrRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, velocityRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, depthImage.image, VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_DEPTH_BIT); + vk_helpers::transitionImage(cmd, drawImage.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + ambient_occlusion::GTAODrawInfo gtaoDrawInfo{ camera, {}, sceneDataDescriptorBuffer.getDescriptorBufferBindingInfo(), - sceneDataDescriptorBuffer.getDescriptorBufferSize() * FRAME_OVERLAP + sceneDataDescriptorBuffer.getDescriptorBufferSize() * currentFrameOverlap }; ambientOcclusionPipeline->draw(cmd, gtaoDrawInfo); - vk_helpers::transitionImage(cmd, normalRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vk_helpers::transitionImage(cmd, albedoRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vk_helpers::transitionImage(cmd, pbrRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vk_helpers::transitionImage(cmd, velocityRenderTarget.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vk_helpers::transitionImage(cmd, depthImage.image, VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_DEPTH_BIT); - vk_helpers::transitionImage(cmd, drawImage.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); const deferred_resolve::DeferredResolveDrawInfo deferredResolveDrawInfo{ deferredDebug, csmPcf, diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index 4178d9f1..b40098b6 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -12,6 +12,7 @@ #include "environment/environment.h" +#include "lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h" #include "lighting/shadows/cascaded_shadow_map.h" #include "lighting/shadows/shadow_constants.h" #include "src/core/engine.h" @@ -867,6 +868,32 @@ void ImguiWrapper::imguiInterface(Engine* engine) } ImGui::End(); + if (ImGui::Begin("Discardable Debug")) { + if (ImGui::Button("Save GTAO depth image")) { + if (file::getOrCreateDirectory(file::imagesSavePath)) { + const std::filesystem::path path = file::imagesSavePath / "gtao_depth.png"; + + auto depthNormalize = [](const float depth) { + return depth; + }; + + vk_helpers::saveImageR32F( + *engine->resourceManager, + *engine->immediate, + engine->ambientOcclusionPipeline->depthPrefilterImage, + //engine->depthImage, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + VK_IMAGE_ASPECT_COLOR_BIT, + path.string().c_str(), + depthNormalize + ); + } + else { + fmt::print(" Failed to find/create image save path directory"); + } + } + } + ImGui::End(); if (selectedItem) { if (IImguiRenderable* imguiRenderable = dynamic_cast(selectedItem)) { diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 26589ab1..55338e44 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -14,6 +14,9 @@ static constexpr int32_t DEPTH_PREFILTER_MIP_COUNT = 5; struct GTAOPushConstants { + glm::vec2 ndcToViewMult; + glm::vec2 ndcToViewAdd; + // Depth prefilter parameters float depthLinearizeMult; float depthLinearizeAdd; @@ -23,13 +26,9 @@ struct GTAOPushConstants float falloff = 0.615f; float radiusMultiplier = 1.457f; - glm::vec2 ndcToViewMult; - glm::vec2 ndcToViewAdd; - // AO parameters float strength; - // Sampling parameters uint32_t numDirections; uint32_t numSteps; diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 6d111e6f..b6e7fad1 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -46,10 +46,11 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth depthPrefilterDescriptorBuffer = resourceManager.createDescriptorBufferSampler(depthPrefilterSetLayout, 1); - VkImageUsageFlags usage{}; usage |= VK_IMAGE_USAGE_STORAGE_BIT; usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(depthPrefilterFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); // 5 mips, suggested by Intel's implementation @@ -66,15 +67,6 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth VkSamplerCreateInfo samplerInfo = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO}; samplerInfo.magFilter = VK_FILTER_NEAREST; samplerInfo.minFilter = VK_FILTER_NEAREST; - samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST; - samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; - samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; - samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; - samplerInfo.anisotropyEnable = VK_FALSE; - samplerInfo.maxAnisotropy = 1.0f; - samplerInfo.compareEnable = VK_FALSE; - samplerInfo.minLod = 0.0f; - samplerInfo.maxLod = 0.0f; depthPrefilterSampler = resourceManager.createSampler(samplerInfo); } @@ -299,7 +291,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupD imageDescriptors.push_back( { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, - {depthPrefilterSampler, depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + {resourceManager.getDefaultSamplerLinear(), depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, false } ); @@ -356,28 +348,30 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V push.depthLinearizeAdd = projMatrix[2][2]; - vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - + //vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::clearColorImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL); // Depth Prefilter {; vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipeline); - vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); + vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; bindingInfos[0] = drawInfo.sceneDataBinding; bindingInfos[1] = depthPrefilterDescriptorBuffer.getDescriptorBufferBindingInfo(); vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - constexpr std::array indices{0, 1}; - const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; + uint32_t index0 = 0; + uint32_t index1 = 1; + VkDeviceSize offset0 = drawInfo.sceneDataOffset; - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 1, &index0, &offset0); + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 1, 1, &index1, &ZERO_DEVICE_SIZE); auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 8.0f)); auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 8.0f)); // divided by 2 because depth prepass operates on 2x2 (still input4 -> output4) - x /= 2; - y /= 2; + x = x / 2 + 1; + y = y / 2 + 1; vkCmdDispatch(cmd, x, y, 1); } diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index 48fa8840..a9d7c110 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -7,6 +7,7 @@ #include +#include "src/renderer/imgui_wrapper.h" #include "src/renderer/resource_manager.h" #include "src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h" @@ -40,7 +41,7 @@ class GroundTruthAmbientOcclusionPipeline VkSampler depthPrefilterSampler{VK_NULL_HANDLE}; // 16 vs 32. look at cost later. - VkFormat depthPrefilterFormat{VK_FORMAT_R16_SFLOAT}; + VkFormat depthPrefilterFormat{VK_FORMAT_R32_SFLOAT}; AllocatedImage depthPrefilterImage{VK_NULL_HANDLE}; std::array depthPrefilterImageViews{}; @@ -83,6 +84,8 @@ class GroundTruthAmbientOcclusionPipeline private: ResourceManager& resourceManager; + + friend void ImguiWrapper::imguiInterface(Engine* engine); }; } diff --git a/src/renderer/pipelines/deferred_mrt/deferred_mrt.cpp b/src/renderer/pipelines/deferred_mrt/deferred_mrt.cpp index 51540dfe..862c8592 100644 --- a/src/renderer/pipelines/deferred_mrt/deferred_mrt.cpp +++ b/src/renderer/pipelines/deferred_mrt/deferred_mrt.cpp @@ -89,7 +89,6 @@ void will_engine::deferred_mrt::DeferredMrtPipeline::draw(VkCommandBuffer cmd, c scissor.extent.width = RENDER_EXTENTS.width; scissor.extent.height = RENDER_EXTENTS.height; vkCmdSetScissor(cmd, 0, 1, &scissor); - constexpr VkDeviceSize zeroOffset{0}; for (RenderObject* renderObject : drawInfo.renderObjects) { diff --git a/src/renderer/vk_helpers.cpp b/src/renderer/vk_helpers.cpp index 3edacf13..91325ca5 100644 --- a/src/renderer/vk_helpers.cpp +++ b/src/renderer/vk_helpers.cpp @@ -773,6 +773,53 @@ void will_engine::vk_helpers::saveImageR32F(const ResourceManager& resourceManag resourceManager.destroyBuffer(receivingBuffer); } +void will_engine::vk_helpers::saveImageR16F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, + VkImageAspectFlags aspectFlag, const char* savePath, const std::function& valueTransform) +{ + using half_float::half; + const size_t dataSize = image.imageExtent.width * image.imageExtent.height * 1 * sizeof(half); + AllocatedBuffer receivingBuffer = resourceManager.createReceivingBuffer(dataSize); + + immediate.submit([&](VkCommandBuffer cmd) { + VkBufferImageCopy bufferCopyRegion{}; + bufferCopyRegion.imageSubresource.aspectMask = aspectFlag; + bufferCopyRegion.imageSubresource.mipLevel = 0; + bufferCopyRegion.imageSubresource.baseArrayLayer = 0; + bufferCopyRegion.imageSubresource.layerCount = 1; + bufferCopyRegion.imageExtent = image.imageExtent; + bufferCopyRegion.bufferOffset = 0; + bufferCopyRegion.bufferRowLength = 0; + bufferCopyRegion.bufferImageHeight = 0; + + vk_helpers::transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, aspectFlag); + + vkCmdCopyImageToBuffer(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, receivingBuffer.buffer, 1, &bufferCopyRegion); + + vk_helpers::transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, aspectFlag); + }); + + void* data = receivingBuffer.info.pMappedData; + const auto imageData = static_cast(data); + + const auto byteImageData = new uint8_t[image.imageExtent.width * image.imageExtent.height * 4]; + const auto powEight = static_cast(pow(2, 8) - 1); + for (size_t i = 0; i < image.imageExtent.width * image.imageExtent.height; ++i) { + half originalData = imageData[i]; + float floatData = half_float::detail::half2float(originalData); + const float halfValue = valueTransform(originalData); + const auto value = static_cast(halfValue * powEight); + byteImageData[i * 4 + 0] = value; + byteImageData[i * 4 + 1] = value; + byteImageData[i * 4 + 2] = value; + byteImageData[i * 4 + 3] = 255; + } + + stbi_write_png(savePath, image.imageExtent.width, image.imageExtent.height, 4, byteImageData, image.imageExtent.width * 4); + + delete[] byteImageData; + resourceManager.destroyBuffer(receivingBuffer); +} + void will_engine::vk_helpers::saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha) { const auto byteImageData = new uint8_t[width * height * 4]; diff --git a/src/renderer/vk_helpers.h b/src/renderer/vk_helpers.h index 18cde1be..6ceaa419 100644 --- a/src/renderer/vk_helpers.h +++ b/src/renderer/vk_helpers.h @@ -12,6 +12,7 @@ #include #include +#include #include "vk_types.h" @@ -115,6 +116,9 @@ namespace vk_helpers void saveImageR32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, const std::function& valueTransform); + void saveImageR16F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, + const char* savePath, const std::function& valueTransform); + void saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha = true); void saveHeightmap(const std::vector& heightData, int width, int height, const std::filesystem::path& filename); From 01654ca0b0b4e45032ebcc13119b0867f6d3251b Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Fri, 28 Mar 2025 16:04:17 +0700 Subject: [PATCH 12/27] Finalize depth prefilter for GTAO. --- .../ground_truth/gtao_depth_prefilter.comp | 50 +++++++++---------- shaders/environment/environment.vert | 2 +- src/core/engine.cpp | 1 + src/renderer/imgui_wrapper.cpp | 11 +++- .../ground_truth_ambient_occlusion.cpp | 14 ++++-- .../ground_truth_ambient_occlusion.h | 3 +- src/renderer/vk_helpers.cpp | 23 +++++---- src/renderer/vk_helpers.h | 2 +- 8 files changed, 61 insertions(+), 45 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index ee53f7bf..f6f01790 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -1,7 +1,12 @@ #version 460 +#extension GL_EXT_nonuniform_qualifier: enable #include "scene.glsl" +layout(local_size_x = 8, local_size_y = 8) in; + +shared float g_scratchDepths[8][8]; + // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl layout (set = 1, binding = 0) uniform sampler2D depthImage; @@ -31,11 +36,6 @@ layout (push_constant) uniform PushConstants { float spatialFilterRadius; } pushConstants; - -shared float g_scratchDepths[8][8]; - -layout(local_size_x = 8, local_size_y = 8) in; - float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float depthLinearizeAdd) { // Optimization by XeGTAO // https://github.com/GameTechDev/XeGTAO/blob/a5b1686c7ea37788eeb3576b5be47f7c03db532c/Source/Rendering/Shaders/XeGTAO.hlsli#L112 @@ -43,6 +43,7 @@ float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float d } float clampDepth(float depth){ + // kind of redundant, the view space depth can only be as far as the depth buffer (which is 1000.0f at time of writing) // using half float precision return clamp(depth, 0.0, 65504.0); } @@ -74,33 +75,28 @@ void main() { ivec2 groupThreadID = ivec2(gl_LocalInvocationID.xy); // MIP 0 - const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy) * 2;// We process 2x2 pixels in MIP 0 + const uvec2 baseCoord = gl_GlobalInvocationID.xy; + const ivec2 screenPos = ivec2(baseCoord.xy) * 2;// We process 2x2 pixels in MIP 0 - vec2 uvCoord = vec2(screenPos) * sceneData.texelSize; - vec4 depths4; - depths4.x = texture(depthImage, uvCoord + vec2(0.0, 0.0) * sceneData.texelSize).r; - depths4.y = texture(depthImage, uvCoord + vec2(1.0, 0.0) * sceneData.texelSize).r; - depths4.z = texture(depthImage, uvCoord + vec2(0.0, 1.0) * sceneData.texelSize).r; - depths4.w = texture(depthImage, uvCoord + vec2(1.0, 1.0) * sceneData.texelSize).r; + vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - float depth0 = clampDepth(screenToViewSpaceDepth(depths4.x, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); - float depth1 = clampDepth(screenToViewSpaceDepth(depths4.y, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); - float depth2 = clampDepth(screenToViewSpaceDepth(depths4.z, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); - float depth3 = clampDepth(screenToViewSpaceDepth(depths4.w, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float rDepth0 = texture(depthImage, uv + vec2(0.0, 0.0) * sceneData.texelSize).r; + float rDepth1 = texture(depthImage, uv + vec2(1.0, 0.0) * sceneData.texelSize).r; + float rDepth2 = texture(depthImage, uv + vec2(0.0, 1.0) * sceneData.texelSize).r; + float rDepth3 = texture(depthImage, uv + vec2(1.0, 1.0) * sceneData.texelSize).r; -// imageStore(outDepth0, pixCoord + ivec2(0, 0), vec4(depth0)); -// imageStore(outDepth0, pixCoord + ivec2(1, 0), vec4(depth1)); -// imageStore(outDepth0, pixCoord + ivec2(0, 1), vec4(depth2)); -// imageStore(outDepth0, pixCoord + ivec2(1, 1), vec4(depth3)); + float depth0 = clampDepth(screenToViewSpaceDepth(rDepth0, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float depth1 = clampDepth(screenToViewSpaceDepth(rDepth1, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float depth2 = clampDepth(screenToViewSpaceDepth(rDepth2, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); + float depth3 = clampDepth(screenToViewSpaceDepth(rDepth3, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); - float test = texture(depthImage, uvCoord).r; - imageStore(outDepth0, screenPos + ivec2(0, 0), vec4(test)); - imageStore(outDepth0, screenPos + ivec2(1, 0), vec4(depths4.y)); - imageStore(outDepth0, screenPos + ivec2(0, 1), vec4(depths4.z)); - imageStore(outDepth0, screenPos + ivec2(1, 1), vec4(depths4.w)); + imageStore(outDepth0, screenPos + ivec2(0, 0), vec4(depth0)); + imageStore(outDepth0, screenPos + ivec2(1, 0), vec4(depth1)); + imageStore(outDepth0, screenPos + ivec2(0, 1), vec4(depth2)); + imageStore(outDepth0, screenPos + ivec2(1, 1), vec4(depth3)); -/* // MIP 1 + // MIP 1 float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); imageStore(outDepth1, ivec2(baseCoord), vec4(dm1)); g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; @@ -148,5 +144,5 @@ void main() { float dm4 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); imageStore(outDepth4, ivec2(baseCoord/8u), vec4(dm4)); //g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; - }*/ + } } diff --git a/shaders/environment/environment.vert b/shaders/environment/environment.vert index af66d555..6ea6da69 100644 --- a/shaders/environment/environment.vert +++ b/shaders/environment/environment.vert @@ -9,7 +9,7 @@ layout (location = 2) out vec4 outPrevMvpPosition; // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl void main() { - const vec3 vertices[3] = vec3[3](vec3(-1, -1, 0.00001), vec3(3, -1, 0.00001), vec3(-1, 3, 0.00001)); + const vec3 vertices[3] = vec3[3](vec3(-1, -1, 0.0002), vec3(3, -1, 0.0002), vec3(-1, 3, 0.0002)); vec4 currClipPos = vec4(vertices[gl_VertexIndex], 1); diff --git a/src/core/engine.cpp b/src/core/engine.cpp index 65f61d9a..fc8ca46b 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -955,6 +955,7 @@ void Engine::hotReloadShaders() const environmentPipeline->reloadShaders(); terrainPipeline->reloadShaders(); deferredMrtPipeline->reloadShaders(); + ambientOcclusionPipeline->reloadShaders(); deferredResolvePipeline->reloadShaders(); temporalAntialiasingPipeline->reloadShaders(); postProcessPipeline->reloadShaders(); diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index b40098b6..421a7cac 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -869,12 +869,18 @@ void ImguiWrapper::imguiInterface(Engine* engine) ImGui::End(); if (ImGui::Begin("Discardable Debug")) { + static int32_t gtaoMip; + constexpr uint32_t minMip = 0; + constexpr uint32_t maxMip = 4; + + ImGui::SliderScalar("GTAO level", ImGuiDataType_S32, >aoMip, &minMip, &maxMip); + if (ImGui::Button("Save GTAO depth image")) { if (file::getOrCreateDirectory(file::imagesSavePath)) { const std::filesystem::path path = file::imagesSavePath / "gtao_depth.png"; auto depthNormalize = [](const float depth) { - return depth; + return depth / 1000.f; }; vk_helpers::saveImageR32F( @@ -885,7 +891,8 @@ void ImguiWrapper::imguiInterface(Engine* engine) VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), - depthNormalize + depthNormalize, + gtaoMip ); } else { diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index b6e7fad1..fc22b271 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -283,7 +283,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTrut resourceManager.destroyDescriptorBuffer(temporalAccumulationDescriptorBuffer); } -void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(const VkImageView depthImageView) +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(const VkImageView& depthImageView) { std::vector imageDescriptors{}; imageDescriptors.reserve(1 + DEPTH_PREFILTER_MIP_COUNT); @@ -370,8 +370,8 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 8.0f)); auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 8.0f)); // divided by 2 because depth prepass operates on 2x2 (still input4 -> output4) - x = x / 2 + 1; - y = y / 2 + 1; + x = x / 2; + y = y / 2; vkCmdDispatch(cmd, x, y, 1); } @@ -408,6 +408,14 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vkCmdEndDebugUtilsLabelEXT(cmd); } +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::reloadShaders() +{ + createDepthPrefilterPipeline(); + createAmbientOcclusionPipeline(); + createSpatialFilteringPipeline(); + createTemporalAccumulationPipeline(); +} + void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createDepthPrefilterPipeline() { resourceManager.destroyPipeline(depthPrefilterPipeline); diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index a9d7c110..f45c6a72 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -21,12 +21,13 @@ class GroundTruthAmbientOcclusionPipeline ~GroundTruthAmbientOcclusionPipeline(); - void setupDepthPrefilterDescriptorBuffer(VkImageView depthImageView); + void setupDepthPrefilterDescriptorBuffer(const VkImageView& depthImageView); void setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView); void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const; + void reloadShaders(); private: void createDepthPrefilterPipeline(); void createAmbientOcclusionPipeline(); diff --git a/src/renderer/vk_helpers.cpp b/src/renderer/vk_helpers.cpp index 91325ca5..73e5e9e0 100644 --- a/src/renderer/vk_helpers.cpp +++ b/src/renderer/vk_helpers.cpp @@ -730,35 +730,38 @@ void will_engine::vk_helpers::savePacked64Bit(const ResourceManager& resourceMan } void will_engine::vk_helpers::saveImageR32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, const std::function& valueTransform) + const char* savePath, const std::function& valueTransform, int32_t mipLevel) { - const size_t dataSize = image.imageExtent.width * image.imageExtent.height * 1 * sizeof(float); + size_t newXSize = image.imageExtent.width / static_cast(std::pow(2, mipLevel)); + size_t newYSize = image.imageExtent.height / static_cast(std::pow(2, mipLevel)); + const size_t texelCount = newXSize * newYSize; + const size_t dataSize = texelCount * 1 * sizeof(float); AllocatedBuffer receivingBuffer = resourceManager.createReceivingBuffer(dataSize); - immediate.submit([&](VkCommandBuffer cmd) { + immediate.submit([&, mipLevel](VkCommandBuffer cmd) { VkBufferImageCopy bufferCopyRegion{}; bufferCopyRegion.imageSubresource.aspectMask = aspectFlag; - bufferCopyRegion.imageSubresource.mipLevel = 0; + bufferCopyRegion.imageSubresource.mipLevel = mipLevel; bufferCopyRegion.imageSubresource.baseArrayLayer = 0; bufferCopyRegion.imageSubresource.layerCount = 1; - bufferCopyRegion.imageExtent = image.imageExtent; + bufferCopyRegion.imageExtent = {static_cast(newXSize), static_cast(newYSize), 1u}; bufferCopyRegion.bufferOffset = 0; bufferCopyRegion.bufferRowLength = 0; bufferCopyRegion.bufferImageHeight = 0; - vk_helpers::transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, aspectFlag); + transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, aspectFlag); vkCmdCopyImageToBuffer(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, receivingBuffer.buffer, 1, &bufferCopyRegion); - vk_helpers::transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, aspectFlag); + transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, aspectFlag); }); void* data = receivingBuffer.info.pMappedData; const auto imageData = static_cast(data); - const auto byteImageData = new uint8_t[image.imageExtent.width * image.imageExtent.height * 4]; + const auto byteImageData = new uint8_t[texelCount * 4]; const auto powEight = static_cast(pow(2, 8) - 1); - for (size_t i = 0; i < image.imageExtent.width * image.imageExtent.height; ++i) { + for (size_t i = 0; i < texelCount; ++i) { const float floatValue = valueTransform(imageData[i]); const auto value = static_cast(floatValue * powEight); byteImageData[i * 4 + 0] = value; @@ -767,7 +770,7 @@ void will_engine::vk_helpers::saveImageR32F(const ResourceManager& resourceManag byteImageData[i * 4 + 3] = 255; } - stbi_write_png(savePath, image.imageExtent.width, image.imageExtent.height, 4, byteImageData, image.imageExtent.width * 4); + stbi_write_png(savePath, static_cast(newXSize), static_cast(newYSize), 4, byteImageData, static_cast(newXSize) * 4); delete[] byteImageData; resourceManager.destroyBuffer(receivingBuffer); diff --git a/src/renderer/vk_helpers.h b/src/renderer/vk_helpers.h index 6ceaa419..3a5358ce 100644 --- a/src/renderer/vk_helpers.h +++ b/src/renderer/vk_helpers.h @@ -114,7 +114,7 @@ namespace vk_helpers * Save the Allocated image as a grayscaled image. The image must be a format with only 1 channel (e.g. R32 or D32) */ void saveImageR32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, const std::function& valueTransform); + const char* savePath, const std::function& valueTransform, int32_t mipLevel = 0); void saveImageR16F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, const std::function& valueTransform); From dea858e2bd997f2078bb35f0d568a6f865af0c4c Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Fri, 28 Mar 2025 17:18:01 +0700 Subject: [PATCH 13/27] removed half library --- .gitmodules | 3 --- extern/half | 1 - 2 files changed, 4 deletions(-) delete mode 160000 extern/half diff --git a/.gitmodules b/.gitmodules index e3987ac8..51bf7966 100644 --- a/.gitmodules +++ b/.gitmodules @@ -7,6 +7,3 @@ [submodule "extern/JoltPhysics"] path = extern/JoltPhysics url = https://github.com/jrouwe/JoltPhysics.git -[submodule "extern/half"] - path = extern/half - url = https://github.com/melowntech/half.git diff --git a/extern/half b/extern/half deleted file mode 160000 index 972e0409..00000000 --- a/extern/half +++ /dev/null @@ -1 +0,0 @@ -Subproject commit 972e040989383764f30c1d45a9b666593e802741 From b695208824599f1e79e8ec2bc8c38a819ccd430a Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Fri, 28 Mar 2025 18:01:46 +0700 Subject: [PATCH 14/27] change depth prefilter format to FP16. --- CMakeLists.txt | 7 - extern/half/half.hpp | 4603 +++++++++++++++++ .../ground_truth/gtao_depth_prefilter.comp | 10 +- src/renderer/imgui_wrapper.cpp | 36 +- .../ground_truth_ambient_occlusion.cpp | 6 +- .../ground_truth_ambient_occlusion.h | 2 +- src/renderer/vk_helpers.cpp | 36 +- src/renderer/vk_helpers.h | 39 +- src/util/math_utils.h | 26 + 9 files changed, 4702 insertions(+), 63 deletions(-) create mode 100644 extern/half/half.hpp diff --git a/CMakeLists.txt b/CMakeLists.txt index 01c37f27..6ea9f6a5 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -62,12 +62,6 @@ set(FASTGLTF_SOURCES ${FASTGLTF_DIR}/src/io.cpp ) -set(HALF_DIR ${CMAKE_CURRENT_SOURCE_DIR}/extern/half) -set(HALF_SOURCES - ${HALF_DIR}/half/dummy.cpp - ${HALF_DIR}/half/half.hpp -) - set(VK_BOOTSTRAP_SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/extern/vk-bootstrap/VkBootstrap.cpp ) @@ -256,7 +250,6 @@ set(TEMP_SOURCES add_executable(WillEngine main.cpp ${IMGUI_SOURCES} ${FASTGLTF_SOURCES} - ${HALF_SOURCES} ${VK_BOOTSTRAP_SOURCES} ${VOLK_SOURCES} ${ENGINE_SOURCES} diff --git a/extern/half/half.hpp b/extern/half/half.hpp new file mode 100644 index 00000000..cb658f22 --- /dev/null +++ b/extern/half/half.hpp @@ -0,0 +1,4603 @@ +// half - IEEE 754-based half-precision floating-point library. +// +// Copyright (c) 2012-2025 Christian Rau +// +// Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation +// files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, +// modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the +// Software is furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE +// WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR +// COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + +// Version 2.2.1 + +/// \file +/// Main header file for half-precision functionality. + +#ifndef HALF_HALF_HPP +#define HALF_HALF_HPP + +#define HALF_GCC_VERSION (__GNUC__*100+__GNUC_MINOR__) + +#if defined(__INTEL_COMPILER) + #define HALF_ICC_VERSION __INTEL_COMPILER +#elif defined(__ICC) + #define HALF_ICC_VERSION __ICC +#elif defined(__ICL) + #define HALF_ICC_VERSION __ICL +#else + #define HALF_ICC_VERSION 0 +#endif + +// check C++11 language features +#if defined(__clang__) // clang + #if __has_feature(cxx_static_assert) && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT) + #define HALF_ENABLE_CPP11_STATIC_ASSERT 1 + #endif + #if __has_feature(cxx_constexpr) && !defined(HALF_ENABLE_CPP11_CONSTEXPR) + #define HALF_ENABLE_CPP11_CONSTEXPR 1 + #endif + #if __has_feature(cxx_noexcept) && !defined(HALF_ENABLE_CPP11_NOEXCEPT) + #define HALF_ENABLE_CPP11_NOEXCEPT 1 + #endif + #if __has_feature(cxx_user_literals) && !defined(HALF_ENABLE_CPP11_USER_LITERALS) + #define HALF_ENABLE_CPP11_USER_LITERALS 1 + #endif + #if __has_feature(cxx_thread_local) && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL) + #define HALF_ENABLE_CPP11_THREAD_LOCAL 1 + #endif + #if (defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L) && !defined(HALF_ENABLE_CPP11_LONG_LONG) + #define HALF_ENABLE_CPP11_LONG_LONG 1 + #endif +#elif HALF_ICC_VERSION && defined(__INTEL_CXX11_MODE__) // Intel C++ + #if HALF_ICC_VERSION >= 1500 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL) + #define HALF_ENABLE_CPP11_THREAD_LOCAL 1 + #endif + #if HALF_ICC_VERSION >= 1500 && !defined(HALF_ENABLE_CPP11_USER_LITERALS) + #define HALF_ENABLE_CPP11_USER_LITERALS 1 + #endif + #if HALF_ICC_VERSION >= 1400 && !defined(HALF_ENABLE_CPP11_CONSTEXPR) + #define HALF_ENABLE_CPP11_CONSTEXPR 1 + #endif + #if HALF_ICC_VERSION >= 1400 && !defined(HALF_ENABLE_CPP11_NOEXCEPT) + #define HALF_ENABLE_CPP11_NOEXCEPT 1 + #endif + #if HALF_ICC_VERSION >= 1110 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT) + #define HALF_ENABLE_CPP11_STATIC_ASSERT 1 + #endif + #if HALF_ICC_VERSION >= 1110 && !defined(HALF_ENABLE_CPP11_LONG_LONG) + #define HALF_ENABLE_CPP11_LONG_LONG 1 + #endif +#elif defined(__GNUC__) // gcc + #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103L + #if HALF_GCC_VERSION >= 408 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL) + #define HALF_ENABLE_CPP11_THREAD_LOCAL 1 + #endif + #if HALF_GCC_VERSION >= 407 && !defined(HALF_ENABLE_CPP11_USER_LITERALS) + #define HALF_ENABLE_CPP11_USER_LITERALS 1 + #endif + #if HALF_GCC_VERSION >= 406 && !defined(HALF_ENABLE_CPP11_CONSTEXPR) + #define HALF_ENABLE_CPP11_CONSTEXPR 1 + #endif + #if HALF_GCC_VERSION >= 406 && !defined(HALF_ENABLE_CPP11_NOEXCEPT) + #define HALF_ENABLE_CPP11_NOEXCEPT 1 + #endif + #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT) + #define HALF_ENABLE_CPP11_STATIC_ASSERT 1 + #endif + #if !defined(HALF_ENABLE_CPP11_LONG_LONG) + #define HALF_ENABLE_CPP11_LONG_LONG 1 + #endif + #endif + #define HALF_TWOS_COMPLEMENT_INT 1 +#elif defined(_MSC_VER) // Visual C++ + #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_THREAD_LOCAL) + #define HALF_ENABLE_CPP11_THREAD_LOCAL 1 + #endif + #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_USER_LITERALS) + #define HALF_ENABLE_CPP11_USER_LITERALS 1 + #endif + #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_CONSTEXPR) + #define HALF_ENABLE_CPP11_CONSTEXPR 1 + #endif + #if _MSC_VER >= 1900 && !defined(HALF_ENABLE_CPP11_NOEXCEPT) + #define HALF_ENABLE_CPP11_NOEXCEPT 1 + #endif + #if _MSC_VER >= 1600 && !defined(HALF_ENABLE_CPP11_STATIC_ASSERT) + #define HALF_ENABLE_CPP11_STATIC_ASSERT 1 + #endif + #if _MSC_VER >= 1310 && !defined(HALF_ENABLE_CPP11_LONG_LONG) + #define HALF_ENABLE_CPP11_LONG_LONG 1 + #endif + #define HALF_TWOS_COMPLEMENT_INT 1 + #define HALF_POP_WARNINGS 1 + #pragma warning(push) + #pragma warning(disable : 4099 4127 4146) //struct vs class, constant in if, negative unsigned +#endif + +// check C++11 library features +#include +#if defined(_LIBCPP_VERSION) // libc++ + #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103 + #ifndef HALF_ENABLE_CPP11_TYPE_TRAITS + #define HALF_ENABLE_CPP11_TYPE_TRAITS 1 + #endif + #ifndef HALF_ENABLE_CPP11_CSTDINT + #define HALF_ENABLE_CPP11_CSTDINT 1 + #endif + #ifndef HALF_ENABLE_CPP11_CMATH + #define HALF_ENABLE_CPP11_CMATH 1 + #endif + #ifndef HALF_ENABLE_CPP11_HASH + #define HALF_ENABLE_CPP11_HASH 1 + #endif + #ifndef HALF_ENABLE_CPP11_CFENV + #define HALF_ENABLE_CPP11_CFENV 1 + #endif + #endif +#elif defined(__GLIBCXX__) // libstdc++ + #if defined(__GXX_EXPERIMENTAL_CXX0X__) || __cplusplus >= 201103 + #ifdef __clang__ + #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS) + #define HALF_ENABLE_CPP11_TYPE_TRAITS 1 + #endif + #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CSTDINT) + #define HALF_ENABLE_CPP11_CSTDINT 1 + #endif + #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CMATH) + #define HALF_ENABLE_CPP11_CMATH 1 + #endif + #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_HASH) + #define HALF_ENABLE_CPP11_HASH 1 + #endif + #if __GLIBCXX__ >= 20080606 && !defined(HALF_ENABLE_CPP11_CFENV) + #define HALF_ENABLE_CPP11_CFENV 1 + #endif + #else + #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS) + #define HALF_ENABLE_CPP11_TYPE_TRAITS 1 + #endif + #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CSTDINT) + #define HALF_ENABLE_CPP11_CSTDINT 1 + #endif + #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CMATH) + #define HALF_ENABLE_CPP11_CMATH 1 + #endif + #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_HASH) + #define HALF_ENABLE_CPP11_HASH 1 + #endif + #if HALF_GCC_VERSION >= 403 && !defined(HALF_ENABLE_CPP11_CFENV) + #define HALF_ENABLE_CPP11_CFENV 1 + #endif + #endif + #endif +#elif defined(_CPPLIB_VER) // Dinkumware/Visual C++ + #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_TYPE_TRAITS) + #define HALF_ENABLE_CPP11_TYPE_TRAITS 1 + #endif + #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_CSTDINT) + #define HALF_ENABLE_CPP11_CSTDINT 1 + #endif + #if _CPPLIB_VER >= 520 && !defined(HALF_ENABLE_CPP11_HASH) + #define HALF_ENABLE_CPP11_HASH 1 + #endif + #if _CPPLIB_VER >= 610 && !defined(HALF_ENABLE_CPP11_CMATH) + #define HALF_ENABLE_CPP11_CMATH 1 + #endif + #if _CPPLIB_VER >= 610 && !defined(HALF_ENABLE_CPP11_CFENV) + #define HALF_ENABLE_CPP11_CFENV 1 + #endif +#endif +#undef HALF_GCC_VERSION +#undef HALF_ICC_VERSION + +// any error throwing C++ exceptions? +#if defined(HALF_ERRHANDLING_THROW_INVALID) || defined(HALF_ERRHANDLING_THROW_DIVBYZERO) || defined(HALF_ERRHANDLING_THROW_OVERFLOW) || defined(HALF_ERRHANDLING_THROW_UNDERFLOW) || defined(HALF_ERRHANDLING_THROW_INEXACT) +#define HALF_ERRHANDLING_THROWS 1 +#endif + +// any error handling enabled? +#define HALF_ERRHANDLING (HALF_ERRHANDLING_FLAGS||HALF_ERRHANDLING_ERRNO||HALF_ERRHANDLING_FENV||HALF_ERRHANDLING_THROWS) + +#if HALF_ERRHANDLING + #define HALF_UNUSED_NOERR(name) name +#else + #define HALF_UNUSED_NOERR(name) +#endif + +// support constexpr +#if HALF_ENABLE_CPP11_CONSTEXPR + #define HALF_CONSTEXPR constexpr + #define HALF_CONSTEXPR_CONST constexpr + #if HALF_ERRHANDLING + #define HALF_CONSTEXPR_NOERR + #else + #define HALF_CONSTEXPR_NOERR constexpr + #endif +#else + #define HALF_CONSTEXPR + #define HALF_CONSTEXPR_CONST const + #define HALF_CONSTEXPR_NOERR +#endif + +// support noexcept +#if HALF_ENABLE_CPP11_NOEXCEPT + #define HALF_NOEXCEPT noexcept + #define HALF_NOTHROW noexcept +#else + #define HALF_NOEXCEPT + #define HALF_NOTHROW throw() +#endif + +// support thread storage +#if HALF_ENABLE_CPP11_THREAD_LOCAL + #define HALF_THREAD_LOCAL thread_local +#else + #define HALF_THREAD_LOCAL static +#endif + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#if HALF_ENABLE_CPP11_TYPE_TRAITS + #include +#endif +#if HALF_ENABLE_CPP11_CSTDINT + #include +#endif +#if HALF_ERRHANDLING_ERRNO + #include +#endif +#if HALF_ENABLE_CPP11_CFENV + #include +#endif +#if HALF_ENABLE_CPP11_HASH + #include +#endif + + +#ifndef HALF_ENABLE_F16C_INTRINSICS + /// Enable F16C intruction set intrinsics. + /// Defining this to 1 enables the use of [F16C compiler intrinsics](https://en.wikipedia.org/wiki/F16C) for converting between + /// half-precision and single-precision values which may result in improved performance. This will not perform additional checks + /// for support of the F16C instruction set, so an appropriate target platform is required when enabling this feature. + /// + /// Unless predefined it will be enabled automatically when the `__F16C__` symbol is defined, which some compilers do on supporting platforms. + #define HALF_ENABLE_F16C_INTRINSICS __F16C__ +#endif +#if HALF_ENABLE_F16C_INTRINSICS + #include +#endif + +#ifdef HALF_DOXYGEN_ONLY +/// Type for internal floating-point computations. +/// This can be predefined to a built-in floating-point type (`float`, `double` or `long double`) to override the internal +/// half-precision implementation to use this type for computing arithmetic operations and mathematical functions (if available). +/// This can result in improved performance for arithmetic operators and mathematical functions but might cause results to +/// deviate from the specified half-precision rounding mode and inhibits proper detection of half-precision exceptions. +#define HALF_ARITHMETIC_TYPE (undefined) + +/// Enable internal exception flags. +/// Defining this to 1 causes operations on half-precision values to raise internal floating-point exception flags according to +/// the IEEE 754 standard. These can then be cleared and checked with clearexcept(), testexcept(). +#define HALF_ERRHANDLING_FLAGS 0 + +/// Enable exception propagation to `errno`. +/// Defining this to 1 causes operations on half-precision values to propagate floating-point exceptions to +/// [errno](https://en.cppreference.com/w/cpp/error/errno) from ``. Specifically this will propagate domain errors as +/// [EDOM](https://en.cppreference.com/w/cpp/error/errno_macros) and pole, overflow and underflow errors as +/// [ERANGE](https://en.cppreference.com/w/cpp/error/errno_macros). Inexact errors won't be propagated. +#define HALF_ERRHANDLING_ERRNO 0 + +/// Enable exception propagation to built-in floating-point platform. +/// Defining this to 1 causes operations on half-precision values to propagate floating-point exceptions to the built-in +/// single- and double-precision implementation's exception flags using the +/// [C++11 floating-point environment control](https://en.cppreference.com/w/cpp/numeric/fenv) from ``. However, this +/// does not work in reverse and single- or double-precision exceptions will not raise the corresponding half-precision +/// exception flags, nor will explicitly clearing flags clear the corresponding built-in flags. +#define HALF_ERRHANDLING_FENV 0 + +/// Throw C++ exception on domain errors. +/// Defining this to a string literal causes operations on half-precision values to throw a +/// [std::domain_error](https://en.cppreference.com/w/cpp/error/domain_error) with the specified message on domain errors. +#define HALF_ERRHANDLING_THROW_INVALID (undefined) + +/// Throw C++ exception on pole errors. +/// Defining this to a string literal causes operations on half-precision values to throw a +/// [std::domain_error](https://en.cppreference.com/w/cpp/error/domain_error) with the specified message on pole errors. +#define HALF_ERRHANDLING_THROW_DIVBYZERO (undefined) + +/// Throw C++ exception on overflow errors. +/// Defining this to a string literal causes operations on half-precision values to throw a +/// [std::overflow_error](https://en.cppreference.com/w/cpp/error/overflow_error) with the specified message on overflows. +#define HALF_ERRHANDLING_THROW_OVERFLOW (undefined) + +/// Throw C++ exception on underflow errors. +/// Defining this to a string literal causes operations on half-precision values to throw a +/// [std::underflow_error](https://en.cppreference.com/w/cpp/error/underflow_error) with the specified message on underflows. +#define HALF_ERRHANDLING_THROW_UNDERFLOW (undefined) + +/// Throw C++ exception on rounding errors. +/// Defining this to 1 causes operations on half-precision values to throw a +/// [std::range_error](https://en.cppreference.com/w/cpp/error/range_error) with the specified message on general rounding errors. +#define HALF_ERRHANDLING_THROW_INEXACT (undefined) +#endif + +#ifndef HALF_ERRHANDLING_OVERFLOW_TO_INEXACT +/// Raise INEXACT exception on overflow. +/// Defining this to 1 (default) causes overflow errors to also raise inexact exceptions. +/// These will be raised after any possible handling of the underflow exception. +#define HALF_ERRHANDLING_OVERFLOW_TO_INEXACT 1 +#endif + +#ifndef HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT +/// Raise INEXACT exception on underflow. +/// Defining this to 1 (default) causes underflow errors to also raise inexact exceptions. +/// These will be raised after any possible handling of the underflow exception. +/// +/// **Note:** This will actually cause underflow (and the accompanying inexact) exceptions to be raised *only* when the result +/// is inexact, while if disabled bare underflow errors will be raised for *any* (possibly exact) subnormal result. +#define HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT 1 +#endif + +/// Default rounding mode. +/// This specifies the rounding mode used for all conversions between [half](\ref half_float::half)s and more precise types +/// (unless using half_cast() and specifying the rounding mode directly) as well as in arithmetic operations and mathematical +/// functions. It can be redefined (before including half.hpp) to one of the standard rounding modes using their respective +/// constants or the equivalent values of +/// [std::float_round_style](https://en.cppreference.com/w/cpp/types/numeric_limits/float_round_style): +/// +/// `std::float_round_style` | value | rounding +/// ---------------------------------|-------|------------------------- +/// `std::round_indeterminate` | -1 | fastest +/// `std::round_toward_zero` | 0 | toward zero +/// `std::round_to_nearest` | 1 | to nearest (default) +/// `std::round_toward_infinity` | 2 | toward positive infinity +/// `std::round_toward_neg_infinity` | 3 | toward negative infinity +/// +/// By default this is set to `1` (`std::round_to_nearest`), which rounds results to the nearest representable value. It can even +/// be set to [std::numeric_limits::round_style](https://en.cppreference.com/w/cpp/types/numeric_limits/round_style) to synchronize +/// the rounding mode with that of the built-in single-precision implementation (which is likely `std::round_to_nearest`, though). +#ifndef HALF_ROUND_STYLE + #define HALF_ROUND_STYLE 1 // = std::round_to_nearest +#endif + +/// Value signaling overflow. +/// In correspondence with `HUGE_VAL[F|L]` from `` this symbol expands to a positive value signaling the overflow of an +/// operation, in particular it just evaluates to positive infinity. +/// +/// **See also:** Documentation for [HUGE_VAL](https://en.cppreference.com/w/cpp/numeric/math/HUGE_VAL) +#define HUGE_VALH std::numeric_limits::infinity() + +/// Fast half-precision fma function. +/// This symbol is defined if the fma() function generally executes as fast as, or faster than, a separate +/// half-precision multiplication followed by an addition, which is always the case. +/// +/// **See also:** Documentation for [FP_FAST_FMA](https://en.cppreference.com/w/cpp/numeric/math/fma) +#define FP_FAST_FMAH 1 + +/// Half rounding mode. +/// In correspondence with `FLT_ROUNDS` from `` this symbol expands to the rounding mode used for +/// half-precision operations. It is an alias for [HALF_ROUND_STYLE](\ref HALF_ROUND_STYLE). +/// +/// **See also:** Documentation for [FLT_ROUNDS](https://en.cppreference.com/w/cpp/types/climits/FLT_ROUNDS) +#define HLF_ROUNDS HALF_ROUND_STYLE + +#ifndef FP_ILOGB0 + #define FP_ILOGB0 INT_MIN +#endif +#ifndef FP_ILOGBNAN + #define FP_ILOGBNAN INT_MAX +#endif +#ifndef FP_SUBNORMAL + #define FP_SUBNORMAL 0 +#endif +#ifndef FP_ZERO + #define FP_ZERO 1 +#endif +#ifndef FP_NAN + #define FP_NAN 2 +#endif +#ifndef FP_INFINITE + #define FP_INFINITE 3 +#endif +#ifndef FP_NORMAL + #define FP_NORMAL 4 +#endif + +#if !HALF_ENABLE_CPP11_CFENV && !defined(FE_ALL_EXCEPT) + #define FE_INVALID 0x10 + #define FE_DIVBYZERO 0x08 + #define FE_OVERFLOW 0x04 + #define FE_UNDERFLOW 0x02 + #define FE_INEXACT 0x01 + #define FE_ALL_EXCEPT (FE_INVALID|FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW|FE_INEXACT) +#endif + + +/// Main namespace for half-precision functionality. +/// This namespace contains all the functionality provided by the library. +namespace half_float +{ + class half; + +#if HALF_ENABLE_CPP11_USER_LITERALS + /// Library-defined half-precision literals. + /// Import this namespace to enable half-precision floating-point literals: + /// ~~~~{.cpp} + /// using namespace half_float::literal; + /// half_float::half = 4.2_h; + /// ~~~~ + namespace literal + { + half operator "" _h(long double); + } +#endif + + /// \internal + /// \brief Implementation details. + namespace detail + { + #if HALF_ENABLE_CPP11_TYPE_TRAITS + /// Conditional type. + template struct conditional : std::conditional {}; + + /// Helper for tag dispatching. + template struct bool_type : std::integral_constant {}; + using std::true_type; + using std::false_type; + + /// Type traits for floating-point types. + template struct is_float : std::is_floating_point {}; + #else + /// Conditional type. + template struct conditional { typedef T type; }; + template struct conditional { typedef F type; }; + + /// Helper for tag dispatching. + template struct bool_type {}; + typedef bool_type true_type; + typedef bool_type false_type; + + /// Type traits for floating-point types. + template struct is_float : false_type {}; + template struct is_float : is_float {}; + template struct is_float : is_float {}; + template struct is_float : is_float {}; + template<> struct is_float : true_type {}; + template<> struct is_float : true_type {}; + template<> struct is_float : true_type {}; + #endif + + /// Type traits for floating-point bits. + template struct bits { typedef unsigned char type; }; + template struct bits : bits {}; + template struct bits : bits {}; + template struct bits : bits {}; + + #if HALF_ENABLE_CPP11_CSTDINT + /// Unsigned integer of (at least) 16 bits width. + typedef std::uint_least16_t uint16; + + /// Fastest unsigned integer of (at least) 32 bits width. + typedef std::uint_fast32_t uint32; + + /// Fastest signed integer of (at least) 32 bits width. + typedef std::int_fast32_t int32; + + /// Unsigned integer of (at least) 32 bits width. + template<> struct bits { typedef std::uint_least32_t type; }; + + /// Unsigned integer of (at least) 64 bits width. + template<> struct bits { typedef std::uint_least64_t type; }; + #else + /// Unsigned integer of (at least) 16 bits width. + typedef unsigned short uint16; + + /// Fastest unsigned integer of (at least) 32 bits width. + typedef unsigned long uint32; + + /// Fastest unsigned integer of (at least) 32 bits width. + typedef long int32; + + /// Unsigned integer of (at least) 32 bits width. + template<> struct bits : conditional::digits>=32,unsigned int,unsigned long> {}; + + #if HALF_ENABLE_CPP11_LONG_LONG + /// Unsigned integer of (at least) 64 bits width. + template<> struct bits : conditional::digits>=64,unsigned long,unsigned long long> {}; + #else + /// Unsigned integer of (at least) 64 bits width. + template<> struct bits { typedef unsigned long type; }; + #endif + #endif + + #ifdef HALF_ARITHMETIC_TYPE + /// Type to use for arithmetic computations and mathematical functions internally. + typedef HALF_ARITHMETIC_TYPE internal_t; + #endif + + /// Tag type for binary construction. + struct binary_t {}; + + /// Tag for binary construction. + HALF_CONSTEXPR_CONST binary_t binary = binary_t(); + + /// \name Implementation defined classification and arithmetic + /// \{ + + /// Check for infinity. + /// \tparam T argument type (builtin floating-point type) + /// \param arg value to query + /// \retval true if infinity + /// \retval false else + template bool builtin_isinf(T arg) + { + #if HALF_ENABLE_CPP11_CMATH + return std::isinf(arg); + #elif defined(_MSC_VER) + return !::_finite(static_cast(arg)) && !::_isnan(static_cast(arg)); + #else + return arg == std::numeric_limits::infinity() || arg == -std::numeric_limits::infinity(); + #endif + } + + /// Check for NaN. + /// \tparam T argument type (builtin floating-point type) + /// \param arg value to query + /// \retval true if not a number + /// \retval false else + template bool builtin_isnan(T arg) + { + #if HALF_ENABLE_CPP11_CMATH + return std::isnan(arg); + #elif defined(_MSC_VER) + return ::_isnan(static_cast(arg)) != 0; + #else + return arg != arg; + #endif + } + + /// Check sign. + /// \tparam T argument type (builtin floating-point type) + /// \param arg value to query + /// \retval true if signbit set + /// \retval false else + template bool builtin_signbit(T arg) + { + #if HALF_ENABLE_CPP11_CMATH + return std::signbit(arg); + #else + return arg < T() || (arg == T() && T(1)/arg < T()); + #endif + } + + /// Platform-independent sign mask. + /// \param arg integer value in two's complement + /// \retval -1 if \a arg negative + /// \retval 0 if \a arg positive + inline uint32 sign_mask(uint32 arg) + { + static const int N = std::numeric_limits::digits - 1; + #if HALF_TWOS_COMPLEMENT_INT + return static_cast(arg) >> N; + #else + return -((arg>>N)&1); + #endif + } + + /// Platform-independent arithmetic right shift. + /// \param arg integer value in two's complement + /// \param i shift amount (at most 31) + /// \return \a arg right shifted for \a i bits with possible sign extension + inline uint32 arithmetic_shift(uint32 arg, int i) + { + #if HALF_TWOS_COMPLEMENT_INT + return static_cast(arg) >> i; + #else + return static_cast(arg)/(static_cast(1)<>(std::numeric_limits::digits-1))&1); + #endif + } + + /// \} + /// \name Error handling + /// \{ + + /// Internal exception flags. + /// \return reference to global exception flags + inline int& errflags() { HALF_THREAD_LOCAL int flags = 0; return flags; } + + /// Raise floating-point exception. + /// \param flags exceptions to raise + /// \param cond condition to raise exceptions for + inline void raise(int HALF_UNUSED_NOERR(flags), bool HALF_UNUSED_NOERR(cond) = true) + { + #if HALF_ERRHANDLING + if(!cond) + return; + #if HALF_ERRHANDLING_FLAGS + errflags() |= flags; + #endif + #if HALF_ERRHANDLING_ERRNO + if(flags & FE_INVALID) + errno = EDOM; + else if(flags & (FE_DIVBYZERO|FE_OVERFLOW|FE_UNDERFLOW)) + errno = ERANGE; + #endif + #if HALF_ERRHANDLING_FENV && HALF_ENABLE_CPP11_CFENV + std::feraiseexcept(flags); + #endif + #ifdef HALF_ERRHANDLING_THROW_INVALID + if(flags & FE_INVALID) + throw std::domain_error(HALF_ERRHANDLING_THROW_INVALID); + #endif + #ifdef HALF_ERRHANDLING_THROW_DIVBYZERO + if(flags & FE_DIVBYZERO) + throw std::domain_error(HALF_ERRHANDLING_THROW_DIVBYZERO); + #endif + #ifdef HALF_ERRHANDLING_THROW_OVERFLOW + if(flags & FE_OVERFLOW) + throw std::overflow_error(HALF_ERRHANDLING_THROW_OVERFLOW); + #endif + #ifdef HALF_ERRHANDLING_THROW_UNDERFLOW + if(flags & FE_UNDERFLOW) + throw std::underflow_error(HALF_ERRHANDLING_THROW_UNDERFLOW); + #endif + #ifdef HALF_ERRHANDLING_THROW_INEXACT + if(flags & FE_INEXACT) + throw std::range_error(HALF_ERRHANDLING_THROW_INEXACT); + #endif + #if HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT + if((flags & FE_UNDERFLOW) && !(flags & FE_INEXACT)) + detail::raise(FE_INEXACT); + #endif + #if HALF_ERRHANDLING_OVERFLOW_TO_INEXACT + if((flags & FE_OVERFLOW) && !(flags & FE_INEXACT)) + detail::raise(FE_INEXACT); + #endif + #endif + } + + /// Check and signal for any NaN. + /// \param x first half-precision value to check + /// \param y second half-precision value to check + /// \retval true if either \a x or \a y is NaN + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool compsignal(unsigned int x, unsigned int y) + { + #if HALF_ERRHANDLING + detail::raise(FE_INVALID, (x&0x7FFF)>0x7C00 || (y&0x7FFF)>0x7C00); + #endif + return (x&0x7FFF) > 0x7C00 || (y&0x7FFF) > 0x7C00; + } + + /// Signal and silence signaling NaN. + /// \param nan half-precision NaN value + /// \return quiet NaN + /// \exception FE_INVALID if \a nan is signaling NaN + inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int nan) + { + #if HALF_ERRHANDLING + detail::raise(FE_INVALID, !(nan&0x200)); + #endif + return nan | 0x200; + } + + /// Signal and silence signaling NaNs. + /// \param x first half-precision value to check + /// \param y second half-precision value to check + /// \return quiet NaN + /// \exception FE_INVALID if \a x or \a y is signaling NaN + inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int x, unsigned int y) + { + #if HALF_ERRHANDLING + detail::raise(FE_INVALID, ((x&0x7FFF)>0x7C00 && !(x&0x200)) || ((y&0x7FFF)>0x7C00 && !(y&0x200))); + #endif + return ((x&0x7FFF)>0x7C00) ? (x|0x200) : (y|0x200); + } + + /// Signal and silence signaling NaNs. + /// \param x first half-precision value to check + /// \param y second half-precision value to check + /// \param z third half-precision value to check + /// \return quiet NaN + /// \exception FE_INVALID if \a x, \a y or \a z is signaling NaN + inline HALF_CONSTEXPR_NOERR unsigned int signal(unsigned int x, unsigned int y, unsigned int z) + { + #if HALF_ERRHANDLING + detail::raise(FE_INVALID, ((x&0x7FFF)>0x7C00 && !(x&0x200)) || ((y&0x7FFF)>0x7C00 && !(y&0x200)) || ((z&0x7FFF)>0x7C00 && !(z&0x200))); + #endif + return ((x&0x7FFF)>0x7C00) ? (x|0x200) : ((y&0x7FFF)>0x7C00) ? (y|0x200) : (z|0x200); + } + + /// Select value or signaling NaN. + /// \param x preferred half-precision value + /// \param y ignored half-precision value except for signaling NaN + /// \return \a y if signaling NaN, \a x otherwise + /// \exception FE_INVALID if \a y is signaling NaN + inline HALF_CONSTEXPR_NOERR unsigned int select(unsigned int x, unsigned int HALF_UNUSED_NOERR(y)) + { + #if HALF_ERRHANDLING + return (((y&0x7FFF)>0x7C00) && !(y&0x200)) ? detail::signal(y) : x; + #else + return x; + #endif + } + + /// Raise domain error and return NaN. + /// return quiet NaN + /// \exception FE_INVALID + inline HALF_CONSTEXPR_NOERR unsigned int invalid() + { + #if HALF_ERRHANDLING + detail::raise(FE_INVALID); + #endif + return 0x7FFF; + } + + /// Raise pole error and return infinity. + /// \param sign half-precision value with sign bit only + /// \return half-precision infinity with sign of \a sign + /// \exception FE_DIVBYZERO + inline HALF_CONSTEXPR_NOERR unsigned int pole(unsigned int sign = 0) + { + #if HALF_ERRHANDLING + detail::raise(FE_DIVBYZERO); + #endif + return sign | 0x7C00; + } + + /// Check value for underflow. + /// \param arg non-zero half-precision value to check + /// \return \a arg + /// \exception FE_UNDERFLOW if arg is subnormal + inline HALF_CONSTEXPR_NOERR unsigned int check_underflow(unsigned int arg) + { + #if HALF_ERRHANDLING && !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT + detail::raise(FE_UNDERFLOW, !(arg&0x7C00)); + #endif + return arg; + } + + /// \} + /// \name Conversion and rounding + /// \{ + + /// Half-precision overflow. + /// \tparam R rounding mode to use + /// \param sign half-precision value with sign bit only + /// \return rounded overflowing half-precision value + /// \exception FE_OVERFLOW + template HALF_CONSTEXPR_NOERR unsigned int overflow(unsigned int sign = 0) + { + #if HALF_ERRHANDLING + detail::raise(FE_OVERFLOW); + #endif + return (R==std::round_toward_infinity) ? (sign+0x7C00-(sign>>15)) : + (R==std::round_toward_neg_infinity) ? (sign+0x7BFF+(sign>>15)) : + (R==std::round_toward_zero) ? (sign|0x7BFF) : + (sign|0x7C00); + } + + /// Half-precision underflow. + /// \tparam R rounding mode to use + /// \param sign half-precision value with sign bit only + /// \return rounded underflowing half-precision value + /// \exception FE_UNDERFLOW + template HALF_CONSTEXPR_NOERR unsigned int underflow(unsigned int sign = 0) + { + #if HALF_ERRHANDLING + detail::raise(FE_UNDERFLOW); + #endif + return (R==std::round_toward_infinity) ? (sign+1-(sign>>15)) : + (R==std::round_toward_neg_infinity) ? (sign+(sign>>15)) : + sign; + } + + /// Round half-precision number. + /// \tparam R rounding mode to use + /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results + /// \param value finite half-precision number to round + /// \param g guard bit (most significant discarded bit) + /// \param s sticky bit (or of all but the most significant discarded bits) + /// \return rounded half-precision value + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded or \a I is `true` + template HALF_CONSTEXPR_NOERR unsigned int rounded(unsigned int value, int g, int s) + { + #if HALF_ERRHANDLING + value += (R==std::round_to_nearest) ? (g&(s|value)) : + (R==std::round_toward_infinity) ? (~(value>>15)&(g|s)) : + (R==std::round_toward_neg_infinity) ? ((value>>15)&(g|s)) : 0; + if((value&0x7C00) == 0x7C00) + detail::raise(FE_OVERFLOW); + else if(value & 0x7C00) + detail::raise(FE_INEXACT, I || (g|s)!=0); + else + detail::raise(FE_UNDERFLOW, !(HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT) || I || (g|s)!=0); + return value; + #else + return (R==std::round_to_nearest) ? (value+(g&(s|value))) : + (R==std::round_toward_infinity) ? (value+(~(value>>15)&(g|s))) : + (R==std::round_toward_neg_infinity) ? (value+((value>>15)&(g|s))) : + value; + #endif + } + + /// Round half-precision number to nearest integer value. + /// \tparam R rounding mode to use + /// \tparam E `true` for round to even, `false` for round away from zero + /// \tparam I `true` to raise INEXACT exception (if inexact), `false` to never raise it + /// \param value half-precision value to round + /// \return nearest integral half-precision value + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT if value had to be rounded and \a I is `true` + template unsigned int integral(unsigned int value) + { + unsigned int abs = value & 0x7FFF; + if(abs < 0x3C00) + { + detail::raise(FE_INEXACT, I); + return ((R==std::round_to_nearest) ? (0x3C00&-static_cast(abs>=(0x3800+E))) : + (R==std::round_toward_infinity) ? (0x3C00&-(~(value>>15)&(abs!=0))) : + (R==std::round_toward_neg_infinity) ? (0x3C00&-static_cast(value>0x8000)) : + 0) | (value&0x8000); + } + if(abs >= 0x6400) + return (abs>0x7C00) ? detail::signal(value) : value; + unsigned int exp = 25 - (abs>>10), mask = (1<>exp)&E)) : + (R==std::round_toward_infinity) ? (mask&((value>>15)-1)) : + (R==std::round_toward_neg_infinity) ? (mask&-(value>>15)) : + 0) + value) & ~mask; + } + + /// Convert fixed point to half-precision floating-point. + /// \tparam R rounding mode to use + /// \tparam F number of fractional bits in [11,31] + /// \tparam S `true` for signed, `false` for unsigned + /// \tparam N `true` for additional normalization step, `false` if already normalized to 1.F + /// \tparam I `true` to always raise INEXACT exception, `false` to raise only for rounded results + /// \param m mantissa in Q1.F fixed point format + /// \param exp biased exponent - 1 + /// \param sign half-precision value with sign bit only + /// \param s sticky bit (or of all but the most significant already discarded bits) + /// \return value converted to half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded or \a I is `true` + template unsigned int fixed2half(uint32 m, int exp = 14, unsigned int sign = 0, int s = 0) + { + if(S) + { + uint32 msign = sign_mask(m); + m = (m^msign) - msign; + sign = msign & 0x8000; + } + if(N) + for(; m<(static_cast(1)<(sign+static_cast(m>>(F-10-exp)), static_cast((m>>(F-11-exp))&1), s|((m&((static_cast(1)<<(F-11-exp))-1))!=0)); + return rounded(sign+(exp<<10)+static_cast(m>>(F-10)), static_cast((m>>(F-11))&1), s|((m&((static_cast(1)<<(F-11))-1))!=0)); + } + + /// Convert IEEE single-precision to half-precision. + /// Credit for this goes to [Jeroen van der Zijp](ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf). + /// \tparam R rounding mode to use + /// \param value single-precision value to convert + /// \return rounded half-precision value + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded + template unsigned int float2half_impl(float value, true_type) + { + #if HALF_ENABLE_F16C_INTRINSICS + return _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(value), + (R==std::round_to_nearest) ? _MM_FROUND_TO_NEAREST_INT : + (R==std::round_toward_zero) ? _MM_FROUND_TO_ZERO : + (R==std::round_toward_infinity) ? _MM_FROUND_TO_POS_INF : + (R==std::round_toward_neg_infinity) ? _MM_FROUND_TO_NEG_INF : + _MM_FROUND_CUR_DIRECTION)); + #else + bits::type fbits; + std::memcpy(&fbits, &value, sizeof(float)); + #if 1 + unsigned int sign = (fbits>>16) & 0x8000; + fbits &= 0x7FFFFFFF; + if(fbits >= 0x7F800000) + return sign | 0x7C00 | ((fbits>0x7F800000) ? (0x200|((fbits>>13)&0x3FF)) : 0); + if(fbits >= 0x47800000) + return overflow(sign); + if(fbits >= 0x38800000) + return rounded(sign|(((fbits>>23)-112)<<10)|((fbits>>13)&0x3FF), (fbits>>12)&1, (fbits&0xFFF)!=0); + if(fbits >= 0x33000000) + { + int i = 125 - (fbits>>23); + fbits = (fbits&0x7FFFFF) | 0x800000; + return rounded(sign|(fbits>>(i+1)), (fbits>>i)&1, (fbits&((static_cast(1)<(sign); + return sign; + #else + static const uint16 base_table[512] = { + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, + 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080, 0x0100, + 0x0200, 0x0400, 0x0800, 0x0C00, 0x1000, 0x1400, 0x1800, 0x1C00, 0x2000, 0x2400, 0x2800, 0x2C00, 0x3000, 0x3400, 0x3800, 0x3C00, + 0x4000, 0x4400, 0x4800, 0x4C00, 0x5000, 0x5400, 0x5800, 0x5C00, 0x6000, 0x6400, 0x6800, 0x6C00, 0x7000, 0x7400, 0x7800, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, + 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7BFF, 0x7C00, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, + 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8000, 0x8001, 0x8002, 0x8004, 0x8008, 0x8010, 0x8020, 0x8040, 0x8080, 0x8100, + 0x8200, 0x8400, 0x8800, 0x8C00, 0x9000, 0x9400, 0x9800, 0x9C00, 0xA000, 0xA400, 0xA800, 0xAC00, 0xB000, 0xB400, 0xB800, 0xBC00, + 0xC000, 0xC400, 0xC800, 0xCC00, 0xD000, 0xD400, 0xD800, 0xDC00, 0xE000, 0xE400, 0xE800, 0xEC00, 0xF000, 0xF400, 0xF800, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, + 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFBFF, 0xFC00 }; + static const unsigned char shift_table[256] = { + 24, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, + 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, + 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, 25, + 25, 25, 25, 25, 25, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, + 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, + 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, + 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, + 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 13 }; + int sexp = fbits >> 23, exp = sexp & 0xFF, i = shift_table[exp]; + fbits &= 0x7FFFFF; + uint32 m = (fbits|((exp!=0)<<23)) & -static_cast(exp!=0xFF); + return rounded(base_table[sexp]+(fbits>>i), (m>>(i-1))&1, (((static_cast(1)<<(i-1))-1)&m)!=0); + #endif + #endif + } + + /// Convert IEEE double-precision to half-precision. + /// \tparam R rounding mode to use + /// \param value double-precision value to convert + /// \return rounded half-precision value + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded + template unsigned int float2half_impl(double value, true_type) + { + #if HALF_ENABLE_F16C_INTRINSICS + if(R == std::round_indeterminate) + return _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_cvtpd_ps(_mm_set_sd(value)), _MM_FROUND_CUR_DIRECTION)); + #endif + bits::type dbits; + std::memcpy(&dbits, &value, sizeof(double)); + uint32 hi = dbits >> 32, lo = dbits & 0xFFFFFFFF; + unsigned int sign = (hi>>16) & 0x8000; + hi &= 0x7FFFFFFF; + if(hi >= 0x7FF00000) + return sign | 0x7C00 | ((dbits&0xFFFFFFFFFFFFF) ? (0x200|((hi>>10)&0x3FF)) : 0); + if(hi >= 0x40F00000) + return overflow(sign); + if(hi >= 0x3F100000) + return rounded(sign|static_cast(((hi>>20)-1008)<<10)|static_cast((hi>>10)&0x3FF), static_cast((hi>>9)&1), ((hi&0x1FF)|lo)!=0); + if(hi >= 0x3E600000) + { + int i = static_cast(1018 - (hi>>20)); + hi = (hi&0xFFFFF) | 0x100000; + return rounded(sign|static_cast(hi>>(i+1)), static_cast((hi>>i)&1), ((hi&((static_cast(1)<(sign); + return sign; + } + + /// Convert non-IEEE floating-point to half-precision. + /// \tparam R rounding mode to use + /// \tparam T source type (builtin floating-point type) + /// \param value floating-point value to convert + /// \return rounded half-precision value + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded + template unsigned int float2half_impl(T value, ...) + { + unsigned int hbits = static_cast(builtin_signbit(value)) << 15; + if(value == T()) + return hbits; + if(builtin_isnan(value)) + return hbits | 0x7FFF; + if(builtin_isinf(value)) + return hbits | 0x7C00; + int exp; + std::frexp(value, &exp); + if(exp > 16) + return overflow(hbits); + if(exp < -13) + value = std::ldexp(value, 25); + else + { + value = std::ldexp(value, 12-exp); + hbits |= ((exp+13)<<10); + } + T ival, frac = std::modf(value, &ival); + int m = std::abs(static_cast(ival)); + return rounded(hbits+(m>>1), m&1, frac!=T()); + } + + /// Convert floating-point to half-precision. + /// \tparam R rounding mode to use + /// \tparam T source type (builtin floating-point type) + /// \param value floating-point value to convert + /// \return rounded half-precision value + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded + template unsigned int float2half(T value) + { + return float2half_impl(value, bool_type::is_iec559&&sizeof(typename bits::type)==sizeof(T)>()); + } + + /// Convert integer to half-precision floating-point. + /// \tparam R rounding mode to use + /// \tparam T type to convert (builtin integer type) + /// \param value integral value to convert + /// \return rounded half-precision value + /// \exception FE_OVERFLOW on overflows + /// \exception FE_INEXACT if value had to be rounded + template unsigned int int2half(T value) + { + unsigned int bits = static_cast(value<0) << 15; + if(!value) + return bits; + if(value > 0xFFE0 || (bits&&value < -0xFFE0)) + return overflow(bits); + unsigned int m = static_cast(value), exp = 24; + if(bits) + m = -m; + for(; m<0x400; m<<=1,--exp) ; + for(; m>0x7FF; m>>=1,++exp) ; + bits |= (exp<<10) + m; + return (exp>24) ? rounded(bits, static_cast(value>>(exp-25))&1, (((1<<(exp-25))-1)&value)!=0) : bits; + } + + /// Convert half-precision to IEEE single-precision. + /// Credit for this goes to [Jeroen van der Zijp](ftp://ftp.fox-toolkit.org/pub/fasthalffloatconversion.pdf). + /// \param value half-precision value to convert + /// \return single-precision value + inline float half2float_impl(unsigned int value, float, true_type) + { + #if HALF_ENABLE_F16C_INTRINSICS + return _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(value))); + #else + #if 0 + bits::type fbits = static_cast::type>(value&0x8000) << 16; + int abs = value & 0x7FFF; + if(abs) + { + fbits |= 0x38000000 << static_cast(abs>=0x7C00); + for(; abs<0x400; abs<<=1,fbits-=0x800000) ; + fbits += static_cast::type>(abs) << 13; + } + #else + static const bits::type mantissa_table[2048] = { + 0x00000000, 0x33800000, 0x34000000, 0x34400000, 0x34800000, 0x34A00000, 0x34C00000, 0x34E00000, 0x35000000, 0x35100000, 0x35200000, 0x35300000, 0x35400000, 0x35500000, 0x35600000, 0x35700000, + 0x35800000, 0x35880000, 0x35900000, 0x35980000, 0x35A00000, 0x35A80000, 0x35B00000, 0x35B80000, 0x35C00000, 0x35C80000, 0x35D00000, 0x35D80000, 0x35E00000, 0x35E80000, 0x35F00000, 0x35F80000, + 0x36000000, 0x36040000, 0x36080000, 0x360C0000, 0x36100000, 0x36140000, 0x36180000, 0x361C0000, 0x36200000, 0x36240000, 0x36280000, 0x362C0000, 0x36300000, 0x36340000, 0x36380000, 0x363C0000, + 0x36400000, 0x36440000, 0x36480000, 0x364C0000, 0x36500000, 0x36540000, 0x36580000, 0x365C0000, 0x36600000, 0x36640000, 0x36680000, 0x366C0000, 0x36700000, 0x36740000, 0x36780000, 0x367C0000, + 0x36800000, 0x36820000, 0x36840000, 0x36860000, 0x36880000, 0x368A0000, 0x368C0000, 0x368E0000, 0x36900000, 0x36920000, 0x36940000, 0x36960000, 0x36980000, 0x369A0000, 0x369C0000, 0x369E0000, + 0x36A00000, 0x36A20000, 0x36A40000, 0x36A60000, 0x36A80000, 0x36AA0000, 0x36AC0000, 0x36AE0000, 0x36B00000, 0x36B20000, 0x36B40000, 0x36B60000, 0x36B80000, 0x36BA0000, 0x36BC0000, 0x36BE0000, + 0x36C00000, 0x36C20000, 0x36C40000, 0x36C60000, 0x36C80000, 0x36CA0000, 0x36CC0000, 0x36CE0000, 0x36D00000, 0x36D20000, 0x36D40000, 0x36D60000, 0x36D80000, 0x36DA0000, 0x36DC0000, 0x36DE0000, + 0x36E00000, 0x36E20000, 0x36E40000, 0x36E60000, 0x36E80000, 0x36EA0000, 0x36EC0000, 0x36EE0000, 0x36F00000, 0x36F20000, 0x36F40000, 0x36F60000, 0x36F80000, 0x36FA0000, 0x36FC0000, 0x36FE0000, + 0x37000000, 0x37010000, 0x37020000, 0x37030000, 0x37040000, 0x37050000, 0x37060000, 0x37070000, 0x37080000, 0x37090000, 0x370A0000, 0x370B0000, 0x370C0000, 0x370D0000, 0x370E0000, 0x370F0000, + 0x37100000, 0x37110000, 0x37120000, 0x37130000, 0x37140000, 0x37150000, 0x37160000, 0x37170000, 0x37180000, 0x37190000, 0x371A0000, 0x371B0000, 0x371C0000, 0x371D0000, 0x371E0000, 0x371F0000, + 0x37200000, 0x37210000, 0x37220000, 0x37230000, 0x37240000, 0x37250000, 0x37260000, 0x37270000, 0x37280000, 0x37290000, 0x372A0000, 0x372B0000, 0x372C0000, 0x372D0000, 0x372E0000, 0x372F0000, + 0x37300000, 0x37310000, 0x37320000, 0x37330000, 0x37340000, 0x37350000, 0x37360000, 0x37370000, 0x37380000, 0x37390000, 0x373A0000, 0x373B0000, 0x373C0000, 0x373D0000, 0x373E0000, 0x373F0000, + 0x37400000, 0x37410000, 0x37420000, 0x37430000, 0x37440000, 0x37450000, 0x37460000, 0x37470000, 0x37480000, 0x37490000, 0x374A0000, 0x374B0000, 0x374C0000, 0x374D0000, 0x374E0000, 0x374F0000, + 0x37500000, 0x37510000, 0x37520000, 0x37530000, 0x37540000, 0x37550000, 0x37560000, 0x37570000, 0x37580000, 0x37590000, 0x375A0000, 0x375B0000, 0x375C0000, 0x375D0000, 0x375E0000, 0x375F0000, + 0x37600000, 0x37610000, 0x37620000, 0x37630000, 0x37640000, 0x37650000, 0x37660000, 0x37670000, 0x37680000, 0x37690000, 0x376A0000, 0x376B0000, 0x376C0000, 0x376D0000, 0x376E0000, 0x376F0000, + 0x37700000, 0x37710000, 0x37720000, 0x37730000, 0x37740000, 0x37750000, 0x37760000, 0x37770000, 0x37780000, 0x37790000, 0x377A0000, 0x377B0000, 0x377C0000, 0x377D0000, 0x377E0000, 0x377F0000, + 0x37800000, 0x37808000, 0x37810000, 0x37818000, 0x37820000, 0x37828000, 0x37830000, 0x37838000, 0x37840000, 0x37848000, 0x37850000, 0x37858000, 0x37860000, 0x37868000, 0x37870000, 0x37878000, + 0x37880000, 0x37888000, 0x37890000, 0x37898000, 0x378A0000, 0x378A8000, 0x378B0000, 0x378B8000, 0x378C0000, 0x378C8000, 0x378D0000, 0x378D8000, 0x378E0000, 0x378E8000, 0x378F0000, 0x378F8000, + 0x37900000, 0x37908000, 0x37910000, 0x37918000, 0x37920000, 0x37928000, 0x37930000, 0x37938000, 0x37940000, 0x37948000, 0x37950000, 0x37958000, 0x37960000, 0x37968000, 0x37970000, 0x37978000, + 0x37980000, 0x37988000, 0x37990000, 0x37998000, 0x379A0000, 0x379A8000, 0x379B0000, 0x379B8000, 0x379C0000, 0x379C8000, 0x379D0000, 0x379D8000, 0x379E0000, 0x379E8000, 0x379F0000, 0x379F8000, + 0x37A00000, 0x37A08000, 0x37A10000, 0x37A18000, 0x37A20000, 0x37A28000, 0x37A30000, 0x37A38000, 0x37A40000, 0x37A48000, 0x37A50000, 0x37A58000, 0x37A60000, 0x37A68000, 0x37A70000, 0x37A78000, + 0x37A80000, 0x37A88000, 0x37A90000, 0x37A98000, 0x37AA0000, 0x37AA8000, 0x37AB0000, 0x37AB8000, 0x37AC0000, 0x37AC8000, 0x37AD0000, 0x37AD8000, 0x37AE0000, 0x37AE8000, 0x37AF0000, 0x37AF8000, + 0x37B00000, 0x37B08000, 0x37B10000, 0x37B18000, 0x37B20000, 0x37B28000, 0x37B30000, 0x37B38000, 0x37B40000, 0x37B48000, 0x37B50000, 0x37B58000, 0x37B60000, 0x37B68000, 0x37B70000, 0x37B78000, + 0x37B80000, 0x37B88000, 0x37B90000, 0x37B98000, 0x37BA0000, 0x37BA8000, 0x37BB0000, 0x37BB8000, 0x37BC0000, 0x37BC8000, 0x37BD0000, 0x37BD8000, 0x37BE0000, 0x37BE8000, 0x37BF0000, 0x37BF8000, + 0x37C00000, 0x37C08000, 0x37C10000, 0x37C18000, 0x37C20000, 0x37C28000, 0x37C30000, 0x37C38000, 0x37C40000, 0x37C48000, 0x37C50000, 0x37C58000, 0x37C60000, 0x37C68000, 0x37C70000, 0x37C78000, + 0x37C80000, 0x37C88000, 0x37C90000, 0x37C98000, 0x37CA0000, 0x37CA8000, 0x37CB0000, 0x37CB8000, 0x37CC0000, 0x37CC8000, 0x37CD0000, 0x37CD8000, 0x37CE0000, 0x37CE8000, 0x37CF0000, 0x37CF8000, + 0x37D00000, 0x37D08000, 0x37D10000, 0x37D18000, 0x37D20000, 0x37D28000, 0x37D30000, 0x37D38000, 0x37D40000, 0x37D48000, 0x37D50000, 0x37D58000, 0x37D60000, 0x37D68000, 0x37D70000, 0x37D78000, + 0x37D80000, 0x37D88000, 0x37D90000, 0x37D98000, 0x37DA0000, 0x37DA8000, 0x37DB0000, 0x37DB8000, 0x37DC0000, 0x37DC8000, 0x37DD0000, 0x37DD8000, 0x37DE0000, 0x37DE8000, 0x37DF0000, 0x37DF8000, + 0x37E00000, 0x37E08000, 0x37E10000, 0x37E18000, 0x37E20000, 0x37E28000, 0x37E30000, 0x37E38000, 0x37E40000, 0x37E48000, 0x37E50000, 0x37E58000, 0x37E60000, 0x37E68000, 0x37E70000, 0x37E78000, + 0x37E80000, 0x37E88000, 0x37E90000, 0x37E98000, 0x37EA0000, 0x37EA8000, 0x37EB0000, 0x37EB8000, 0x37EC0000, 0x37EC8000, 0x37ED0000, 0x37ED8000, 0x37EE0000, 0x37EE8000, 0x37EF0000, 0x37EF8000, + 0x37F00000, 0x37F08000, 0x37F10000, 0x37F18000, 0x37F20000, 0x37F28000, 0x37F30000, 0x37F38000, 0x37F40000, 0x37F48000, 0x37F50000, 0x37F58000, 0x37F60000, 0x37F68000, 0x37F70000, 0x37F78000, + 0x37F80000, 0x37F88000, 0x37F90000, 0x37F98000, 0x37FA0000, 0x37FA8000, 0x37FB0000, 0x37FB8000, 0x37FC0000, 0x37FC8000, 0x37FD0000, 0x37FD8000, 0x37FE0000, 0x37FE8000, 0x37FF0000, 0x37FF8000, + 0x38000000, 0x38004000, 0x38008000, 0x3800C000, 0x38010000, 0x38014000, 0x38018000, 0x3801C000, 0x38020000, 0x38024000, 0x38028000, 0x3802C000, 0x38030000, 0x38034000, 0x38038000, 0x3803C000, + 0x38040000, 0x38044000, 0x38048000, 0x3804C000, 0x38050000, 0x38054000, 0x38058000, 0x3805C000, 0x38060000, 0x38064000, 0x38068000, 0x3806C000, 0x38070000, 0x38074000, 0x38078000, 0x3807C000, + 0x38080000, 0x38084000, 0x38088000, 0x3808C000, 0x38090000, 0x38094000, 0x38098000, 0x3809C000, 0x380A0000, 0x380A4000, 0x380A8000, 0x380AC000, 0x380B0000, 0x380B4000, 0x380B8000, 0x380BC000, + 0x380C0000, 0x380C4000, 0x380C8000, 0x380CC000, 0x380D0000, 0x380D4000, 0x380D8000, 0x380DC000, 0x380E0000, 0x380E4000, 0x380E8000, 0x380EC000, 0x380F0000, 0x380F4000, 0x380F8000, 0x380FC000, + 0x38100000, 0x38104000, 0x38108000, 0x3810C000, 0x38110000, 0x38114000, 0x38118000, 0x3811C000, 0x38120000, 0x38124000, 0x38128000, 0x3812C000, 0x38130000, 0x38134000, 0x38138000, 0x3813C000, + 0x38140000, 0x38144000, 0x38148000, 0x3814C000, 0x38150000, 0x38154000, 0x38158000, 0x3815C000, 0x38160000, 0x38164000, 0x38168000, 0x3816C000, 0x38170000, 0x38174000, 0x38178000, 0x3817C000, + 0x38180000, 0x38184000, 0x38188000, 0x3818C000, 0x38190000, 0x38194000, 0x38198000, 0x3819C000, 0x381A0000, 0x381A4000, 0x381A8000, 0x381AC000, 0x381B0000, 0x381B4000, 0x381B8000, 0x381BC000, + 0x381C0000, 0x381C4000, 0x381C8000, 0x381CC000, 0x381D0000, 0x381D4000, 0x381D8000, 0x381DC000, 0x381E0000, 0x381E4000, 0x381E8000, 0x381EC000, 0x381F0000, 0x381F4000, 0x381F8000, 0x381FC000, + 0x38200000, 0x38204000, 0x38208000, 0x3820C000, 0x38210000, 0x38214000, 0x38218000, 0x3821C000, 0x38220000, 0x38224000, 0x38228000, 0x3822C000, 0x38230000, 0x38234000, 0x38238000, 0x3823C000, + 0x38240000, 0x38244000, 0x38248000, 0x3824C000, 0x38250000, 0x38254000, 0x38258000, 0x3825C000, 0x38260000, 0x38264000, 0x38268000, 0x3826C000, 0x38270000, 0x38274000, 0x38278000, 0x3827C000, + 0x38280000, 0x38284000, 0x38288000, 0x3828C000, 0x38290000, 0x38294000, 0x38298000, 0x3829C000, 0x382A0000, 0x382A4000, 0x382A8000, 0x382AC000, 0x382B0000, 0x382B4000, 0x382B8000, 0x382BC000, + 0x382C0000, 0x382C4000, 0x382C8000, 0x382CC000, 0x382D0000, 0x382D4000, 0x382D8000, 0x382DC000, 0x382E0000, 0x382E4000, 0x382E8000, 0x382EC000, 0x382F0000, 0x382F4000, 0x382F8000, 0x382FC000, + 0x38300000, 0x38304000, 0x38308000, 0x3830C000, 0x38310000, 0x38314000, 0x38318000, 0x3831C000, 0x38320000, 0x38324000, 0x38328000, 0x3832C000, 0x38330000, 0x38334000, 0x38338000, 0x3833C000, + 0x38340000, 0x38344000, 0x38348000, 0x3834C000, 0x38350000, 0x38354000, 0x38358000, 0x3835C000, 0x38360000, 0x38364000, 0x38368000, 0x3836C000, 0x38370000, 0x38374000, 0x38378000, 0x3837C000, + 0x38380000, 0x38384000, 0x38388000, 0x3838C000, 0x38390000, 0x38394000, 0x38398000, 0x3839C000, 0x383A0000, 0x383A4000, 0x383A8000, 0x383AC000, 0x383B0000, 0x383B4000, 0x383B8000, 0x383BC000, + 0x383C0000, 0x383C4000, 0x383C8000, 0x383CC000, 0x383D0000, 0x383D4000, 0x383D8000, 0x383DC000, 0x383E0000, 0x383E4000, 0x383E8000, 0x383EC000, 0x383F0000, 0x383F4000, 0x383F8000, 0x383FC000, + 0x38400000, 0x38404000, 0x38408000, 0x3840C000, 0x38410000, 0x38414000, 0x38418000, 0x3841C000, 0x38420000, 0x38424000, 0x38428000, 0x3842C000, 0x38430000, 0x38434000, 0x38438000, 0x3843C000, + 0x38440000, 0x38444000, 0x38448000, 0x3844C000, 0x38450000, 0x38454000, 0x38458000, 0x3845C000, 0x38460000, 0x38464000, 0x38468000, 0x3846C000, 0x38470000, 0x38474000, 0x38478000, 0x3847C000, + 0x38480000, 0x38484000, 0x38488000, 0x3848C000, 0x38490000, 0x38494000, 0x38498000, 0x3849C000, 0x384A0000, 0x384A4000, 0x384A8000, 0x384AC000, 0x384B0000, 0x384B4000, 0x384B8000, 0x384BC000, + 0x384C0000, 0x384C4000, 0x384C8000, 0x384CC000, 0x384D0000, 0x384D4000, 0x384D8000, 0x384DC000, 0x384E0000, 0x384E4000, 0x384E8000, 0x384EC000, 0x384F0000, 0x384F4000, 0x384F8000, 0x384FC000, + 0x38500000, 0x38504000, 0x38508000, 0x3850C000, 0x38510000, 0x38514000, 0x38518000, 0x3851C000, 0x38520000, 0x38524000, 0x38528000, 0x3852C000, 0x38530000, 0x38534000, 0x38538000, 0x3853C000, + 0x38540000, 0x38544000, 0x38548000, 0x3854C000, 0x38550000, 0x38554000, 0x38558000, 0x3855C000, 0x38560000, 0x38564000, 0x38568000, 0x3856C000, 0x38570000, 0x38574000, 0x38578000, 0x3857C000, + 0x38580000, 0x38584000, 0x38588000, 0x3858C000, 0x38590000, 0x38594000, 0x38598000, 0x3859C000, 0x385A0000, 0x385A4000, 0x385A8000, 0x385AC000, 0x385B0000, 0x385B4000, 0x385B8000, 0x385BC000, + 0x385C0000, 0x385C4000, 0x385C8000, 0x385CC000, 0x385D0000, 0x385D4000, 0x385D8000, 0x385DC000, 0x385E0000, 0x385E4000, 0x385E8000, 0x385EC000, 0x385F0000, 0x385F4000, 0x385F8000, 0x385FC000, + 0x38600000, 0x38604000, 0x38608000, 0x3860C000, 0x38610000, 0x38614000, 0x38618000, 0x3861C000, 0x38620000, 0x38624000, 0x38628000, 0x3862C000, 0x38630000, 0x38634000, 0x38638000, 0x3863C000, + 0x38640000, 0x38644000, 0x38648000, 0x3864C000, 0x38650000, 0x38654000, 0x38658000, 0x3865C000, 0x38660000, 0x38664000, 0x38668000, 0x3866C000, 0x38670000, 0x38674000, 0x38678000, 0x3867C000, + 0x38680000, 0x38684000, 0x38688000, 0x3868C000, 0x38690000, 0x38694000, 0x38698000, 0x3869C000, 0x386A0000, 0x386A4000, 0x386A8000, 0x386AC000, 0x386B0000, 0x386B4000, 0x386B8000, 0x386BC000, + 0x386C0000, 0x386C4000, 0x386C8000, 0x386CC000, 0x386D0000, 0x386D4000, 0x386D8000, 0x386DC000, 0x386E0000, 0x386E4000, 0x386E8000, 0x386EC000, 0x386F0000, 0x386F4000, 0x386F8000, 0x386FC000, + 0x38700000, 0x38704000, 0x38708000, 0x3870C000, 0x38710000, 0x38714000, 0x38718000, 0x3871C000, 0x38720000, 0x38724000, 0x38728000, 0x3872C000, 0x38730000, 0x38734000, 0x38738000, 0x3873C000, + 0x38740000, 0x38744000, 0x38748000, 0x3874C000, 0x38750000, 0x38754000, 0x38758000, 0x3875C000, 0x38760000, 0x38764000, 0x38768000, 0x3876C000, 0x38770000, 0x38774000, 0x38778000, 0x3877C000, + 0x38780000, 0x38784000, 0x38788000, 0x3878C000, 0x38790000, 0x38794000, 0x38798000, 0x3879C000, 0x387A0000, 0x387A4000, 0x387A8000, 0x387AC000, 0x387B0000, 0x387B4000, 0x387B8000, 0x387BC000, + 0x387C0000, 0x387C4000, 0x387C8000, 0x387CC000, 0x387D0000, 0x387D4000, 0x387D8000, 0x387DC000, 0x387E0000, 0x387E4000, 0x387E8000, 0x387EC000, 0x387F0000, 0x387F4000, 0x387F8000, 0x387FC000, + 0x38000000, 0x38002000, 0x38004000, 0x38006000, 0x38008000, 0x3800A000, 0x3800C000, 0x3800E000, 0x38010000, 0x38012000, 0x38014000, 0x38016000, 0x38018000, 0x3801A000, 0x3801C000, 0x3801E000, + 0x38020000, 0x38022000, 0x38024000, 0x38026000, 0x38028000, 0x3802A000, 0x3802C000, 0x3802E000, 0x38030000, 0x38032000, 0x38034000, 0x38036000, 0x38038000, 0x3803A000, 0x3803C000, 0x3803E000, + 0x38040000, 0x38042000, 0x38044000, 0x38046000, 0x38048000, 0x3804A000, 0x3804C000, 0x3804E000, 0x38050000, 0x38052000, 0x38054000, 0x38056000, 0x38058000, 0x3805A000, 0x3805C000, 0x3805E000, + 0x38060000, 0x38062000, 0x38064000, 0x38066000, 0x38068000, 0x3806A000, 0x3806C000, 0x3806E000, 0x38070000, 0x38072000, 0x38074000, 0x38076000, 0x38078000, 0x3807A000, 0x3807C000, 0x3807E000, + 0x38080000, 0x38082000, 0x38084000, 0x38086000, 0x38088000, 0x3808A000, 0x3808C000, 0x3808E000, 0x38090000, 0x38092000, 0x38094000, 0x38096000, 0x38098000, 0x3809A000, 0x3809C000, 0x3809E000, + 0x380A0000, 0x380A2000, 0x380A4000, 0x380A6000, 0x380A8000, 0x380AA000, 0x380AC000, 0x380AE000, 0x380B0000, 0x380B2000, 0x380B4000, 0x380B6000, 0x380B8000, 0x380BA000, 0x380BC000, 0x380BE000, + 0x380C0000, 0x380C2000, 0x380C4000, 0x380C6000, 0x380C8000, 0x380CA000, 0x380CC000, 0x380CE000, 0x380D0000, 0x380D2000, 0x380D4000, 0x380D6000, 0x380D8000, 0x380DA000, 0x380DC000, 0x380DE000, + 0x380E0000, 0x380E2000, 0x380E4000, 0x380E6000, 0x380E8000, 0x380EA000, 0x380EC000, 0x380EE000, 0x380F0000, 0x380F2000, 0x380F4000, 0x380F6000, 0x380F8000, 0x380FA000, 0x380FC000, 0x380FE000, + 0x38100000, 0x38102000, 0x38104000, 0x38106000, 0x38108000, 0x3810A000, 0x3810C000, 0x3810E000, 0x38110000, 0x38112000, 0x38114000, 0x38116000, 0x38118000, 0x3811A000, 0x3811C000, 0x3811E000, + 0x38120000, 0x38122000, 0x38124000, 0x38126000, 0x38128000, 0x3812A000, 0x3812C000, 0x3812E000, 0x38130000, 0x38132000, 0x38134000, 0x38136000, 0x38138000, 0x3813A000, 0x3813C000, 0x3813E000, + 0x38140000, 0x38142000, 0x38144000, 0x38146000, 0x38148000, 0x3814A000, 0x3814C000, 0x3814E000, 0x38150000, 0x38152000, 0x38154000, 0x38156000, 0x38158000, 0x3815A000, 0x3815C000, 0x3815E000, + 0x38160000, 0x38162000, 0x38164000, 0x38166000, 0x38168000, 0x3816A000, 0x3816C000, 0x3816E000, 0x38170000, 0x38172000, 0x38174000, 0x38176000, 0x38178000, 0x3817A000, 0x3817C000, 0x3817E000, + 0x38180000, 0x38182000, 0x38184000, 0x38186000, 0x38188000, 0x3818A000, 0x3818C000, 0x3818E000, 0x38190000, 0x38192000, 0x38194000, 0x38196000, 0x38198000, 0x3819A000, 0x3819C000, 0x3819E000, + 0x381A0000, 0x381A2000, 0x381A4000, 0x381A6000, 0x381A8000, 0x381AA000, 0x381AC000, 0x381AE000, 0x381B0000, 0x381B2000, 0x381B4000, 0x381B6000, 0x381B8000, 0x381BA000, 0x381BC000, 0x381BE000, + 0x381C0000, 0x381C2000, 0x381C4000, 0x381C6000, 0x381C8000, 0x381CA000, 0x381CC000, 0x381CE000, 0x381D0000, 0x381D2000, 0x381D4000, 0x381D6000, 0x381D8000, 0x381DA000, 0x381DC000, 0x381DE000, + 0x381E0000, 0x381E2000, 0x381E4000, 0x381E6000, 0x381E8000, 0x381EA000, 0x381EC000, 0x381EE000, 0x381F0000, 0x381F2000, 0x381F4000, 0x381F6000, 0x381F8000, 0x381FA000, 0x381FC000, 0x381FE000, + 0x38200000, 0x38202000, 0x38204000, 0x38206000, 0x38208000, 0x3820A000, 0x3820C000, 0x3820E000, 0x38210000, 0x38212000, 0x38214000, 0x38216000, 0x38218000, 0x3821A000, 0x3821C000, 0x3821E000, + 0x38220000, 0x38222000, 0x38224000, 0x38226000, 0x38228000, 0x3822A000, 0x3822C000, 0x3822E000, 0x38230000, 0x38232000, 0x38234000, 0x38236000, 0x38238000, 0x3823A000, 0x3823C000, 0x3823E000, + 0x38240000, 0x38242000, 0x38244000, 0x38246000, 0x38248000, 0x3824A000, 0x3824C000, 0x3824E000, 0x38250000, 0x38252000, 0x38254000, 0x38256000, 0x38258000, 0x3825A000, 0x3825C000, 0x3825E000, + 0x38260000, 0x38262000, 0x38264000, 0x38266000, 0x38268000, 0x3826A000, 0x3826C000, 0x3826E000, 0x38270000, 0x38272000, 0x38274000, 0x38276000, 0x38278000, 0x3827A000, 0x3827C000, 0x3827E000, + 0x38280000, 0x38282000, 0x38284000, 0x38286000, 0x38288000, 0x3828A000, 0x3828C000, 0x3828E000, 0x38290000, 0x38292000, 0x38294000, 0x38296000, 0x38298000, 0x3829A000, 0x3829C000, 0x3829E000, + 0x382A0000, 0x382A2000, 0x382A4000, 0x382A6000, 0x382A8000, 0x382AA000, 0x382AC000, 0x382AE000, 0x382B0000, 0x382B2000, 0x382B4000, 0x382B6000, 0x382B8000, 0x382BA000, 0x382BC000, 0x382BE000, + 0x382C0000, 0x382C2000, 0x382C4000, 0x382C6000, 0x382C8000, 0x382CA000, 0x382CC000, 0x382CE000, 0x382D0000, 0x382D2000, 0x382D4000, 0x382D6000, 0x382D8000, 0x382DA000, 0x382DC000, 0x382DE000, + 0x382E0000, 0x382E2000, 0x382E4000, 0x382E6000, 0x382E8000, 0x382EA000, 0x382EC000, 0x382EE000, 0x382F0000, 0x382F2000, 0x382F4000, 0x382F6000, 0x382F8000, 0x382FA000, 0x382FC000, 0x382FE000, + 0x38300000, 0x38302000, 0x38304000, 0x38306000, 0x38308000, 0x3830A000, 0x3830C000, 0x3830E000, 0x38310000, 0x38312000, 0x38314000, 0x38316000, 0x38318000, 0x3831A000, 0x3831C000, 0x3831E000, + 0x38320000, 0x38322000, 0x38324000, 0x38326000, 0x38328000, 0x3832A000, 0x3832C000, 0x3832E000, 0x38330000, 0x38332000, 0x38334000, 0x38336000, 0x38338000, 0x3833A000, 0x3833C000, 0x3833E000, + 0x38340000, 0x38342000, 0x38344000, 0x38346000, 0x38348000, 0x3834A000, 0x3834C000, 0x3834E000, 0x38350000, 0x38352000, 0x38354000, 0x38356000, 0x38358000, 0x3835A000, 0x3835C000, 0x3835E000, + 0x38360000, 0x38362000, 0x38364000, 0x38366000, 0x38368000, 0x3836A000, 0x3836C000, 0x3836E000, 0x38370000, 0x38372000, 0x38374000, 0x38376000, 0x38378000, 0x3837A000, 0x3837C000, 0x3837E000, + 0x38380000, 0x38382000, 0x38384000, 0x38386000, 0x38388000, 0x3838A000, 0x3838C000, 0x3838E000, 0x38390000, 0x38392000, 0x38394000, 0x38396000, 0x38398000, 0x3839A000, 0x3839C000, 0x3839E000, + 0x383A0000, 0x383A2000, 0x383A4000, 0x383A6000, 0x383A8000, 0x383AA000, 0x383AC000, 0x383AE000, 0x383B0000, 0x383B2000, 0x383B4000, 0x383B6000, 0x383B8000, 0x383BA000, 0x383BC000, 0x383BE000, + 0x383C0000, 0x383C2000, 0x383C4000, 0x383C6000, 0x383C8000, 0x383CA000, 0x383CC000, 0x383CE000, 0x383D0000, 0x383D2000, 0x383D4000, 0x383D6000, 0x383D8000, 0x383DA000, 0x383DC000, 0x383DE000, + 0x383E0000, 0x383E2000, 0x383E4000, 0x383E6000, 0x383E8000, 0x383EA000, 0x383EC000, 0x383EE000, 0x383F0000, 0x383F2000, 0x383F4000, 0x383F6000, 0x383F8000, 0x383FA000, 0x383FC000, 0x383FE000, + 0x38400000, 0x38402000, 0x38404000, 0x38406000, 0x38408000, 0x3840A000, 0x3840C000, 0x3840E000, 0x38410000, 0x38412000, 0x38414000, 0x38416000, 0x38418000, 0x3841A000, 0x3841C000, 0x3841E000, + 0x38420000, 0x38422000, 0x38424000, 0x38426000, 0x38428000, 0x3842A000, 0x3842C000, 0x3842E000, 0x38430000, 0x38432000, 0x38434000, 0x38436000, 0x38438000, 0x3843A000, 0x3843C000, 0x3843E000, + 0x38440000, 0x38442000, 0x38444000, 0x38446000, 0x38448000, 0x3844A000, 0x3844C000, 0x3844E000, 0x38450000, 0x38452000, 0x38454000, 0x38456000, 0x38458000, 0x3845A000, 0x3845C000, 0x3845E000, + 0x38460000, 0x38462000, 0x38464000, 0x38466000, 0x38468000, 0x3846A000, 0x3846C000, 0x3846E000, 0x38470000, 0x38472000, 0x38474000, 0x38476000, 0x38478000, 0x3847A000, 0x3847C000, 0x3847E000, + 0x38480000, 0x38482000, 0x38484000, 0x38486000, 0x38488000, 0x3848A000, 0x3848C000, 0x3848E000, 0x38490000, 0x38492000, 0x38494000, 0x38496000, 0x38498000, 0x3849A000, 0x3849C000, 0x3849E000, + 0x384A0000, 0x384A2000, 0x384A4000, 0x384A6000, 0x384A8000, 0x384AA000, 0x384AC000, 0x384AE000, 0x384B0000, 0x384B2000, 0x384B4000, 0x384B6000, 0x384B8000, 0x384BA000, 0x384BC000, 0x384BE000, + 0x384C0000, 0x384C2000, 0x384C4000, 0x384C6000, 0x384C8000, 0x384CA000, 0x384CC000, 0x384CE000, 0x384D0000, 0x384D2000, 0x384D4000, 0x384D6000, 0x384D8000, 0x384DA000, 0x384DC000, 0x384DE000, + 0x384E0000, 0x384E2000, 0x384E4000, 0x384E6000, 0x384E8000, 0x384EA000, 0x384EC000, 0x384EE000, 0x384F0000, 0x384F2000, 0x384F4000, 0x384F6000, 0x384F8000, 0x384FA000, 0x384FC000, 0x384FE000, + 0x38500000, 0x38502000, 0x38504000, 0x38506000, 0x38508000, 0x3850A000, 0x3850C000, 0x3850E000, 0x38510000, 0x38512000, 0x38514000, 0x38516000, 0x38518000, 0x3851A000, 0x3851C000, 0x3851E000, + 0x38520000, 0x38522000, 0x38524000, 0x38526000, 0x38528000, 0x3852A000, 0x3852C000, 0x3852E000, 0x38530000, 0x38532000, 0x38534000, 0x38536000, 0x38538000, 0x3853A000, 0x3853C000, 0x3853E000, + 0x38540000, 0x38542000, 0x38544000, 0x38546000, 0x38548000, 0x3854A000, 0x3854C000, 0x3854E000, 0x38550000, 0x38552000, 0x38554000, 0x38556000, 0x38558000, 0x3855A000, 0x3855C000, 0x3855E000, + 0x38560000, 0x38562000, 0x38564000, 0x38566000, 0x38568000, 0x3856A000, 0x3856C000, 0x3856E000, 0x38570000, 0x38572000, 0x38574000, 0x38576000, 0x38578000, 0x3857A000, 0x3857C000, 0x3857E000, + 0x38580000, 0x38582000, 0x38584000, 0x38586000, 0x38588000, 0x3858A000, 0x3858C000, 0x3858E000, 0x38590000, 0x38592000, 0x38594000, 0x38596000, 0x38598000, 0x3859A000, 0x3859C000, 0x3859E000, + 0x385A0000, 0x385A2000, 0x385A4000, 0x385A6000, 0x385A8000, 0x385AA000, 0x385AC000, 0x385AE000, 0x385B0000, 0x385B2000, 0x385B4000, 0x385B6000, 0x385B8000, 0x385BA000, 0x385BC000, 0x385BE000, + 0x385C0000, 0x385C2000, 0x385C4000, 0x385C6000, 0x385C8000, 0x385CA000, 0x385CC000, 0x385CE000, 0x385D0000, 0x385D2000, 0x385D4000, 0x385D6000, 0x385D8000, 0x385DA000, 0x385DC000, 0x385DE000, + 0x385E0000, 0x385E2000, 0x385E4000, 0x385E6000, 0x385E8000, 0x385EA000, 0x385EC000, 0x385EE000, 0x385F0000, 0x385F2000, 0x385F4000, 0x385F6000, 0x385F8000, 0x385FA000, 0x385FC000, 0x385FE000, + 0x38600000, 0x38602000, 0x38604000, 0x38606000, 0x38608000, 0x3860A000, 0x3860C000, 0x3860E000, 0x38610000, 0x38612000, 0x38614000, 0x38616000, 0x38618000, 0x3861A000, 0x3861C000, 0x3861E000, + 0x38620000, 0x38622000, 0x38624000, 0x38626000, 0x38628000, 0x3862A000, 0x3862C000, 0x3862E000, 0x38630000, 0x38632000, 0x38634000, 0x38636000, 0x38638000, 0x3863A000, 0x3863C000, 0x3863E000, + 0x38640000, 0x38642000, 0x38644000, 0x38646000, 0x38648000, 0x3864A000, 0x3864C000, 0x3864E000, 0x38650000, 0x38652000, 0x38654000, 0x38656000, 0x38658000, 0x3865A000, 0x3865C000, 0x3865E000, + 0x38660000, 0x38662000, 0x38664000, 0x38666000, 0x38668000, 0x3866A000, 0x3866C000, 0x3866E000, 0x38670000, 0x38672000, 0x38674000, 0x38676000, 0x38678000, 0x3867A000, 0x3867C000, 0x3867E000, + 0x38680000, 0x38682000, 0x38684000, 0x38686000, 0x38688000, 0x3868A000, 0x3868C000, 0x3868E000, 0x38690000, 0x38692000, 0x38694000, 0x38696000, 0x38698000, 0x3869A000, 0x3869C000, 0x3869E000, + 0x386A0000, 0x386A2000, 0x386A4000, 0x386A6000, 0x386A8000, 0x386AA000, 0x386AC000, 0x386AE000, 0x386B0000, 0x386B2000, 0x386B4000, 0x386B6000, 0x386B8000, 0x386BA000, 0x386BC000, 0x386BE000, + 0x386C0000, 0x386C2000, 0x386C4000, 0x386C6000, 0x386C8000, 0x386CA000, 0x386CC000, 0x386CE000, 0x386D0000, 0x386D2000, 0x386D4000, 0x386D6000, 0x386D8000, 0x386DA000, 0x386DC000, 0x386DE000, + 0x386E0000, 0x386E2000, 0x386E4000, 0x386E6000, 0x386E8000, 0x386EA000, 0x386EC000, 0x386EE000, 0x386F0000, 0x386F2000, 0x386F4000, 0x386F6000, 0x386F8000, 0x386FA000, 0x386FC000, 0x386FE000, + 0x38700000, 0x38702000, 0x38704000, 0x38706000, 0x38708000, 0x3870A000, 0x3870C000, 0x3870E000, 0x38710000, 0x38712000, 0x38714000, 0x38716000, 0x38718000, 0x3871A000, 0x3871C000, 0x3871E000, + 0x38720000, 0x38722000, 0x38724000, 0x38726000, 0x38728000, 0x3872A000, 0x3872C000, 0x3872E000, 0x38730000, 0x38732000, 0x38734000, 0x38736000, 0x38738000, 0x3873A000, 0x3873C000, 0x3873E000, + 0x38740000, 0x38742000, 0x38744000, 0x38746000, 0x38748000, 0x3874A000, 0x3874C000, 0x3874E000, 0x38750000, 0x38752000, 0x38754000, 0x38756000, 0x38758000, 0x3875A000, 0x3875C000, 0x3875E000, + 0x38760000, 0x38762000, 0x38764000, 0x38766000, 0x38768000, 0x3876A000, 0x3876C000, 0x3876E000, 0x38770000, 0x38772000, 0x38774000, 0x38776000, 0x38778000, 0x3877A000, 0x3877C000, 0x3877E000, + 0x38780000, 0x38782000, 0x38784000, 0x38786000, 0x38788000, 0x3878A000, 0x3878C000, 0x3878E000, 0x38790000, 0x38792000, 0x38794000, 0x38796000, 0x38798000, 0x3879A000, 0x3879C000, 0x3879E000, + 0x387A0000, 0x387A2000, 0x387A4000, 0x387A6000, 0x387A8000, 0x387AA000, 0x387AC000, 0x387AE000, 0x387B0000, 0x387B2000, 0x387B4000, 0x387B6000, 0x387B8000, 0x387BA000, 0x387BC000, 0x387BE000, + 0x387C0000, 0x387C2000, 0x387C4000, 0x387C6000, 0x387C8000, 0x387CA000, 0x387CC000, 0x387CE000, 0x387D0000, 0x387D2000, 0x387D4000, 0x387D6000, 0x387D8000, 0x387DA000, 0x387DC000, 0x387DE000, + 0x387E0000, 0x387E2000, 0x387E4000, 0x387E6000, 0x387E8000, 0x387EA000, 0x387EC000, 0x387EE000, 0x387F0000, 0x387F2000, 0x387F4000, 0x387F6000, 0x387F8000, 0x387FA000, 0x387FC000, 0x387FE000 }; + static const bits::type exponent_table[64] = { + 0x00000000, 0x00800000, 0x01000000, 0x01800000, 0x02000000, 0x02800000, 0x03000000, 0x03800000, 0x04000000, 0x04800000, 0x05000000, 0x05800000, 0x06000000, 0x06800000, 0x07000000, 0x07800000, + 0x08000000, 0x08800000, 0x09000000, 0x09800000, 0x0A000000, 0x0A800000, 0x0B000000, 0x0B800000, 0x0C000000, 0x0C800000, 0x0D000000, 0x0D800000, 0x0E000000, 0x0E800000, 0x0F000000, 0x47800000, + 0x80000000, 0x80800000, 0x81000000, 0x81800000, 0x82000000, 0x82800000, 0x83000000, 0x83800000, 0x84000000, 0x84800000, 0x85000000, 0x85800000, 0x86000000, 0x86800000, 0x87000000, 0x87800000, + 0x88000000, 0x88800000, 0x89000000, 0x89800000, 0x8A000000, 0x8A800000, 0x8B000000, 0x8B800000, 0x8C000000, 0x8C800000, 0x8D000000, 0x8D800000, 0x8E000000, 0x8E800000, 0x8F000000, 0xC7800000 }; + static const unsigned short offset_table[64] = { + 0, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, + 0, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024, 1024 }; + bits::type fbits = mantissa_table[offset_table[value>>10]+(value&0x3FF)] + exponent_table[value>>10]; + #endif + float out; + std::memcpy(&out, &fbits, sizeof(float)); + return out; + #endif + } + + /// Convert half-precision to IEEE double-precision. + /// \param value half-precision value to convert + /// \return double-precision value + inline double half2float_impl(unsigned int value, double, true_type) + { + #if HALF_ENABLE_F16C_INTRINSICS + return _mm_cvtsd_f64(_mm_cvtps_pd(_mm_cvtph_ps(_mm_cvtsi32_si128(value)))); + #else + uint32 hi = static_cast(value&0x8000) << 16; + unsigned int abs = value & 0x7FFF; + if(abs) + { + hi |= 0x3F000000 << static_cast(abs>=0x7C00); + for(; abs<0x400; abs<<=1,hi-=0x100000) ; + hi += static_cast(abs) << 10; + } + bits::type dbits = static_cast::type>(hi) << 32; + double out; + std::memcpy(&out, &dbits, sizeof(double)); + return out; + #endif + } + + /// Convert half-precision to non-IEEE floating-point. + /// \tparam T type to convert to (builtin integer type) + /// \param value half-precision value to convert + /// \return floating-point value + template T half2float_impl(unsigned int value, T, ...) + { + T out; + unsigned int abs = value & 0x7FFF; + if(abs > 0x7C00) + out = (std::numeric_limits::has_signaling_NaN && !(abs&0x200)) ? std::numeric_limits::signaling_NaN() : + std::numeric_limits::has_quiet_NaN ? std::numeric_limits::quiet_NaN() : T(); + else if(abs == 0x7C00) + out = std::numeric_limits::has_infinity ? std::numeric_limits::infinity() : std::numeric_limits::max(); + else if(abs > 0x3FF) + out = std::ldexp(static_cast((abs&0x3FF)|0x400), (abs>>10)-25); + else + out = std::ldexp(static_cast(abs), -24); + return (value&0x8000) ? -out : out; + } + + /// Convert half-precision to floating-point. + /// \tparam T type to convert to (builtin integer type) + /// \param value half-precision value to convert + /// \return floating-point value + template T half2float(unsigned int value) + { + return half2float_impl(value, T(), bool_type::is_iec559&&sizeof(typename bits::type)==sizeof(T)>()); + } + + /// Convert half-precision floating-point to integer. + /// \tparam R rounding mode to use + /// \tparam E `true` for round to even, `false` for round away from zero + /// \tparam I `true` to raise INEXACT exception (if inexact), `false` to never raise it + /// \tparam T type to convert to (buitlin integer type with at least 16 bits precision, excluding any implicit sign bits) + /// \param value half-precision value to convert + /// \return rounded integer value + /// \exception FE_INVALID if value is not representable in type \a T + /// \exception FE_INEXACT if value had to be rounded and \a I is `true` + template T half2int(unsigned int value) + { + unsigned int abs = value & 0x7FFF; + if(abs >= 0x7C00) + { + detail::raise(FE_INVALID); + return (value&0x8000) ? std::numeric_limits::min() : std::numeric_limits::max(); + } + if(abs < 0x3800) + { + detail::raise(FE_INEXACT, I); + return (R==std::round_toward_infinity) ? T(~(value>>15)&(abs!=0)) : + (R==std::round_toward_neg_infinity) ? -T(value>0x8000) : + T(); + } + int exp = 25 - (abs>>10); + unsigned int m = (value&0x3FF) | 0x400; + int32 i = static_cast((exp<=0) ? (m<<-exp) : ((m+( + (R==std::round_to_nearest) ? ((1<<(exp-1))-(~(m>>exp)&E)) : + (R==std::round_toward_infinity) ? (((1<>15)-1)) : + (R==std::round_toward_neg_infinity) ? (((1<>15)) : 0))>>exp)); + if((!std::numeric_limits::is_signed && (value&0x8000)) || (std::numeric_limits::digits<16 && + ((value&0x8000) ? (-i::min()) : (i>std::numeric_limits::max())))) + detail::raise(FE_INVALID); + else if(I && exp > 0 && (m&((1<((value&0x8000) ? -i : i); + } + + /// \} + /// \name Mathematics + /// \{ + + /// upper part of 64-bit multiplication. + /// \tparam R rounding mode to use + /// \param x first factor + /// \param y second factor + /// \return upper 32 bit of \a x * \a y + template uint32 mulhi(uint32 x, uint32 y) + { + uint32 xy = (x>>16) * (y&0xFFFF), yx = (x&0xFFFF) * (y>>16), c = (xy&0xFFFF) + (yx&0xFFFF) + (((x&0xFFFF)*(y&0xFFFF))>>16); + return (x>>16)*(y>>16) + (xy>>16) + (yx>>16) + (c>>16) + + ((R==std::round_to_nearest) ? ((c>>15)&1) : (R==std::round_toward_infinity) ? ((c&0xFFFF)!=0) : 0); + } + + /// 64-bit multiplication. + /// \param x first factor + /// \param y second factor + /// \return upper 32 bit of \a x * \a y rounded to nearest + inline uint32 multiply64(uint32 x, uint32 y) + { + #if HALF_ENABLE_CPP11_LONG_LONG + return static_cast((static_cast(x)*static_cast(y)+0x80000000)>>32); + #else + return mulhi(x, y); + #endif + } + + /// 64-bit division. + /// \param x upper 32 bit of dividend + /// \param y divisor + /// \param s variable to store sticky bit for rounding + /// \return (\a x << 32) / \a y + inline uint32 divide64(uint32 x, uint32 y, int &s) + { + #if HALF_ENABLE_CPP11_LONG_LONG + unsigned long long xx = static_cast(x) << 32; + return s = (xx%y!=0), static_cast(xx/y); + #else + y >>= 1; + uint32 rem = x, div = 0; + for(unsigned int i=0; i<32; ++i) + { + div <<= 1; + if(rem >= y) + { + rem -= y; + div |= 1; + } + rem <<= 1; + } + return s = rem > 1, div; + #endif + } + + /// Half precision positive modulus. + /// \tparam Q `true` to compute full quotient, `false` else + /// \tparam R `true` to compute signed remainder, `false` for positive remainder + /// \param x first operand as positive finite half-precision value + /// \param y second operand as positive finite half-precision value + /// \param quo adress to store quotient at, `nullptr` if \a Q `false` + /// \return modulus of \a x / \a y + template unsigned int mod(unsigned int x, unsigned int y, int *quo = NULL) + { + unsigned int q = 0; + if(x > y) + { + int absx = x, absy = y, expx = 0, expy = 0; + for(; absx<0x400; absx<<=1,--expx) ; + for(; absy<0x400; absy<<=1,--expy) ; + expx += absx >> 10; + expy += absy >> 10; + int mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400; + for(int d=expx-expy; d; --d) + { + if(!Q && mx == my) + return 0; + if(mx >= my) + { + mx -= my; + q += Q; + } + mx <<= 1; + q <<= static_cast(Q); + } + if(!Q && mx == my) + return 0; + if(mx >= my) + { + mx -= my; + ++q; + } + if(Q) + { + q &= (1<<(std::numeric_limits::digits-1)) - 1; + if(!mx) + return *quo = q, 0; + } + for(; mx<0x400; mx<<=1,--expy) ; + x = (expy>0) ? ((expy<<10)|(mx&0x3FF)) : (mx>>(1-expy)); + } + if(R) + { + unsigned int a, b; + if(y < 0x800) + { + a = (x<0x400) ? (x<<1) : (x+0x400); + b = y; + } + else + { + a = x; + b = y - 0x400; + } + if(a > b || (a == b && (q&1))) + { + int exp = (y>>10) + (y<=0x3FF), d = exp - (x>>10) - (x<=0x3FF); + int m = (((y&0x3FF)|((y>0x3FF)<<10))<<1) - (((x&0x3FF)|((x>0x3FF)<<10))<<(1-d)); + for(; m<0x800 && exp>1; m<<=1,--exp) ; + x = 0x8000 + ((exp-1)<<10) + (m>>1); + q += Q; + } + } + if(Q) + *quo = q; + return x; + } + + /// Fixed point square root. + /// \tparam F number of fractional bits + /// \param r radicand in Q1.F fixed point format + /// \param exp exponent + /// \return square root as Q1.F/2 + template uint32 sqrt(uint32 &r, int &exp) + { + int i = exp & 1; + r <<= i; + exp = (exp-i) / 2; + uint32 m = 0; + for(uint32 bit=static_cast(1)<>=2) + { + if(r < m+bit) + m >>= 1; + else + { + r -= m + bit; + m = (m>>1) + bit; + } + } + return m; + } + + /// Fixed point binary exponential. + /// This uses the BKM algorithm in E-mode. + /// \param m exponent in [0,1) as Q0.31 + /// \param n number of iterations (at most 32) + /// \return 2 ^ \a m as Q1.31 + inline uint32 exp2(uint32 m, unsigned int n = 32) + { + static const uint32 logs[] = { + 0x80000000, 0x4AE00D1D, 0x2934F098, 0x15C01A3A, 0x0B31FB7D, 0x05AEB4DD, 0x02DCF2D1, 0x016FE50B, + 0x00B84E23, 0x005C3E10, 0x002E24CA, 0x001713D6, 0x000B8A47, 0x0005C53B, 0x0002E2A3, 0x00017153, + 0x0000B8AA, 0x00005C55, 0x00002E2B, 0x00001715, 0x00000B8B, 0x000005C5, 0x000002E3, 0x00000171, + 0x000000B9, 0x0000005C, 0x0000002E, 0x00000017, 0x0000000C, 0x00000006, 0x00000003, 0x00000001 }; + if(!m) + return 0x80000000; + uint32 mx = 0x80000000, my = 0; + for(unsigned int i=1; i> i; + } + } + return mx; + } + + /// Fixed point binary logarithm. + /// This uses the BKM algorithm in L-mode. + /// \param m mantissa in [1,2) as Q1.30 + /// \param n number of iterations (at most 32) + /// \return log2(\a m) as Q0.31 + inline uint32 log2(uint32 m, unsigned int n = 32) + { + static const uint32 logs[] = { + 0x80000000, 0x4AE00D1D, 0x2934F098, 0x15C01A3A, 0x0B31FB7D, 0x05AEB4DD, 0x02DCF2D1, 0x016FE50B, + 0x00B84E23, 0x005C3E10, 0x002E24CA, 0x001713D6, 0x000B8A47, 0x0005C53B, 0x0002E2A3, 0x00017153, + 0x0000B8AA, 0x00005C55, 0x00002E2B, 0x00001715, 0x00000B8B, 0x000005C5, 0x000002E3, 0x00000171, + 0x000000B9, 0x0000005C, 0x0000002E, 0x00000017, 0x0000000C, 0x00000006, 0x00000003, 0x00000001 }; + if(m == 0x40000000) + return 0; + uint32 mx = 0x40000000, my = 0; + for(unsigned int i=1; i>i); + if(mz <= m) + { + mx = mz; + my += logs[i]; + } + } + return my; + } + + /// Fixed point sine and cosine. + /// This uses the CORDIC algorithm in rotation mode. + /// \param mz angle in [-pi/2,pi/2] as Q1.30 + /// \param n number of iterations (at most 31) + /// \return sine and cosine of \a mz as Q1.30 + inline std::pair sincos(uint32 mz, unsigned int n = 31) + { + static const uint32 angles[] = { + 0x3243F6A9, 0x1DAC6705, 0x0FADBAFD, 0x07F56EA7, 0x03FEAB77, 0x01FFD55C, 0x00FFFAAB, 0x007FFF55, + 0x003FFFEB, 0x001FFFFD, 0x00100000, 0x00080000, 0x00040000, 0x00020000, 0x00010000, 0x00008000, + 0x00004000, 0x00002000, 0x00001000, 0x00000800, 0x00000400, 0x00000200, 0x00000100, 0x00000080, + 0x00000040, 0x00000020, 0x00000010, 0x00000008, 0x00000004, 0x00000002, 0x00000001 }; + uint32 mx = 0x26DD3B6A, my = 0; + for(unsigned int i=0; i0x3FF)<<10); + int exp = (abs>>10) + (abs<=0x3FF) - 15; + if(abs < 0x3A48) + return k = 0, m << (exp+20); + #if HALF_ENABLE_CPP11_LONG_LONG + unsigned long long y = m * 0xA2F9836E4E442, mask = (1ULL<<(62-exp)) - 1, yi = (y+(mask>>1)) & ~mask, f = y - yi; + uint32 sign = -static_cast(f>>63); + k = static_cast(yi>>(62-exp)); + return (multiply64(static_cast((sign ? -f : f)>>(31-exp)), 0xC90FDAA2)^sign) - sign; + #else + uint32 yh = m*0xA2F98 + mulhi(m, 0x36E4E442), yl = (m*0x36E4E442) & 0xFFFFFFFF; + uint32 mask = (static_cast(1)<<(30-exp)) - 1, yi = (yh+(mask>>1)) & ~mask, sign = -static_cast(yi>yh); + k = static_cast(yi>>(30-exp)); + uint32 fh = (yh^sign) + (yi^~sign) - ~sign, fl = (yl^sign) - sign; + return (multiply64((exp>-1) ? (((fh<<(1+exp))&0xFFFFFFFF)|((fl&0xFFFFFFFF)>>(31-exp))) : fh, 0xC90FDAA2)^sign) - sign; + #endif + } + + /// Get arguments for atan2 function. + /// \param abs half-precision floating-point value + /// \return \a abs and sqrt(1 - \a abs^2) as Q0.30 + inline std::pair atan2_args(unsigned int abs) + { + int exp = -15; + for(; abs<0x400; abs<<=1,--exp) ; + exp += abs >> 10; + uint32 my = ((abs&0x3FF)|0x400) << 5, r = my * my; + int rexp = 2 * exp; + r = 0x40000000 - ((rexp>-31) ? ((r>>-rexp)|((r&((static_cast(1)<<-rexp)-1))!=0)) : 1); + for(rexp=0; r<0x40000000; r<<=1,--rexp) ; + uint32 mx = sqrt<30>(r, rexp); + int d = exp - rexp; + if(d < 0) + return std::make_pair((d<-14) ? ((my>>(-d-14))+((my>>(-d-15))&1)) : (my<<(14+d)), (mx<<14)+(r<<13)/mx); + if(d > 0) + return std::make_pair(my<<14, (d>14) ? ((mx>>(d-14))+((mx>>(d-15))&1)) : ((d==14) ? mx : ((mx<<(14-d))+(r<<(13-d))/mx))); + return std::make_pair(my<<13, (mx<<13)+(r<<12)/mx); + } + + /// Get exponentials for hyperbolic computation + /// \param abs half-precision floating-point value + /// \param exp variable to take unbiased exponent of larger result + /// \param n number of BKM iterations (at most 32) + /// \return exp(abs) and exp(-\a abs) as Q1.31 with same exponent + inline std::pair hyperbolic_args(unsigned int abs, int &exp, unsigned int n = 32) + { + uint32 mx = detail::multiply64(static_cast((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29), my; + int e = (abs>>10) + (abs<=0x3FF); + if(e < 14) + { + exp = 0; + mx >>= 14 - e; + } + else + { + exp = static_cast(mx >> (45-e)); + mx = (mx<<(e-14)) & 0x7FFFFFFF; + } + mx = exp2(mx, n); + int d = exp << 1, s; + if(mx > 0x80000000) + { + my = divide64(0x80000000, mx, s); + my |= s; + ++d; + } + else + my = mx; + return std::make_pair(mx, (d<31) ? ((my>>d)|((my&((static_cast(1)< unsigned int exp2_post(uint32 m, int exp, bool esign, unsigned int sign = 0, unsigned int n = 32) + { + if(esign) + { + exp = -exp - (m!=0); + if(exp < -25) + return underflow(sign); + else if(exp == -25) + return rounded(sign, 1, m!=0); + } + else if(exp > 15) + return overflow(sign); + if(!m) + return sign | (((exp+=15)>0) ? (exp<<10) : check_underflow(0x200>>-exp)); + m = exp2(m, n); + int s = 0; + if(esign) + m = divide64(0x80000000, m, s); + return fixed2half(m, exp+14, sign, s); + } + + /// Postprocessing for binary logarithm. + /// \tparam R rounding mode to use + /// \tparam L logarithm for base transformation as Q1.31 + /// \param m fractional part of logarithm as Q0.31 + /// \param ilog signed integer part of logarithm + /// \param exp biased exponent of result + /// \param sign sign bit of result + /// \return value base-transformed and converted to half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if no other exception occurred + template unsigned int log2_post(uint32 m, int ilog, int exp, unsigned int sign = 0) + { + uint32 msign = sign_mask(ilog); + m = (((static_cast(ilog)<<27)+(m>>4))^msign) - msign; + if(!m) + return 0; + for(; m<0x80000000; m<<=1,--exp) ; + int i = m >= L, s; + exp += i; + m >>= 1 + i; + sign ^= msign & 0x8000; + if(exp < -11) + return underflow(sign); + m = divide64(m, L, s); + return fixed2half(m, exp, sign, 1); + } + + /// Hypotenuse square root and postprocessing. + /// \tparam R rounding mode to use + /// \param r mantissa as Q2.30 + /// \param exp biased exponent + /// \return square root converted to half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if value had to be rounded + template unsigned int hypot_post(uint32 r, int exp) + { + int i = static_cast(r >> 31); + if((exp+=i) > 46) + return overflow(); + if(exp < -34) + return underflow(); + r = (r>>i) | (r&i); + uint32 m = sqrt<30>(r, exp+=15); + return fixed2half(m, exp-1, 0, r!=0); + } + + /// Division and postprocessing for tangents. + /// \tparam R rounding mode to use + /// \param my dividend as Q1.31 + /// \param mx divisor as Q1.31 + /// \param exp biased exponent of result + /// \param sign sign bit of result + /// \return quotient converted to half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if no other exception occurred + template unsigned int tangent_post(uint32 my, uint32 mx, int exp, unsigned int sign = 0) + { + int i = my >= mx, s; + exp += i; + if(exp > 29) + return overflow(sign); + if(exp < -11) + return underflow(sign); + uint32 m = divide64(my>>(i+1), mx, s); + return fixed2half(m, exp, sign, s); + } + + /// Area function and postprocessing. + /// This computes the value directly in Q2.30 using the representation `asinh|acosh(x) = log(x+sqrt(x^2+|-1))`. + /// \tparam R rounding mode to use + /// \tparam S `true` for asinh, `false` for acosh + /// \param arg half-precision argument + /// \return asinh|acosh(\a arg) converted to half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if no other exception occurred + template unsigned int area(unsigned int arg) + { + int abs = arg & 0x7FFF, expx = (abs>>10) + (abs<=0x3FF) - 15, expy = -15, ilog; + uint32 mx = static_cast((abs&0x3FF)|((abs>0x3FF)<<10)) << 20, my, r; + for(; abs<0x400; abs<<=1,--expy) ; + expy += abs >> 10; + r = ((abs&0x3FF)|0x400) << 5; + r *= r; + int i = static_cast(r >> 31); + expy = 2*expy + i; + r >>= i; + if(S) + { + if(expy < 0) + { + r = 0x40000000 + ((expy>-30) ? ((r>>-expy)|((r&((static_cast(1)<<-expy)-1))!=0)) : 1); + expy = 0; + } + else + { + r += 0x40000000 >> expy; + i = static_cast(r >> 31); + r = (r>>i) | (r&i); + expy += i; + } + } + else + { + r -= 0x40000000 >> expy; + for(; r<0x40000000; r<<=1,--expy) ; + } + my = sqrt<30>(r, expy); + my = (my<<15) + (r<<14)/my; + if(S) + { + mx >>= expy - expx; + ilog = expy; + } + else + { + my >>= expx - expy; + ilog = expx; + } + my += mx; + i = static_cast(my >> 31); + static const int G = S && (R==std::round_to_nearest); + return log2_post(log2(my>>i, 26+S+G)+(G<<3), ilog+i, 17, arg&(static_cast(S)<<15)); + } + + /// Class for 1.31 unsigned floating-point computation + struct f31 + { + /// Constructor. + /// \param mant mantissa as 1.31 + /// \param e exponent + HALF_CONSTEXPR f31(uint32 mant, int e) : m(mant), exp(e) {} + + /// Constructor. + /// \param abs unsigned half-precision value + f31(unsigned int abs) : exp(-15) + { + for(; abs<0x400; abs<<=1,--exp) ; + m = static_cast((abs&0x3FF)|0x400) << 21; + exp += (abs>>10); + } + + /// Addition operator. + /// \param a first operand + /// \param b second operand + /// \return \a a + \a b + friend f31 operator+(f31 a, f31 b) + { + if(b.exp > a.exp) + std::swap(a, b); + int d = a.exp - b.exp; + uint32 m = a.m + ((d<32) ? (b.m>>d) : 0); + int i = (m&0xFFFFFFFF) < a.m; + return f31(((m+i)>>i)|0x80000000, a.exp+i); + } + + /// Subtraction operator. + /// \param a first operand + /// \param b second operand + /// \return \a a - \a b + friend f31 operator-(f31 a, f31 b) + { + int d = a.exp - b.exp, exp = a.exp; + uint32 m = a.m - ((d<32) ? (b.m>>d) : 0); + if(!m) + return f31(0, -32); + for(; m<0x80000000; m<<=1,--exp) ; + return f31(m, exp); + } + + /// Multiplication operator. + /// \param a first operand + /// \param b second operand + /// \return \a a * \a b + friend f31 operator*(f31 a, f31 b) + { + uint32 m = multiply64(a.m, b.m); + int i = static_cast(m >> 31); + return f31(m<<(1-i), a.exp + b.exp + i); + } + + /// Division operator. + /// \param a first operand + /// \param b second operand + /// \return \a a / \a b + friend f31 operator/(f31 a, f31 b) + { + int i = a.m >= b.m, s; + uint32 m = divide64((a.m+i)>>i, b.m, s); + return f31(m, a.exp - b.exp + i - 1); + } + + uint32 m; ///< mantissa as 1.31. + int exp; ///< exponent. + }; + + /// Error function and postprocessing. + /// This computes the value directly in Q1.31 using the approximations given + /// [here](https://en.wikipedia.org/wiki/Error_function#Approximation_with_elementary_functions). + /// \tparam R rounding mode to use + /// \tparam C `true` for comlementary error function, `false` else + /// \param arg half-precision function argument + /// \return approximated value of error function in half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if no other exception occurred + template unsigned int erf(unsigned int arg) + { + unsigned int abs = arg & 0x7FFF, sign = arg & 0x8000; + f31 x(abs), x2 = x * x * f31(0xB8AA3B29, 0), t = f31(0x80000000, 0) / (f31(0x80000000, 0)+f31(0xA7BA054A, -2)*x), t2 = t * t; + f31 e = ((f31(0x87DC2213, 0)*t2+f31(0xB5F0E2AE, 0))*t2+f31(0x82790637, -2)-(f31(0xBA00E2B8, 0)*t2+f31(0x91A98E62, -2))*t) * t / + ((x2.exp<0) ? f31(exp2((x2.exp>-32) ? (x2.m>>-x2.exp) : 0, 30), 0) : f31(exp2((x2.m<(x2.m>>(31-x2.exp)))); + return (!C || sign) ? fixed2half(0x80000000-(e.m>>(C-e.exp)), 14+C, sign&(C-1U)) : + (e.exp<-25) ? underflow() : fixed2half(e.m>>1, e.exp+14, 0, e.m&1); + } + + /// Gamma function and postprocessing. + /// This approximates the value of either the gamma function or its logarithm directly in Q1.31. + /// \tparam R rounding mode to use + /// \tparam L `true` for lograithm of gamma function, `false` for gamma function + /// \param arg half-precision floating-point value + /// \return lgamma/tgamma(\a arg) in half-precision + /// \exception FE_OVERFLOW on overflows + /// \exception FE_UNDERFLOW on underflows + /// \exception FE_INEXACT if \a arg is not a positive integer + template unsigned int gamma(unsigned int arg) + { +/* static const double p[] ={ 2.50662827563479526904, 225.525584619175212544, -268.295973841304927459, 80.9030806934622512966, -5.00757863970517583837, 0.0114684895434781459556 }; + double t = arg + 4.65, s = p[0]; + for(unsigned int i=0; i<5; ++i) + s += p[i+1] / (arg+i); + return std::log(s) + (arg-0.5)*std::log(t) - t; +*/ static const f31 pi(0xC90FDAA2, 1), lbe(0xB8AA3B29, 0); + unsigned int abs = arg & 0x7FFF, sign = arg & 0x8000; + bool bsign = sign != 0; + f31 z(abs), x = sign ? (z+f31(0x80000000, 0)) : z, t = x + f31(0x94CCCCCD, 2), s = + f31(0xA06C9901, 1) + f31(0xBBE654E2, -7)/(x+f31(0x80000000, 2)) + f31(0xA1CE6098, 6)/(x+f31(0x80000000, 1)) + + f31(0xE1868CB7, 7)/x - f31(0x8625E279, 8)/(x+f31(0x80000000, 0)) - f31(0xA03E158F, 2)/(x+f31(0xC0000000, 1)); + int i = (s.exp>=2) + (s.exp>=4) + (s.exp>=8) + (s.exp>=16); + s = f31((static_cast(s.exp)<<(31-i))+(log2(s.m>>1, 28)>>i), i) / lbe; + if(x.exp != -1 || x.m != 0x80000000) + { + i = (t.exp>=2) + (t.exp>=4) + (t.exp>=8); + f31 l = f31((static_cast(t.exp)<<(31-i))+(log2(t.m>>1, 30)>>i), i) / lbe; + s = (x.exp<-1) ? (s-(f31(0x80000000, -1)-x)*l) : (s+(x-f31(0x80000000, -1))*l); + } + s = x.exp ? (s-t) : (t-s); + if(bsign) + { + if(z.exp >= 0) + { + sign &= (L|((z.m>>(31-z.exp))&1)) - 1; + for(z=f31((z.m<<(1+z.exp))&0xFFFFFFFF, -1); z.m<0x80000000; z.m<<=1,--z.exp) ; + } + if(z.exp == -1) + z = f31(0x80000000, 0) - z; + if(z.exp < -1) + { + z = z * pi; + z.m = sincos(z.m>>(1-z.exp), 30).first; + for(z.exp=1; z.m<0x80000000; z.m<<=1,--z.exp) ; + } + else + z = f31(0x80000000, 0); + } + if(L) + { + if(bsign) + { + f31 l(0x92868247, 0); + if(z.exp < 0) + { + uint32 m = log2((z.m+1)>>1, 27); + z = f31(-((static_cast(z.exp)<<26)+(m>>5)), 5); + for(; z.m<0x80000000; z.m<<=1,--z.exp) ; + l = l + z / lbe; + } + sign = static_cast(x.exp&&(l.exp(x.exp==0) << 15; + if(s.exp < -24) + return underflow(sign); + if(s.exp > 15) + return overflow(sign); + } + } + else + { + s = s * lbe; + uint32 m; + if(s.exp < 0) + { + m = s.m >> -s.exp; + s.exp = 0; + } + else + { + m = (s.m<(s.m>>(31-s.exp)); + } + s.m = exp2(m, 27); + if(!x.exp) + s = f31(0x80000000, 0) / s; + if(bsign) + { + if(z.exp < 0) + s = s * z; + s = pi / s; + if(s.exp < -24) + return underflow(sign); + } + else if(z.exp > 0 && !(z.m&((1<<(31-z.exp))-1))) + return ((s.exp+14)<<10) + static_cast(s.m>>21); + if(s.exp > 15) + return overflow(sign); + } + return fixed2half(s.m, s.exp+14, sign); + } + /// \} + + template struct half_caster; + } + + /// Half-precision floating-point type. + /// This class implements an IEEE-conformant half-precision floating-point type with the usual arithmetic + /// operators and conversions. It is implicitly convertible to single-precision floating-point, which makes artihmetic + /// expressions and functions with mixed-type operands to be of the most precise operand type. + /// + /// According to the C++98/03 definition, the half type is not a POD type. But according to C++11's less strict and + /// extended definitions it is both a standard layout type and a trivially copyable type (even if not a POD type), which + /// means it can be standard-conformantly copied using raw binary copies. But in this context some more words about the + /// actual size of the type. Although the half is representing an IEEE 16-bit type, it does not neccessarily have to be of + /// exactly 16-bits size. But on any reasonable implementation the actual binary representation of this type will most + /// probably not ivolve any additional "magic" or padding beyond the simple binary representation of the underlying 16-bit + /// IEEE number, even if not strictly guaranteed by the standard. But even then it only has an actual size of 16 bits if + /// your C++ implementation supports an unsigned integer type of exactly 16 bits width. But this should be the case on + /// nearly any reasonable platform. + /// + /// So if your C++ implementation is not totally exotic or imposes special alignment requirements, it is a reasonable + /// assumption that the data of a half is just comprised of the 2 bytes of the underlying IEEE representation. + class half + { + public: + /// \name Construction and assignment + /// \{ + + /// Default constructor. + /// This initializes the half to 0. Although this does not match the builtin types' default-initialization semantics + /// and may be less efficient than no initialization, it is needed to provide proper value-initialization semantics. + HALF_CONSTEXPR half() HALF_NOEXCEPT : data_() {} + + /// Conversion constructor. + /// \param rhs float to convert + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + explicit half(float rhs) : data_(static_cast(detail::float2half(rhs))) {} + + /// Conversion to single-precision. + /// \return single precision value representing expression value + operator float() const { return detail::half2float(data_); } + + /// Assignment operator. + /// \param rhs single-precision value to copy from + /// \return reference to this half + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + half& operator=(float rhs) { data_ = static_cast(detail::float2half(rhs)); return *this; } + + /// \} + /// \name Arithmetic updates + /// \{ + + /// Arithmetic assignment. + /// \tparam T type of concrete half expression + /// \param rhs half expression to add + /// \return reference to this half + /// \exception FE_... according to operator+(half,half) + half& operator+=(half rhs) { return *this = *this + rhs; } + + /// Arithmetic assignment. + /// \tparam T type of concrete half expression + /// \param rhs half expression to subtract + /// \return reference to this half + /// \exception FE_... according to operator-(half,half) + half& operator-=(half rhs) { return *this = *this - rhs; } + + /// Arithmetic assignment. + /// \tparam T type of concrete half expression + /// \param rhs half expression to multiply with + /// \return reference to this half + /// \exception FE_... according to operator*(half,half) + half& operator*=(half rhs) { return *this = *this * rhs; } + + /// Arithmetic assignment. + /// \tparam T type of concrete half expression + /// \param rhs half expression to divide by + /// \return reference to this half + /// \exception FE_... according to operator/(half,half) + half& operator/=(half rhs) { return *this = *this / rhs; } + + /// Arithmetic assignment. + /// \param rhs single-precision value to add + /// \return reference to this half + /// \exception FE_... according to operator=() + half& operator+=(float rhs) { return *this = *this + rhs; } + + /// Arithmetic assignment. + /// \param rhs single-precision value to subtract + /// \return reference to this half + /// \exception FE_... according to operator=() + half& operator-=(float rhs) { return *this = *this - rhs; } + + /// Arithmetic assignment. + /// \param rhs single-precision value to multiply with + /// \return reference to this half + /// \exception FE_... according to operator=() + half& operator*=(float rhs) { return *this = *this * rhs; } + + /// Arithmetic assignment. + /// \param rhs single-precision value to divide by + /// \return reference to this half + /// \exception FE_... according to operator=() + half& operator/=(float rhs) { return *this = *this / rhs; } + + /// \} + /// \name Increment and decrement + /// \{ + + /// Prefix increment. + /// \return incremented half value + /// \exception FE_... according to operator+(half,half) + half& operator++() { return *this = *this + half(detail::binary, 0x3C00); } + + /// Prefix decrement. + /// \return decremented half value + /// \exception FE_... according to operator-(half,half) + half& operator--() { return *this = *this + half(detail::binary, 0xBC00); } + + /// Postfix increment. + /// \return non-incremented half value + /// \exception FE_... according to operator+(half,half) + half operator++(int) { half out(*this); ++*this; return out; } + + /// Postfix decrement. + /// \return non-decremented half value + /// \exception FE_... according to operator-(half,half) + half operator--(int) { half out(*this); --*this; return out; } + /// \} + + private: + /// Rounding mode to use + static const std::float_round_style round_style = (std::float_round_style)(HALF_ROUND_STYLE); + + /// Constructor. + /// \param bits binary representation to set half to + HALF_CONSTEXPR half(detail::binary_t, unsigned int bits) HALF_NOEXCEPT : data_(static_cast(bits)) {} + + /// Internal binary representation + detail::uint16 data_; + + #ifndef HALF_DOXYGEN_ONLY + friend HALF_CONSTEXPR_NOERR bool operator==(half, half); + friend HALF_CONSTEXPR_NOERR bool operator!=(half, half); + friend HALF_CONSTEXPR_NOERR bool operator<(half, half); + friend HALF_CONSTEXPR_NOERR bool operator>(half, half); + friend HALF_CONSTEXPR_NOERR bool operator<=(half, half); + friend HALF_CONSTEXPR_NOERR bool operator>=(half, half); + friend HALF_CONSTEXPR half operator-(half); + friend half operator+(half, half); + friend half operator-(half, half); + friend half operator*(half, half); + friend half operator/(half, half); + template friend std::basic_ostream& operator<<(std::basic_ostream&, half); + template friend std::basic_istream& operator>>(std::basic_istream&, half&); + friend HALF_CONSTEXPR half fabs(half); + friend half fmod(half, half); + friend half remainder(half, half); + friend half remquo(half, half, int*); + friend half fma(half, half, half); + friend HALF_CONSTEXPR_NOERR half fmax(half, half); + friend HALF_CONSTEXPR_NOERR half fmin(half, half); + friend half fdim(half, half); + friend half nanh(const char*); + friend half exp(half); + friend half exp2(half); + friend half expm1(half); + friend half log(half); + friend half log10(half); + friend half log2(half); + friend half log1p(half); + friend half sqrt(half); + friend half rsqrt(half); + friend half cbrt(half); + friend half hypot(half, half); + friend half hypot(half, half, half); + friend half pow(half, half); + friend void sincos(half, half*, half*); + friend half sin(half); + friend half cos(half); + friend half tan(half); + friend half asin(half); + friend half acos(half); + friend half atan(half); + friend half atan2(half, half); + friend half sinh(half); + friend half cosh(half); + friend half tanh(half); + friend half asinh(half); + friend half acosh(half); + friend half atanh(half); + friend half erf(half); + friend half erfc(half); + friend half lgamma(half); + friend half tgamma(half); + friend half ceil(half); + friend half floor(half); + friend half trunc(half); + friend half round(half); + friend long lround(half); + friend half rint(half); + friend long lrint(half); + friend half nearbyint(half); + #ifdef HALF_ENABLE_CPP11_LONG_LONG + friend long long llround(half); + friend long long llrint(half); + #endif + friend half frexp(half, int*); + friend half scalbln(half, long); + friend half modf(half, half*); + friend int ilogb(half); + friend half logb(half); + friend half nextafter(half, half); + friend half nexttoward(half, long double); + friend HALF_CONSTEXPR half copysign(half, half); + friend HALF_CONSTEXPR int fpclassify(half); + friend HALF_CONSTEXPR bool isfinite(half); + friend HALF_CONSTEXPR bool isinf(half); + friend HALF_CONSTEXPR bool isnan(half); + friend HALF_CONSTEXPR bool isnormal(half); + friend HALF_CONSTEXPR bool signbit(half); + friend HALF_CONSTEXPR bool isgreater(half, half); + friend HALF_CONSTEXPR bool isgreaterequal(half, half); + friend HALF_CONSTEXPR bool isless(half, half); + friend HALF_CONSTEXPR bool islessequal(half, half); + friend HALF_CONSTEXPR bool islessgreater(half, half); + template friend struct detail::half_caster; + friend class std::numeric_limits; + #if HALF_ENABLE_CPP11_HASH + friend struct std::hash; + #endif + #if HALF_ENABLE_CPP11_USER_LITERALS + friend half literal::operator "" _h(long double); + #endif + #endif + }; + +#if HALF_ENABLE_CPP11_USER_LITERALS + namespace literal + { + /// Half literal. + /// While this returns a properly rounded half-precision value, half literals can unfortunately not be constant + /// expressions due to rather involved conversions. So don't expect this to be a literal literal without involving + /// conversion operations at runtime. It is a convenience feature, not a performance optimization. + /// \param value literal value + /// \return half with of given value (possibly rounded) + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half operator "" _h(long double value) { return half(detail::binary, detail::float2half(value)); } + } +#endif + + namespace detail + { + /// Helper class for half casts. + /// This class template has to be specialized for all valid cast arguments to define an appropriate static + /// `cast` member function and a corresponding `type` member denoting its return type. + /// \tparam T destination type + /// \tparam U source type + /// \tparam R rounding mode to use + template struct half_caster {}; + template struct half_caster + { + #if HALF_ENABLE_CPP11_STATIC_ASSERT && HALF_ENABLE_CPP11_TYPE_TRAITS + static_assert(std::is_arithmetic::value, "half_cast from non-arithmetic type not supported"); + #endif + + static half cast(U arg) { return cast_impl(arg, is_float()); }; + + private: + static half cast_impl(U arg, true_type) { return half(binary, float2half(arg)); } + static half cast_impl(U arg, false_type) { return half(binary, int2half(arg)); } + }; + template struct half_caster + { + #if HALF_ENABLE_CPP11_STATIC_ASSERT && HALF_ENABLE_CPP11_TYPE_TRAITS + static_assert(std::is_arithmetic::value, "half_cast to non-arithmetic type not supported"); + #endif + + static T cast(half arg) { return cast_impl(arg, is_float()); } + + private: + static T cast_impl(half arg, true_type) { return half2float(arg.data_); } + static T cast_impl(half arg, false_type) { return half2int(arg.data_); } + }; + template struct half_caster + { + static half cast(half arg) { return arg; } + }; + } +} + +/// Extensions to the C++ standard library. +namespace std +{ + /// Numeric limits for half-precision floats. + /// **See also:** Documentation for [std::numeric_limits](https://en.cppreference.com/w/cpp/types/numeric_limits) + template<> class numeric_limits + { + public: + /// Is template specialization. + static HALF_CONSTEXPR_CONST bool is_specialized = true; + + /// Supports signed values. + static HALF_CONSTEXPR_CONST bool is_signed = true; + + /// Is not an integer type. + static HALF_CONSTEXPR_CONST bool is_integer = false; + + /// Is not exact. + static HALF_CONSTEXPR_CONST bool is_exact = false; + + /// Doesn't provide modulo arithmetic. + static HALF_CONSTEXPR_CONST bool is_modulo = false; + + /// Has a finite set of values. + static HALF_CONSTEXPR_CONST bool is_bounded = true; + + /// IEEE conformant. + static HALF_CONSTEXPR_CONST bool is_iec559 = true; + + /// Supports infinity. + static HALF_CONSTEXPR_CONST bool has_infinity = true; + + /// Supports quiet NaNs. + static HALF_CONSTEXPR_CONST bool has_quiet_NaN = true; + + /// Supports signaling NaNs. + static HALF_CONSTEXPR_CONST bool has_signaling_NaN = true; + + /// Supports subnormal values. + static HALF_CONSTEXPR_CONST float_denorm_style has_denorm = denorm_present; + + /// Does not support denormalization detection. + static HALF_CONSTEXPR_CONST bool has_denorm_loss = false; + + #if HALF_ERRHANDLING_THROWS + static HALF_CONSTEXPR_CONST bool traps = true; + #else + /// Traps only if [HALF_ERRHANDLING_THROW_...](\ref HALF_ERRHANDLING_THROW_INVALID) is acitvated. + static HALF_CONSTEXPR_CONST bool traps = false; + #endif + + /// Does not support pre-rounding underflow detection. + static HALF_CONSTEXPR_CONST bool tinyness_before = false; + + /// Rounding mode. + static HALF_CONSTEXPR_CONST float_round_style round_style = half_float::half::round_style; + + /// Significant digits. + static HALF_CONSTEXPR_CONST int digits = 11; + + /// Significant decimal digits. + static HALF_CONSTEXPR_CONST int digits10 = 3; + + /// Required decimal digits to represent all possible values. + static HALF_CONSTEXPR_CONST int max_digits10 = 5; + + /// Number base. + static HALF_CONSTEXPR_CONST int radix = 2; + + /// One more than smallest exponent. + static HALF_CONSTEXPR_CONST int min_exponent = -13; + + /// Smallest normalized representable power of 10. + static HALF_CONSTEXPR_CONST int min_exponent10 = -4; + + /// One more than largest exponent + static HALF_CONSTEXPR_CONST int max_exponent = 16; + + /// Largest finitely representable power of 10. + static HALF_CONSTEXPR_CONST int max_exponent10 = 4; + + /// Smallest positive normal value. + static HALF_CONSTEXPR half_float::half min() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x0400); } + + /// Smallest finite value. + static HALF_CONSTEXPR half_float::half lowest() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0xFBFF); } + + /// Largest finite value. + static HALF_CONSTEXPR half_float::half max() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7BFF); } + + /// Difference between 1 and next representable value. + static HALF_CONSTEXPR half_float::half epsilon() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x1400); } + + /// Maximum rounding error in ULP (units in the last place). + static HALF_CONSTEXPR half_float::half round_error() HALF_NOTHROW + { return half_float::half(half_float::detail::binary, (round_style==std::round_to_nearest) ? 0x3800 : 0x3C00); } + + /// Positive infinity. + static HALF_CONSTEXPR half_float::half infinity() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7C00); } + + /// Quiet NaN. + static HALF_CONSTEXPR half_float::half quiet_NaN() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7FFF); } + + /// Signaling NaN. + static HALF_CONSTEXPR half_float::half signaling_NaN() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x7DFF); } + + /// Smallest positive subnormal value. + static HALF_CONSTEXPR half_float::half denorm_min() HALF_NOTHROW { return half_float::half(half_float::detail::binary, 0x0001); } + }; + +#if HALF_ENABLE_CPP11_HASH + /// Hash function for half-precision floats. + /// This is only defined if C++11 `std::hash` is supported and enabled. + /// + /// **See also:** Documentation for [std::hash](https://en.cppreference.com/w/cpp/utility/hash) + template<> struct hash + { + /// Type of function argument. + typedef half_float::half argument_type; + + /// Function return type. + typedef size_t result_type; + + /// Compute hash function. + /// \param arg half to hash + /// \return hash value + result_type operator()(argument_type arg) const { return hash()(arg.data_&-static_cast(arg.data_!=0x8000)); } + }; +#endif +} + +namespace half_float +{ + /// \anchor compop + /// \name Comparison operators + /// \{ + + /// Comparison for equality. + /// \param x first operand + /// \param y second operand + /// \retval true if operands equal + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool operator==(half x, half y) + { + return !detail::compsignal(x.data_, y.data_) && (x.data_==y.data_ || !((x.data_|y.data_)&0x7FFF)); + } + + /// Comparison for inequality. + /// \param x first operand + /// \param y second operand + /// \retval true if operands not equal + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool operator!=(half x, half y) + { + return detail::compsignal(x.data_, y.data_) || (x.data_!=y.data_ && ((x.data_|y.data_)&0x7FFF)); + } + + /// Comparison for less than. + /// \param x first operand + /// \param y second operand + /// \retval true if \a x less than \a y + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool operator<(half x, half y) + { + return !detail::compsignal(x.data_, y.data_) && + ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) < ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)); + } + + /// Comparison for greater than. + /// \param x first operand + /// \param y second operand + /// \retval true if \a x greater than \a y + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool operator>(half x, half y) + { + return !detail::compsignal(x.data_, y.data_) && + ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) > ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)); + } + + /// Comparison for less equal. + /// \param x first operand + /// \param y second operand + /// \retval true if \a x less equal \a y + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool operator<=(half x, half y) + { + return !detail::compsignal(x.data_, y.data_) && + ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) <= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)); + } + + /// Comparison for greater equal. + /// \param x first operand + /// \param y second operand + /// \retval true if \a x greater equal \a y + /// \retval false else + /// \exception FE_INVALID if \a x or \a y is NaN + inline HALF_CONSTEXPR_NOERR bool operator>=(half x, half y) + { + return !detail::compsignal(x.data_, y.data_) && + ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) >= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)); + } + + /// \} + /// \anchor arithmetics + /// \name Arithmetic operators + /// \{ + + /// Identity. + /// \param arg operand + /// \return unchanged operand + inline HALF_CONSTEXPR half operator+(half arg) { return arg; } + + /// Negation. + /// \param arg operand + /// \return negated operand + inline HALF_CONSTEXPR half operator-(half arg) { return half(detail::binary, arg.data_^0x8000); } + + /// Addition. + /// This operation is exact to rounding for all rounding modes. + /// \param x left operand + /// \param y right operand + /// \return sum of half expressions + /// \exception FE_INVALID if \a x and \a y are infinities with different signs or signaling NaNs + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half operator+(half x, half y) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(detail::half2float(x.data_)+detail::half2float(y.data_))); + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF; + bool sub = ((x.data_^y.data_)&0x8000) != 0; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : (absy!=0x7C00) ? x.data_ : + (sub && absx==0x7C00) ? detail::invalid() : y.data_); + if(!absx) + return absy ? y : half(detail::binary, (half::round_style==std::round_toward_neg_infinity) ? (x.data_|y.data_) : (x.data_&y.data_)); + if(!absy) + return x; + unsigned int sign = ((sub && absy>absx) ? y.data_ : x.data_) & 0x8000; + if(absy > absx) + std::swap(absx, absy); + int exp = (absx>>10) + (absx<=0x3FF), d = exp - (absy>>10) - (absy<=0x3FF), mx = ((absx&0x3FF)|((absx>0x3FF)<<10)) << 3, my; + if(d < 13) + { + my = ((absy&0x3FF)|((absy>0x3FF)<<10)) << 3; + my = (my>>d) | ((my&((1<(half::round_style==std::round_toward_neg_infinity)<<15); + for(; mx<0x2000 && exp>1; mx<<=1,--exp) ; + } + else + { + mx += my; + int i = mx >> 14; + if((exp+=i) > 30) + return half(detail::binary, detail::overflow(sign)); + mx = (mx>>i) | (mx&i); + } + return half(detail::binary, detail::rounded(sign+((exp-1)<<10)+(mx>>3), (mx>>2)&1, (mx&0x3)!=0)); + #endif + } + + /// Subtraction. + /// This operation is exact to rounding for all rounding modes. + /// \param x left operand + /// \param y right operand + /// \return difference of half expressions + /// \exception FE_INVALID if \a x and \a y are infinities with equal signs or signaling NaNs + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half operator-(half x, half y) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(detail::half2float(x.data_)-detail::half2float(y.data_))); + #else + return x + -y; + #endif + } + + /// Multiplication. + /// This operation is exact to rounding for all rounding modes. + /// \param x left operand + /// \param y right operand + /// \return product of half expressions + /// \exception FE_INVALID if multiplying 0 with infinity or if \a x or \a y is signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half operator*(half x, half y) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(detail::half2float(x.data_)*detail::half2float(y.data_))); + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = -16; + unsigned int sign = (x.data_^y.data_) & 0x8000; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : + ((absx==0x7C00 && !absy)||(absy==0x7C00 && !absx)) ? detail::invalid() : (sign|0x7C00)); + if(!absx || !absy) + return half(detail::binary, sign); + for(; absx<0x400; absx<<=1,--exp) ; + for(; absy<0x400; absy<<=1,--exp) ; + detail::uint32 m = static_cast((absx&0x3FF)|0x400) * static_cast((absy&0x3FF)|0x400); + int i = static_cast(m >> 21), s = static_cast(m & i); + exp += (absx>>10) + (absy>>10) + i; + if(exp > 29) + return half(detail::binary, detail::overflow(sign)); + else if(exp < -11) + return half(detail::binary, detail::underflow(sign)); + return half(detail::binary, detail::fixed2half(m>>i, exp, sign, s)); + #endif + } + + /// Division. + /// This operation is exact to rounding for all rounding modes. + /// \param x left operand + /// \param y right operand + /// \return quotient of half expressions + /// \exception FE_INVALID if dividing 0s or infinities with each other or if \a x or \a y is signaling NaN + /// \exception FE_DIVBYZERO if dividing finite value by 0 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half operator/(half x, half y) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(detail::half2float(x.data_)/detail::half2float(y.data_))); + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = 14; + unsigned int sign = (x.data_^y.data_) & 0x8000; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : + (absx==absy) ? detail::invalid() : (sign|((absx==0x7C00) ? 0x7C00 : 0))); + if(!absx) + return half(detail::binary, absy ? sign : detail::invalid()); + if(!absy) + return half(detail::binary, detail::pole(sign)); + for(; absx<0x400; absx<<=1,--exp) ; + for(; absy<0x400; absy<<=1,++exp) ; + detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400; + int i = mx < my; + exp += (absx>>10) - (absy>>10) - i; + if(exp > 29) + return half(detail::binary, detail::overflow(sign)); + else if(exp < -11) + return half(detail::binary, detail::underflow(sign)); + mx <<= 12 + i; + my <<= 1; + return half(detail::binary, detail::fixed2half(mx/my, exp, sign, mx%my!=0)); + #endif + } + + /// \} + /// \anchor streaming + /// \name Input and output + /// \{ + + /// Output operator. + /// This uses the built-in functionality for streaming out floating-point numbers. + /// \param out output stream to write into + /// \param arg half expression to write + /// \return reference to output stream + template std::basic_ostream& operator<<(std::basic_ostream &out, half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return out << detail::half2float(arg.data_); + #else + return out << detail::half2float(arg.data_); + #endif + } + + /// Input operator. + /// This uses the built-in functionality for streaming in floating-point numbers, specifically double precision floating + /// point numbers (unless overridden with [HALF_ARITHMETIC_TYPE](\ref HALF_ARITHMETIC_TYPE)). So the input string is first + /// rounded to double precision using the underlying platform's current floating-point rounding mode before being rounded + /// to half-precision using the library's half-precision rounding mode. + /// \param in input stream to read from + /// \param arg half to read into + /// \return reference to input stream + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + template std::basic_istream& operator>>(std::basic_istream &in, half &arg) + { + #ifdef HALF_ARITHMETIC_TYPE + detail::internal_t f; + #else + double f; + #endif + if(in >> f) + arg.data_ = detail::float2half(f); + return in; + } + + /// \} + /// \anchor basic + /// \name Basic mathematical operations + /// \{ + + /// Absolute value. + /// **See also:** Documentation for [std::fabs](https://en.cppreference.com/w/cpp/numeric/math/fabs). + /// \param arg operand + /// \return absolute value of \a arg + inline HALF_CONSTEXPR half fabs(half arg) { return half(detail::binary, arg.data_&0x7FFF); } + + /// Absolute value. + /// **See also:** Documentation for [std::abs](https://en.cppreference.com/w/cpp/numeric/math/fabs). + /// \param arg operand + /// \return absolute value of \a arg + inline HALF_CONSTEXPR half abs(half arg) { return fabs(arg); } + + /// Remainder of division. + /// **See also:** Documentation for [std::fmod](https://en.cppreference.com/w/cpp/numeric/math/fmod). + /// \param x first operand + /// \param y second operand + /// \return remainder of floating-point division. + /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN + inline half fmod(half x, half y) + { + unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, sign = x.data_ & 0x8000; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : + (absx==0x7C00) ? detail::invalid() : x.data_); + if(!absy) + return half(detail::binary, detail::invalid()); + if(!absx) + return x; + if(absx == absy) + return half(detail::binary, sign); + return half(detail::binary, sign|detail::mod(absx, absy)); + } + + /// Remainder of division. + /// **See also:** Documentation for [std::remainder](https://en.cppreference.com/w/cpp/numeric/math/remainder). + /// \param x first operand + /// \param y second operand + /// \return remainder of floating-point division. + /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN + inline half remainder(half x, half y) + { + unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, sign = x.data_ & 0x8000; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : + (absx==0x7C00) ? detail::invalid() : x.data_); + if(!absy) + return half(detail::binary, detail::invalid()); + if(absx == absy) + return half(detail::binary, sign); + return half(detail::binary, sign^detail::mod(absx, absy)); + } + + /// Remainder of division. + /// **See also:** Documentation for [std::remquo](https://en.cppreference.com/w/cpp/numeric/math/remquo). + /// \param x first operand + /// \param y second operand + /// \param quo address to store some bits of quotient at + /// \return remainder of floating-point division. + /// \exception FE_INVALID if \a x is infinite or \a y is 0 or if \a x or \a y is signaling NaN + inline half remquo(half x, half y, int *quo) + { + unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, value = x.data_ & 0x8000; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : + (absx==0x7C00) ? detail::invalid() : (*quo = 0, x.data_)); + if(!absy) + return half(detail::binary, detail::invalid()); + bool qsign = ((value^y.data_)&0x8000) != 0; + int q = 1; + if(absx != absy) + value ^= detail::mod(absx, absy, &q); + return *quo = qsign ? -q : q, half(detail::binary, value); + } + + /// Fused multiply add. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::fma](https://en.cppreference.com/w/cpp/numeric/math/fma). + /// \param x first operand + /// \param y second operand + /// \param z third operand + /// \return ( \a x * \a y ) + \a z rounded as one operation. + /// \exception FE_INVALID according to operator*() and operator+() unless any argument is a quiet NaN and no argument is a signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding the final addition + inline half fma(half x, half y, half z) + { + #ifdef HALF_ARITHMETIC_TYPE + detail::internal_t fx = detail::half2float(x.data_), fy = detail::half2float(y.data_), fz = detail::half2float(z.data_); + #if HALF_ENABLE_CPP11_CMATH && FP_FAST_FMA + return half(detail::binary, detail::float2half(std::fma(fx, fy, fz))); + #else + return half(detail::binary, detail::float2half(fx*fy+fz)); + #endif + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, absz = z.data_ & 0x7FFF, exp = -15; + unsigned int sign = (x.data_^y.data_) & 0x8000; + bool sub = ((sign^z.data_)&0x8000) != 0; + if(absx >= 0x7C00 || absy >= 0x7C00 || absz >= 0x7C00) + return (absx>0x7C00 || absy>0x7C00 || absz>0x7C00) ? half(detail::binary, detail::signal(x.data_, y.data_, z.data_)) : + (absx==0x7C00) ? half(detail::binary, (!absy || (sub && absz==0x7C00)) ? detail::invalid() : (sign|0x7C00)) : + (absy==0x7C00) ? half(detail::binary, (!absx || (sub && absz==0x7C00)) ? detail::invalid() : (sign|0x7C00)) : z; + if(!absx || !absy) + return absz ? z : half(detail::binary, (half::round_style==std::round_toward_neg_infinity) ? (z.data_|sign) : (z.data_&sign)); + for(; absx<0x400; absx<<=1,--exp) ; + for(; absy<0x400; absy<<=1,--exp) ; + detail::uint32 m = static_cast((absx&0x3FF)|0x400) * static_cast((absy&0x3FF)|0x400); + int i = static_cast(m >> 21); + exp += (absx>>10) + (absy>>10) + i; + m <<= 3 - i; + if(absz) + { + int expz = 0; + for(; absz<0x400; absz<<=1,--expz) ; + expz += absz >> 10; + detail::uint32 mz = static_cast((absz&0x3FF)|0x400) << 13; + if(expz > exp || (expz == exp && mz > m)) + { + std::swap(m, mz); + std::swap(exp, expz); + if(sub) + sign = z.data_ & 0x8000; + } + int d = exp - expz; + mz = (d<23) ? ((mz>>d)|((mz&((static_cast(1)<(half::round_style==std::round_toward_neg_infinity)<<15); + for(; m<0x800000; m<<=1,--exp) ; + } + else + { + m += mz; + i = static_cast(m >> 24); + m = (m>>i) | (m&i); + exp += i; + } + } + if(exp > 30) + return half(detail::binary, detail::overflow(sign)); + else if(exp < -10) + return half(detail::binary, detail::underflow(sign)); + return half(detail::binary, detail::fixed2half(m, exp-1, sign)); + #endif + } + + /// Maximum of half expressions. + /// **See also:** Documentation for [std::fmax](https://en.cppreference.com/w/cpp/numeric/math/fmax). + /// \param x first operand + /// \param y second operand + /// \return maximum of operands, ignoring quiet NaNs + /// \exception FE_INVALID if \a x or \a y is signaling NaN + inline HALF_CONSTEXPR_NOERR half fmax(half x, half y) + { + return half(detail::binary, (!isnan(y) && (isnan(x) || (x.data_^(0x8000|(0x8000-(x.data_>>15)))) < + (y.data_^(0x8000|(0x8000-(y.data_>>15)))))) ? detail::select(y.data_, x.data_) : detail::select(x.data_, y.data_)); + } + + /// Minimum of half expressions. + /// **See also:** Documentation for [std::fmin](https://en.cppreference.com/w/cpp/numeric/math/fmin). + /// \param x first operand + /// \param y second operand + /// \return minimum of operands, ignoring quiet NaNs + /// \exception FE_INVALID if \a x or \a y is signaling NaN + inline HALF_CONSTEXPR_NOERR half fmin(half x, half y) + { + return half(detail::binary, (!isnan(y) && (isnan(x) || (x.data_^(0x8000|(0x8000-(x.data_>>15)))) > + (y.data_^(0x8000|(0x8000-(y.data_>>15)))))) ? detail::select(y.data_, x.data_) : detail::select(x.data_, y.data_)); + } + + /// Positive difference. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::fdim](https://en.cppreference.com/w/cpp/numeric/math/fdim). + /// \param x first operand + /// \param y second operand + /// \return \a x - \a y or 0 if difference negative + /// \exception FE_... according to operator-(half,half) + inline half fdim(half x, half y) + { + if(isnan(x) || isnan(y)) + return half(detail::binary, detail::signal(x.data_, y.data_)); + return (x.data_^(0x8000|(0x8000-(x.data_>>15)))) <= (y.data_^(0x8000|(0x8000-(y.data_>>15)))) ? half(detail::binary, 0) : (x-y); + } + + /// Get NaN value. + /// **See also:** Documentation for [std::nan](https://en.cppreference.com/w/cpp/numeric/math/nan). + /// \param arg string code + /// \return quiet NaN + inline half nanh(const char *arg) + { + unsigned int value = 0x7FFF; + while(*arg) + value ^= static_cast(*arg++) & 0xFF; + return half(detail::binary, value); + } + + /// \} + /// \anchor exponential + /// \name Exponential functions + /// \{ + + /// Exponential function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::exp](https://en.cppreference.com/w/cpp/numeric/math/exp). + /// \param arg function argument + /// \return e raised to \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half exp(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::exp(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, e = (abs>>10) + (abs<=0x3FF), exp; + if(!abs) + return half(detail::binary, 0x3C00); + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? (0x7C00&((arg.data_>>15)-1U)) : detail::signal(arg.data_)); + if(abs >= 0x4C80) + return half(detail::binary, (arg.data_&0x8000) ? detail::underflow() : detail::overflow()); + detail::uint32 m = detail::multiply64(static_cast((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29); + if(e < 14) + { + exp = 0; + m >>= 14 - e; + } + else + { + exp = static_cast(m >> (45-e)); + m = (m<<(e-14)) & 0x7FFFFFFF; + } + return half(detail::binary, detail::exp2_post(m, exp, (arg.data_&0x8000)!=0, 0, 26)); + #endif + } + + /// Binary exponential. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::exp2](https://en.cppreference.com/w/cpp/numeric/math/exp2). + /// \param arg function argument + /// \return 2 raised to \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half exp2(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::exp2(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, e = (abs>>10) + (abs<=0x3FF), exp = (abs&0x3FF) + ((abs>0x3FF)<<10); + if(!abs) + return half(detail::binary, 0x3C00); + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? (0x7C00&((arg.data_>>15)-1U)) : detail::signal(arg.data_)); + if(abs >= 0x4E40) + return half(detail::binary, (arg.data_&0x8000) ? detail::underflow() : detail::overflow()); + return half(detail::binary, detail::exp2_post( + (static_cast(exp)<<(6+e))&0x7FFFFFFF, exp>>(25-e), (arg.data_&0x8000)!=0, 0, 28)); + #endif + } + + /// Exponential minus one. + /// This function may be 1 ULP off the correctly rounded exact result in <0.05% of inputs for `std::round_to_nearest` + /// and in <1% of inputs for any other rounding mode. + /// + /// **See also:** Documentation for [std::expm1](https://en.cppreference.com/w/cpp/numeric/math/expm1). + /// \param arg function argument + /// \return e raised to \a arg and subtracted by 1 + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half expm1(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::expm1(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000, e = (abs>>10) + (abs<=0x3FF), exp; + if(!abs) + return arg; + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? (0x7C00+(sign>>1)) : detail::signal(arg.data_)); + if(abs >= 0x4A00) + return half(detail::binary, (arg.data_&0x8000) ? detail::rounded(0xBBFF, 1, 1) : detail::overflow()); + detail::uint32 m = detail::multiply64(static_cast((abs&0x3FF)+((abs>0x3FF)<<10))<<21, 0xB8AA3B29); + if(e < 14) + { + exp = 0; + m >>= 14 - e; + } + else + { + exp = static_cast(m >> (45-e)); + m = (m<<(e-14)) & 0x7FFFFFFF; + } + m = detail::exp2(m); + if(sign) + { + int s = 0; + if(m > 0x80000000) + { + ++exp; + m = detail::divide64(0x80000000, m, s); + } + m = 0x80000000 - ((m>>exp)|((m&((static_cast(1)<>exp) : 1; + for(exp+=14; m<0x80000000 && exp; m<<=1,--exp) ; + if(exp > 29) + return half(detail::binary, detail::overflow()); + return half(detail::binary, detail::rounded(sign+(exp<<10)+static_cast(m>>21), static_cast((m>>20)&1), (m&0xFFFFF)!=0)); + #endif + } + + /// Natural logarithm. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::log](https://en.cppreference.com/w/cpp/numeric/math/log). + /// \param arg function argument + /// \return logarithm of \a arg to base e + /// \exception FE_INVALID for signaling NaN or negative argument + /// \exception FE_DIVBYZERO for 0 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half log(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::log(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = -15; + if(!abs) + return half(detail::binary, detail::pole(0x8000)); + if(arg.data_ & 0x8000) + return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs >= 0x7C00) + return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_)); + for(; abs<0x400; abs<<=1,--exp) ; + exp += abs >> 10; + return half(detail::binary, detail::log2_post( + detail::log2(static_cast((abs&0x3FF)|0x400)<<20, 27)+8, exp, 17)); + #endif + } + + /// Common logarithm. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::log10](https://en.cppreference.com/w/cpp/numeric/math/log10). + /// \param arg function argument + /// \return logarithm of \a arg to base 10 + /// \exception FE_INVALID for signaling NaN or negative argument + /// \exception FE_DIVBYZERO for 0 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half log10(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::log10(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = -15; + if(!abs) + return half(detail::binary, detail::pole(0x8000)); + if(arg.data_ & 0x8000) + return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs >= 0x7C00) + return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_)); + switch(abs) + { + case 0x4900: return half(detail::binary, 0x3C00); + case 0x5640: return half(detail::binary, 0x4000); + case 0x63D0: return half(detail::binary, 0x4200); + case 0x70E2: return half(detail::binary, 0x4400); + } + for(; abs<0x400; abs<<=1,--exp) ; + exp += abs >> 10; + return half(detail::binary, detail::log2_post( + detail::log2(static_cast((abs&0x3FF)|0x400)<<20, 27)+8, exp, 16)); + #endif + } + + /// Binary logarithm. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::log2](https://en.cppreference.com/w/cpp/numeric/math/log2). + /// \param arg function argument + /// \return logarithm of \a arg to base 2 + /// \exception FE_INVALID for signaling NaN or negative argument + /// \exception FE_DIVBYZERO for 0 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half log2(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::log2(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = -15, s = 0; + if(!abs) + return half(detail::binary, detail::pole(0x8000)); + if(arg.data_ & 0x8000) + return half(detail::binary, (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs >= 0x7C00) + return (abs==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_)); + if(abs == 0x3C00) + return half(detail::binary, 0); + for(; abs<0x400; abs<<=1,--exp) ; + exp += (abs>>10); + if(!(abs&0x3FF)) + { + unsigned int value = static_cast(exp<0) << 15, m = std::abs(exp) << 6; + for(exp=18; m<0x400; m<<=1,--exp) ; + return half(detail::binary, value+(exp<<10)+m); + } + detail::uint32 ilog = exp, sign = detail::sign_mask(ilog), m = + (((ilog<<27)+(detail::log2(static_cast((abs&0x3FF)|0x400)<<20, 28)>>4))^sign) - sign; + if(!m) + return half(detail::binary, 0); + for(exp=14; m<0x8000000 && exp; m<<=1,--exp) ; + for(; m>0xFFFFFFF; m>>=1,++exp) + s |= m & 1; + return half(detail::binary, detail::fixed2half(m, exp, sign&0x8000, s)); + #endif + } + + /// Natural logarithm plus one. + /// This function may be 1 ULP off the correctly rounded exact result in <0.05% of inputs for `std::round_to_nearest` + /// and in ~1% of inputs for any other rounding mode. + /// + /// **See also:** Documentation for [std::log1p](https://en.cppreference.com/w/cpp/numeric/math/log1p). + /// \param arg function argument + /// \return logarithm of \a arg plus 1 to base e + /// \exception FE_INVALID for signaling NaN or argument <-1 + /// \exception FE_DIVBYZERO for -1 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half log1p(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::log1p(detail::half2float(arg.data_)))); + #else + if(arg.data_ >= 0xBC00) + return half(detail::binary, (arg.data_==0xBC00) ? detail::pole(0x8000) : (arg.data_<=0xFC00) ? detail::invalid() : detail::signal(arg.data_)); + int abs = arg.data_ & 0x7FFF, exp = -15; + if(!abs || abs >= 0x7C00) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + for(; abs<0x400; abs<<=1,--exp) ; + exp += abs >> 10; + detail::uint32 m = static_cast((abs&0x3FF)|0x400) << 20; + if(arg.data_ & 0x8000) + { + m = 0x40000000 - (m>>-exp); + for(exp=0; m<0x40000000; m<<=1,--exp) ; + } + else + { + if(exp < 0) + { + m = 0x40000000 + (m>>-exp); + exp = 0; + } + else + { + m += 0x40000000 >> exp; + int i = static_cast(m >> 31); + m >>= i; + exp += i; + } + } + return half(detail::binary, detail::log2_post(detail::log2(m), exp, 17)); + #endif + } + + /// \} + /// \anchor power + /// \name Power functions + /// \{ + + /// Square root. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::sqrt](https://en.cppreference.com/w/cpp/numeric/math/sqrt). + /// \param arg function argument + /// \return square root of \a arg + /// \exception FE_INVALID for signaling NaN and negative arguments + /// \exception FE_INEXACT according to rounding + inline half sqrt(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::sqrt(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = 15; + if(!abs || arg.data_ >= 0x7C00) + return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_>0x8000) ? detail::invalid() : arg.data_); + for(; abs<0x400; abs<<=1,--exp) ; + detail::uint32 r = static_cast((abs&0x3FF)|0x400) << 10, m = detail::sqrt<20>(r, exp+=abs>>10); + return half(detail::binary, detail::rounded((exp<<10)+(m&0x3FF), r>m, r!=0)); + #endif + } + + /// Inverse square root. + /// This function is exact to rounding for all rounding modes and thus generally more accurate than directly computing + /// 1 / sqrt(\a arg) in half-precision, in addition to also being faster. + /// \param arg function argument + /// \return reciprocal of square root of \a arg + /// \exception FE_INVALID for signaling NaN and negative arguments + /// \exception FE_INEXACT according to rounding + inline half rsqrt(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(detail::internal_t(1)/std::sqrt(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF, bias = 0x4000; + if(!abs || arg.data_ >= 0x7C00) + return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_>0x8000) ? + detail::invalid() : !abs ? detail::pole(arg.data_&0x8000) : 0); + for(; abs<0x400; abs<<=1,bias-=0x400) ; + unsigned int frac = (abs+=bias) & 0x7FF; + if(frac == 0x400) + return half(detail::binary, 0x7A00-(abs>>1)); + if((half::round_style == std::round_to_nearest && (frac == 0x3FE || frac == 0x76C)) || + (half::round_style != std::round_to_nearest && (frac == 0x15A || frac == 0x3FC || frac == 0x401 || frac == 0x402 || frac == 0x67B))) + return pow(arg, half(detail::binary, 0xB800)); + detail::uint32 f = 0x17376 - abs, mx = (abs&0x3FF) | 0x400, my = ((f>>1)&0x3FF) | 0x400, mz = my * my; + int expy = static_cast(f>>11) - 31, expx = 32 - (abs>>10), i = static_cast(mz >> 21); + for(mz=0x60000000-(((mz>>i)*mx)>>(expx-2*expy-i)); mz<0x40000000; mz<<=1,--expy) ; + i = static_cast((my*=mz>>10) >> 31); + expy += i; + my = (my>>(20+i)) + 1; + i = static_cast((mz=my*my) >> 21); + for(mz=0x60000000-(((mz>>i)*mx)>>(expx-2*expy-i)); mz<0x40000000; mz<<=1,--expy) ; + i = static_cast((my*=(mz>>10)+1) >> 31); + return half(detail::binary, detail::fixed2half(my>>i, expy+i+14)); + #endif + } + + /// Cubic root. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::cbrt](https://en.cppreference.com/w/cpp/numeric/math/cbrt). + /// \param arg function argument + /// \return cubic root of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT according to rounding + inline half cbrt(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::cbrt(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = -15; + if(!abs || abs == 0x3C00 || abs >= 0x7C00) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + for(; abs<0x400; abs<<=1, --exp); + detail::uint32 ilog = exp + (abs>>10), sign = detail::sign_mask(ilog), f, m = + (((ilog<<27)+(detail::log2(static_cast((abs&0x3FF)|0x400)<<20, 24)>>4))^sign) - sign; + for(exp=2; m<0x80000000; m<<=1,--exp) ; + m = detail::multiply64(m, 0xAAAAAAAB); + int i = static_cast(m >> 31), s; + exp += i; + m <<= 1 - i; + if(exp < 0) + { + f = m >> -exp; + exp = 0; + } + else + { + f = (m<(m >> (31-exp)); + } + m = detail::exp2(f, (half::round_style==std::round_to_nearest) ? 29 : 26); + if(sign) + { + if(m > 0x80000000) + { + m = detail::divide64(0x80000000, m, s); + ++exp; + } + exp = -exp; + } + return half(detail::binary, (half::round_style==std::round_to_nearest) ? + detail::fixed2half(m, exp+14, arg.data_&0x8000) : + detail::fixed2half((m+0x80)>>8, exp+14, arg.data_&0x8000)); + #endif + } + + /// Hypotenuse function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::hypot](https://en.cppreference.com/w/cpp/numeric/math/hypot). + /// \param x first argument + /// \param y second argument + /// \return square root of sum of squares without internal over- or underflows + /// \exception FE_INVALID if \a x or \a y is signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding of the final square root + inline half hypot(half x, half y) + { + #ifdef HALF_ARITHMETIC_TYPE + detail::internal_t fx = detail::half2float(x.data_), fy = detail::half2float(y.data_); + #if HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::hypot(fx, fy))); + #else + return half(detail::binary, detail::float2half(std::sqrt(fx*fx+fy*fy))); + #endif + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, expx = 0, expy = 0; + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx==0x7C00) ? detail::select(0x7C00, y.data_) : + (absy==0x7C00) ? detail::select(0x7C00, x.data_) : detail::signal(x.data_, y.data_)); + if(!absx) + return half(detail::binary, absy ? detail::check_underflow(absy) : 0); + if(!absy) + return half(detail::binary, detail::check_underflow(absx)); + if(absy > absx) + std::swap(absx, absy); + for(; absx<0x400; absx<<=1,--expx) ; + for(; absy<0x400; absy<<=1,--expy) ; + detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400; + mx *= mx; + my *= my; + int ix = static_cast(mx >> 21), iy = static_cast(my >> 21); + expx = 2*(expx+(absx>>10)) - 15 + ix; + expy = 2*(expy+(absy>>10)) - 15 + iy; + mx <<= 10 - ix; + my <<= 10 - iy; + int d = expx - expy; + my = (d<30) ? ((my>>d)|((my&((static_cast(1)<(mx+my, expx)); + #endif + } + + /// Hypotenuse function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::hypot](https://en.cppreference.com/w/cpp/numeric/math/hypot). + /// \param x first argument + /// \param y second argument + /// \param z third argument + /// \return square root of sum of squares without internal over- or underflows + /// \exception FE_INVALID if \a x, \a y or \a z is signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding of the final square root + inline half hypot(half x, half y, half z) + { + #ifdef HALF_ARITHMETIC_TYPE + detail::internal_t fx = detail::half2float(x.data_), fy = detail::half2float(y.data_), fz = detail::half2float(z.data_); + return half(detail::binary, detail::float2half(std::sqrt(fx*fx+fy*fy+fz*fz))); + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, absz = z.data_ & 0x7FFF, expx = 0, expy = 0, expz = 0; + if(!absx) + return hypot(y, z); + if(!absy) + return hypot(x, z); + if(!absz) + return hypot(x, y); + if(absx >= 0x7C00 || absy >= 0x7C00 || absz >= 0x7C00) + return half(detail::binary, (absx==0x7C00) ? detail::select(0x7C00, detail::select(y.data_, z.data_)) : + (absy==0x7C00) ? detail::select(0x7C00, detail::select(x.data_, z.data_)) : + (absz==0x7C00) ? detail::select(0x7C00, detail::select(x.data_, y.data_)) : + detail::signal(x.data_, y.data_, z.data_)); + if(absz > absy) + std::swap(absy, absz); + if(absy > absx) + std::swap(absx, absy); + if(absz > absy) + std::swap(absy, absz); + for(; absx<0x400; absx<<=1,--expx) ; + for(; absy<0x400; absy<<=1,--expy) ; + for(; absz<0x400; absz<<=1,--expz) ; + detail::uint32 mx = (absx&0x3FF) | 0x400, my = (absy&0x3FF) | 0x400, mz = (absz&0x3FF) | 0x400; + mx *= mx; + my *= my; + mz *= mz; + int ix = static_cast(mx >> 21), iy = static_cast(my >> 21), iz = static_cast(mz >> 21); + expx = 2*(expx+(absx>>10)) - 15 + ix; + expy = 2*(expy+(absy>>10)) - 15 + iy; + expz = 2*(expz+(absz>>10)) - 15 + iz; + mx <<= 10 - ix; + my <<= 10 - iy; + mz <<= 10 - iz; + int d = expy - expz; + mz = (d<30) ? ((mz>>d)|((mz&((static_cast(1)<>1) | (my&1); + if(++expy > expx) + { + std::swap(mx, my); + std::swap(expx, expy); + } + } + d = expx - expy; + my = (d<30) ? ((my>>d)|((my&((static_cast(1)<(mx+my, expx)); + #endif + } + + /// Power function. + /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.00025% of inputs. + /// + /// **See also:** Documentation for [std::pow](https://en.cppreference.com/w/cpp/numeric/math/pow). + /// \param x base + /// \param y exponent + /// \return \a x raised to \a y + /// \exception FE_INVALID if \a x or \a y is signaling NaN or if \a x is finite an negative and \a y is finite and not integral + /// \exception FE_DIVBYZERO if \a x is 0 and \a y is negative + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half pow(half x, half y) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::pow(detail::half2float(x.data_), detail::half2float(y.data_)))); + #else + int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, exp = -15; + if(!absy || x.data_ == 0x3C00) + return half(detail::binary, detail::select(0x3C00, (x.data_==0x3C00) ? y.data_ : x.data_)); + bool is_int = absy >= 0x6400 || (absy>=0x3C00 && !(absy&((1<<(25-(absy>>10)))-1))); + unsigned int sign = x.data_ & (static_cast((absy<0x6800)&&is_int&&((absy>>(25-(absy>>10)))&1))<<15); + if(absx >= 0x7C00 || absy >= 0x7C00) + return half(detail::binary, (absx>0x7C00 || absy>0x7C00) ? detail::signal(x.data_, y.data_) : + (absy==0x7C00) ? ((absx==0x3C00) ? 0x3C00 : (!absx && y.data_==0xFC00) ? detail::pole() : + (0x7C00&-((y.data_>>15)^(absx>0x3C00)))) : (sign|(0x7C00&((y.data_>>15)-1U)))); + if(!absx) + return half(detail::binary, (y.data_&0x8000) ? detail::pole(sign) : sign); + if((x.data_&0x8000) && !is_int) + return half(detail::binary, detail::invalid()); + if(x.data_ == 0xBC00) + return half(detail::binary, sign|0x3C00); + switch(y.data_) + { + case 0x3800: return sqrt(x); + case 0x3C00: return half(detail::binary, detail::check_underflow(x.data_)); + case 0x4000: return x * x; + case 0xBC00: return half(detail::binary, 0x3C00) / x; + } + for(; absx<0x400; absx<<=1,--exp) ; + detail::uint32 ilog = exp + (absx>>10), msign = detail::sign_mask(ilog), f, m = + (((ilog<<27)+((detail::log2(static_cast((absx&0x3FF)|0x400)<<20)+8)>>4))^msign) - msign; + for(exp=-11; m<0x80000000; m<<=1,--exp) ; + for(; absy<0x400; absy<<=1,--exp) ; + m = detail::multiply64(m, static_cast((absy&0x3FF)|0x400)<<21); + int i = static_cast(m >> 31); + exp += (absy>>10) + i; + m <<= 1 - i; + if(exp < 0) + { + f = m >> -exp; + exp = 0; + } + else + { + f = (m<(m >> (31-exp)); + } + return half(detail::binary, detail::exp2_post(f, exp, ((msign&1)^(y.data_>>15))!=0, sign)); + #endif + } + + /// \} + /// \anchor trigonometric + /// \name Trigonometric functions + /// \{ + + /// Compute sine and cosine simultaneously. + /// This returns the same results as sin() and cos() but is faster than calling each function individually. + /// + /// This function is exact to rounding for all rounding modes. + /// \param arg function argument + /// \param sin variable to take sine of \a arg + /// \param cos variable to take cosine of \a arg + /// \exception FE_INVALID for signaling NaN or infinity + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline void sincos(half arg, half *sin, half *cos) + { + #ifdef HALF_ARITHMETIC_TYPE + detail::internal_t f = detail::half2float(arg.data_); + *sin = half(detail::binary, detail::float2half(std::sin(f))); + *cos = half(detail::binary, detail::float2half(std::cos(f))); + #else + int abs = arg.data_ & 0x7FFF, sign = arg.data_ >> 15, k; + if(abs >= 0x7C00) + *sin = *cos = half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_)); + else if(!abs) + { + *sin = arg; + *cos = half(detail::binary, 0x3C00); + } + else if(abs < 0x2500) + { + *sin = half(detail::binary, detail::rounded(arg.data_-1, 1, 1)); + *cos = half(detail::binary, detail::rounded(0x3BFF, 1, 1)); + } + else + { + if(half::round_style != std::round_to_nearest) + { + switch(abs) + { + case 0x48B7: + *sin = half(detail::binary, detail::rounded((~arg.data_&0x8000)|0x1D07, 1, 1)); + *cos = half(detail::binary, detail::rounded(0xBBFF, 1, 1)); + return; + case 0x598C: + *sin = half(detail::binary, detail::rounded((arg.data_&0x8000)|0x3BFF, 1, 1)); + *cos = half(detail::binary, detail::rounded(0x80FC, 1, 1)); + return; + case 0x6A64: + *sin = half(detail::binary, detail::rounded((~arg.data_&0x8000)|0x3BFE, 1, 1)); + *cos = half(detail::binary, detail::rounded(0x27FF, 1, 1)); + return; + case 0x6D8C: + *sin = half(detail::binary, detail::rounded((arg.data_&0x8000)|0x0FE6, 1, 1)); + *cos = half(detail::binary, detail::rounded(0x3BFF, 1, 1)); + return; + } + } + std::pair sc = detail::sincos(detail::angle_arg(abs, k), 28); + switch(k & 3) + { + case 1: sc = std::make_pair(sc.second, -sc.first); break; + case 2: sc = std::make_pair(-sc.first, -sc.second); break; + case 3: sc = std::make_pair(-sc.second, sc.first); break; + } + *sin = half(detail::binary, detail::fixed2half((sc.first^-static_cast(sign))+sign)); + *cos = half(detail::binary, detail::fixed2half(sc.second)); + } + #endif + } + + /// Sine function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::sin](https://en.cppreference.com/w/cpp/numeric/math/sin). + /// \param arg function argument + /// \return sine value of \a arg + /// \exception FE_INVALID for signaling NaN or infinity + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half sin(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::sin(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, k; + if(!abs) + return arg; + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs < 0x2900) + return half(detail::binary, detail::rounded(arg.data_-1, 1, 1)); + if(half::round_style != std::round_to_nearest) + switch(abs) + { + case 0x48B7: return half(detail::binary, detail::rounded((~arg.data_&0x8000)|0x1D07, 1, 1)); + case 0x6A64: return half(detail::binary, detail::rounded((~arg.data_&0x8000)|0x3BFE, 1, 1)); + case 0x6D8C: return half(detail::binary, detail::rounded((arg.data_&0x8000)|0x0FE6, 1, 1)); + } + std::pair sc = detail::sincos(detail::angle_arg(abs, k), 28); + detail::uint32 sign = -static_cast(((k>>1)&1)^(arg.data_>>15)); + return half(detail::binary, detail::fixed2half((((k&1) ? sc.second : sc.first)^sign) - sign)); + #endif + } + + /// Cosine function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::cos](https://en.cppreference.com/w/cpp/numeric/math/cos). + /// \param arg function argument + /// \return cosine value of \a arg + /// \exception FE_INVALID for signaling NaN or infinity + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half cos(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::cos(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, k; + if(!abs) + return half(detail::binary, 0x3C00); + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs < 0x2500) + return half(detail::binary, detail::rounded(0x3BFF, 1, 1)); + if(half::round_style != std::round_to_nearest && abs == 0x598C) + return half(detail::binary, detail::rounded(0x80FC, 1, 1)); + std::pair sc = detail::sincos(detail::angle_arg(abs, k), 28); + detail::uint32 sign = -static_cast(((k>>1)^k)&1); + return half(detail::binary, detail::fixed2half((((k&1) ? sc.first : sc.second)^sign) - sign)); + #endif + } + + /// Tangent function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::tan](https://en.cppreference.com/w/cpp/numeric/math/tan). + /// \param arg function argument + /// \return tangent value of \a arg + /// \exception FE_INVALID for signaling NaN or infinity + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half tan(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::tan(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = 13, k; + if(!abs) + return arg; + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs < 0x2700) + return half(detail::binary, detail::rounded(arg.data_, 0, 1)); + if(half::round_style != std::round_to_nearest) + switch(abs) + { + case 0x658C: return half(detail::binary, detail::rounded((arg.data_&0x8000)|0x07E6, 1, 1)); + case 0x7330: return half(detail::binary, detail::rounded((~arg.data_&0x8000)|0x4B62, 1, 1)); + } + std::pair sc = detail::sincos(detail::angle_arg(abs, k), 30); + if(k & 1) + sc = std::make_pair(-sc.second, sc.first); + detail::uint32 signy = detail::sign_mask(sc.first), signx = detail::sign_mask(sc.second); + detail::uint32 my = (sc.first^signy) - signy, mx = (sc.second^signx) - signx; + for(; my<0x80000000; my<<=1,--exp) ; + for(; mx<0x80000000; mx<<=1,++exp) ; + return half(detail::binary, detail::tangent_post(my, mx, exp, (signy^signx^arg.data_)&0x8000)); + #endif + } + + /// Arc sine. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::asin](https://en.cppreference.com/w/cpp/numeric/math/asin). + /// \param arg function argument + /// \return arc sine value of \a arg + /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half asin(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::asin(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000; + if(!abs) + return arg; + if(abs >= 0x3C00) + return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (abs>0x3C00) ? detail::invalid() : + detail::rounded(sign|0x3E48, 0, 1)); + if(abs < 0x2900) + return half(detail::binary, detail::rounded(arg.data_, 0, 1)); + if(half::round_style != std::round_to_nearest && (abs == 0x2B44 || abs == 0x2DC3)) + return half(detail::binary, detail::rounded(arg.data_+1, 1, 1)); + std::pair sc = detail::atan2_args(abs); + detail::uint32 m = detail::atan2(sc.first, sc.second, (half::round_style==std::round_to_nearest) ? 27 : 26); + return half(detail::binary, detail::fixed2half(m, 14, sign)); + #endif + } + + /// Arc cosine function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::acos](https://en.cppreference.com/w/cpp/numeric/math/acos). + /// \param arg function argument + /// \return arc cosine value of \a arg + /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half acos(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::acos(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ >> 15; + if(!abs) + return half(detail::binary, detail::rounded(0x3E48, 0, 1)); + if(abs >= 0x3C00) + return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (abs>0x3C00) ? detail::invalid() : + sign ? detail::rounded(0x4248, 0, 1) : 0); + std::pair cs = detail::atan2_args(abs); + detail::uint32 m = detail::atan2(cs.second, cs.first, 28); + return half(detail::binary, detail::fixed2half(sign ? (0xC90FDAA2-m) : m, 15, 0, sign)); + #endif + } + + /// Arc tangent function. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::atan](https://en.cppreference.com/w/cpp/numeric/math/atan). + /// \param arg function argument + /// \return arc tangent value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half atan(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::atan(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000; + if(!abs) + return arg; + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? detail::rounded(sign|0x3E48, 0, 1) : detail::signal(arg.data_)); + if(abs <= 0x2700) + return half(detail::binary, detail::rounded(arg.data_-1, 1, 1)); + int exp = (abs>>10) + (abs<=0x3FF); + detail::uint32 my = (abs&0x3FF) | ((abs>0x3FF)<<10); + detail::uint32 m = (exp>15) ? detail::atan2(my<<19, 0x20000000>>(exp-15), (half::round_style==std::round_to_nearest) ? 26 : 24) : + detail::atan2(my<<(exp+4), 0x20000000, (half::round_style==std::round_to_nearest) ? 30 : 28); + return half(detail::binary, detail::fixed2half(m, 14, sign)); + #endif + } + + /// Arc tangent function. + /// This function may be 1 ULP off the correctly rounded exact result in ~0.005% of inputs for `std::round_to_nearest`, + /// in ~0.1% of inputs for `std::round_toward_zero` and in ~0.02% of inputs for any other rounding mode. + /// + /// **See also:** Documentation for [std::atan2](https://en.cppreference.com/w/cpp/numeric/math/atan2). + /// \param y numerator + /// \param x denominator + /// \return arc tangent value + /// \exception FE_INVALID if \a x or \a y is signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half atan2(half y, half x) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::atan2(detail::half2float(y.data_), detail::half2float(x.data_)))); + #else + unsigned int absx = x.data_ & 0x7FFF, absy = y.data_ & 0x7FFF, signx = x.data_ >> 15, signy = y.data_ & 0x8000; + if(absx >= 0x7C00 || absy >= 0x7C00) + { + if(absx > 0x7C00 || absy > 0x7C00) + return half(detail::binary, detail::signal(x.data_, y.data_)); + if(absy == 0x7C00) + return half(detail::binary, (absx<0x7C00) ? detail::rounded(signy|0x3E48, 0, 1) : + signx ? detail::rounded(signy|0x40B6, 0, 1) : + detail::rounded(signy|0x3A48, 0, 1)); + return (x.data_==0x7C00) ? half(detail::binary, signy) : half(detail::binary, detail::rounded(signy|0x4248, 0, 1)); + } + if(!absy) + return signx ? half(detail::binary, detail::rounded(signy|0x4248, 0, 1)) : y; + if(!absx) + return half(detail::binary, detail::rounded(signy|0x3E48, 0, 1)); + int d = (absy>>10) + (absy<=0x3FF) - (absx>>10) - (absx<=0x3FF); + if(d > (signx ? 18 : 12)) + return half(detail::binary, detail::rounded(signy|0x3E48, 0, 1)); + if(signx && d < -11) + return half(detail::binary, detail::rounded(signy|0x4248, 0, 1)); + if(!signx && d < ((half::round_style==std::round_toward_zero) ? -15 : -9)) + { + for(; absy<0x400; absy<<=1,--d) ; + detail::uint32 mx = ((absx<<1)&0x7FF) | 0x800, my = ((absy<<1)&0x7FF) | 0x800; + int i = my < mx; + d -= i; + if(d < -25) + return half(detail::binary, detail::underflow(signy)); + my <<= 11 + i; + return half(detail::binary, detail::fixed2half(my/mx, d+14, signy, my%mx!=0)); + } + detail::uint32 m = detail::atan2( ((absy&0x3FF)|((absy>0x3FF)<<10))<<(19+((d<0) ? d : (d>0) ? 0 : -1)), + ((absx&0x3FF)|((absx>0x3FF)<<10))<<(19-((d>0) ? d : (d<0) ? 0 : 1))); + return half(detail::binary, detail::fixed2half(signx ? (0xC90FDAA2-m) : m, 15, signy, signx)); + #endif + } + + /// \} + /// \anchor hyperbolic + /// \name Hyperbolic functions + /// \{ + + /// Hyperbolic sine. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::sinh](https://en.cppreference.com/w/cpp/numeric/math/sinh). + /// \param arg function argument + /// \return hyperbolic sine value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half sinh(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::sinh(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp; + if(!abs || abs >= 0x7C00) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + if(abs <= 0x2900) + return half(detail::binary, detail::rounded(arg.data_, 0, 1)); + std::pair mm = detail::hyperbolic_args(abs, exp, (half::round_style==std::round_to_nearest) ? 29 : 27); + detail::uint32 m = mm.first - mm.second; + for(exp+=13; m<0x80000000 && exp; m<<=1,--exp) ; + unsigned int sign = arg.data_ & 0x8000; + if(exp > 29) + return half(detail::binary, detail::overflow(sign)); + return half(detail::binary, detail::fixed2half(m, exp, sign)); + #endif + } + + /// Hyperbolic cosine. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::cosh](https://en.cppreference.com/w/cpp/numeric/math/cosh). + /// \param arg function argument + /// \return hyperbolic cosine value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half cosh(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::cosh(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp; + if(!abs) + return half(detail::binary, 0x3C00); + if(abs >= 0x7C00) + return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : 0x7C00); + std::pair mm = detail::hyperbolic_args(abs, exp, (half::round_style==std::round_to_nearest) ? 23 : 26); + detail::uint32 m = mm.first + mm.second; + int i = static_cast((~m & 0xFFFFFFFF) >> 31); + m = (m>>i) | (m&i) | 0x80000000; + if((exp+=13+i) > 29) + return half(detail::binary, detail::overflow()); + return half(detail::binary, detail::fixed2half(m, exp)); + #endif + } + + /// Hyperbolic tangent. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::tanh](https://en.cppreference.com/w/cpp/numeric/math/tanh). + /// \param arg function argument + /// \return hyperbolic tangent value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half tanh(half arg) + { + #ifdef HALF_ARITHMETIC_TYPE + return half(detail::binary, detail::float2half(std::tanh(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp; + if(!abs) + return arg; + if(abs >= 0x7C00) + return half(detail::binary, (abs>0x7C00) ? detail::signal(arg.data_) : (arg.data_-0x4000)); + if(abs >= 0x4500) + return half(detail::binary, detail::rounded((arg.data_&0x8000)|0x3BFF, 1, 1)); + if(abs < 0x2700) + return half(detail::binary, detail::rounded(arg.data_-1, 1, 1)); + if(half::round_style != std::round_to_nearest && abs == 0x2D3F) + return half(detail::binary, detail::rounded(arg.data_-3, 0, 1)); + std::pair mm = detail::hyperbolic_args(abs, exp, 27); + detail::uint32 my = mm.first - mm.second - (half::round_style!=std::round_to_nearest), mx = mm.first + mm.second; + int i = static_cast((~mx&0xFFFFFFFF) >> 31); + for(exp=13; my<0x80000000; my<<=1,--exp) ; + mx = (mx>>i) | 0x80000000; + return half(detail::binary, detail::tangent_post(my, mx, exp-i, arg.data_&0x8000)); + #endif + } + + /// Hyperbolic area sine. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::asinh](https://en.cppreference.com/w/cpp/numeric/math/asinh). + /// \param arg function argument + /// \return area sine value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half asinh(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::asinh(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF; + if(!abs || abs >= 0x7C00) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + if(abs <= 0x2900) + return half(detail::binary, detail::rounded(arg.data_-1, 1, 1)); + if(half::round_style != std::round_to_nearest) + switch(abs) + { + case 0x32D4: return half(detail::binary, detail::rounded(arg.data_-13, 1, 1)); + case 0x3B5B: return half(detail::binary, detail::rounded(arg.data_-197, 1, 1)); + } + return half(detail::binary, detail::area(arg.data_)); + #endif + } + + /// Hyperbolic area cosine. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::acosh](https://en.cppreference.com/w/cpp/numeric/math/acosh). + /// \param arg function argument + /// \return area cosine value of \a arg + /// \exception FE_INVALID for signaling NaN or arguments <1 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half acosh(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::acosh(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF; + if((arg.data_&0x8000) || abs < 0x3C00) + return half(detail::binary, (abs<=0x7C00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs == 0x3C00) + return half(detail::binary, 0); + if(arg.data_ >= 0x7C00) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + return half(detail::binary, detail::area(arg.data_)); + #endif + } + + /// Hyperbolic area tangent. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::atanh](https://en.cppreference.com/w/cpp/numeric/math/atanh). + /// \param arg function argument + /// \return area tangent value of \a arg + /// \exception FE_INVALID for signaling NaN or if abs(\a arg) > 1 + /// \exception FE_DIVBYZERO for +/-1 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half atanh(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::atanh(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF, exp = 0; + if(!abs) + return arg; + if(abs >= 0x3C00) + return half(detail::binary, (abs==0x3C00) ? detail::pole(arg.data_&0x8000) : (abs<=0x7C00) ? detail::invalid() : detail::signal(arg.data_)); + if(abs < 0x2700) + return half(detail::binary, detail::rounded(arg.data_, 0, 1)); + detail::uint32 m = static_cast((abs&0x3FF)|((abs>0x3FF)<<10)) << ((abs>>10)+(abs<=0x3FF)+6), my = 0x80000000 + m, mx = 0x80000000 - m; + for(; mx<0x80000000; mx<<=1,++exp) ; + int i = my >= mx, s; + return half(detail::binary, detail::log2_post(detail::log2( + (detail::divide64(my>>i, mx, s)+1)>>1, 27)+0x10, exp+i-1, 16, arg.data_&0x8000)); + #endif + } + + /// \} + /// \anchor special + /// \name Error and gamma functions + /// \{ + + /// Error function. + /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.5% of inputs. + /// + /// **See also:** Documentation for [std::erf](https://en.cppreference.com/w/cpp/numeric/math/erf). + /// \param arg function argument + /// \return error function value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half erf(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::erf(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF; + if(!abs || abs >= 0x7C00) + return (abs>=0x7C00) ? half(detail::binary, (abs==0x7C00) ? (arg.data_-0x4000) : detail::signal(arg.data_)) : arg; + if(abs >= 0x4200) + return half(detail::binary, detail::rounded((arg.data_&0x8000)|0x3BFF, 1, 1)); + return half(detail::binary, detail::erf(arg.data_)); + #endif + } + + /// Complementary error function. + /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.5% of inputs. + /// + /// **See also:** Documentation for [std::erfc](https://en.cppreference.com/w/cpp/numeric/math/erfc). + /// \param arg function argument + /// \return 1 minus error function value of \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half erfc(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::erfc(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000; + if(abs >= 0x7C00) + return (abs>=0x7C00) ? half(detail::binary, (abs==0x7C00) ? (sign>>1) : detail::signal(arg.data_)) : arg; + if(!abs) + return half(detail::binary, 0x3C00); + if(abs >= 0x4400) + return half(detail::binary, detail::rounded((sign>>1)-(sign>>15), sign>>15, 1)); + return half(detail::binary, detail::erf(arg.data_)); + #endif + } + + /// Natural logarithm of gamma function. + /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in ~0.025% of inputs. + /// + /// **See also:** Documentation for [std::lgamma](https://en.cppreference.com/w/cpp/numeric/math/lgamma). + /// \param arg function argument + /// \return natural logarith of gamma function for \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_DIVBYZERO for 0 or negative integer arguments + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half lgamma(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::lgamma(detail::half2float(arg.data_)))); + #else + int abs = arg.data_ & 0x7FFF; + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? 0x7C00 : detail::signal(arg.data_)); + if(!abs || arg.data_ >= 0xE400 || (arg.data_ >= 0xBC00 && !(abs&((1<<(25-(abs>>10)))-1)))) + return half(detail::binary, detail::pole()); + if(arg.data_ == 0x3C00 || arg.data_ == 0x4000) + return half(detail::binary, 0); + return half(detail::binary, detail::gamma(arg.data_)); + #endif + } + + /// Gamma function. + /// This function may be 1 ULP off the correctly rounded exact result for any rounding mode in <0.25% of inputs. + /// + /// **See also:** Documentation for [std::tgamma](https://en.cppreference.com/w/cpp/numeric/math/tgamma). + /// \param arg function argument + /// \return gamma function value of \a arg + /// \exception FE_INVALID for signaling NaN, negative infinity or negative integer arguments + /// \exception FE_DIVBYZERO for 0 + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half tgamma(half arg) + { + #if defined(HALF_ARITHMETIC_TYPE) && HALF_ENABLE_CPP11_CMATH + return half(detail::binary, detail::float2half(std::tgamma(detail::half2float(arg.data_)))); + #else + unsigned int abs = arg.data_ & 0x7FFF; + if(!abs) + return half(detail::binary, detail::pole(arg.data_)); + if(abs >= 0x7C00) + return (arg.data_==0x7C00) ? arg : half(detail::binary, detail::signal(arg.data_)); + if(arg.data_ >= 0xE400 || (arg.data_ >= 0xBC00 && !(abs&((1<<(25-(abs>>10)))-1)))) + return half(detail::binary, detail::invalid()); + if(arg.data_ >= 0xCA80) + return half(detail::binary, detail::underflow((1-((abs>>(25-(abs>>10)))&1))<<15)); + if(arg.data_ <= 0x100 || (arg.data_ >= 0x4900 && arg.data_ < 0x8000)) + return half(detail::binary, detail::overflow()); + if(arg.data_ == 0x3C00) + return arg; + return half(detail::binary, detail::gamma(arg.data_)); + #endif + } + + /// \} + /// \anchor rounding + /// \name Rounding + /// \{ + + /// Nearest integer not less than half value. + /// **See also:** Documentation for [std::ceil](https://en.cppreference.com/w/cpp/numeric/math/ceil). + /// \param arg half to round + /// \return nearest integer not less than \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT if value had to be rounded + inline half ceil(half arg) { return half(detail::binary, detail::integral(arg.data_)); } + + /// Nearest integer not greater than half value. + /// **See also:** Documentation for [std::floor](https://en.cppreference.com/w/cpp/numeric/math/floor). + /// \param arg half to round + /// \return nearest integer not greater than \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT if value had to be rounded + inline half floor(half arg) { return half(detail::binary, detail::integral(arg.data_)); } + + /// Nearest integer not greater in magnitude than half value. + /// **See also:** Documentation for [std::trunc](https://en.cppreference.com/w/cpp/numeric/math/trunc). + /// \param arg half to round + /// \return nearest integer not greater in magnitude than \a arg + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT if value had to be rounded + inline half trunc(half arg) { return half(detail::binary, detail::integral(arg.data_)); } + + /// Nearest integer. + /// **See also:** Documentation for [std::round](https://en.cppreference.com/w/cpp/numeric/math/round). + /// \param arg half to round + /// \return nearest integer, rounded away from zero in half-way cases + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT if value had to be rounded + inline half round(half arg) { return half(detail::binary, detail::integral(arg.data_)); } + + /// Nearest integer. + /// **See also:** Documentation for [std::lround](https://en.cppreference.com/w/cpp/numeric/math/round). + /// \param arg half to round + /// \return nearest integer, rounded away from zero in half-way cases + /// \exception FE_INVALID if value is not representable as `long` + inline long lround(half arg) { return detail::half2int(arg.data_); } + + /// Nearest integer using half's internal rounding mode. + /// **See also:** Documentation for [std::rint](https://en.cppreference.com/w/cpp/numeric/math/rint). + /// \param arg half expression to round + /// \return nearest integer using default rounding mode + /// \exception FE_INVALID for signaling NaN + /// \exception FE_INEXACT if value had to be rounded + inline half rint(half arg) { return half(detail::binary, detail::integral(arg.data_)); } + + /// Nearest integer using half's internal rounding mode. + /// **See also:** Documentation for [std::lrint](https://en.cppreference.com/w/cpp/numeric/math/rint). + /// \param arg half expression to round + /// \return nearest integer using default rounding mode + /// \exception FE_INVALID if value is not representable as `long` + /// \exception FE_INEXACT if value had to be rounded + inline long lrint(half arg) { return detail::half2int(arg.data_); } + + /// Nearest integer using half's internal rounding mode. + /// **See also:** Documentation for [std::nearbyint](https://en.cppreference.com/w/cpp/numeric/math/nearbyint). + /// \param arg half expression to round + /// \return nearest integer using default rounding mode + /// \exception FE_INVALID for signaling NaN + inline half nearbyint(half arg) { return half(detail::binary, detail::integral(arg.data_)); } +#if HALF_ENABLE_CPP11_LONG_LONG + /// Nearest integer. + /// **See also:** Documentation for [std::llround](https://en.cppreference.com/w/cpp/numeric/math/round). + /// \param arg half to round + /// \return nearest integer, rounded away from zero in half-way cases + /// \exception FE_INVALID if value is not representable as `long long` + inline long long llround(half arg) { return detail::half2int(arg.data_); } + + /// Nearest integer using half's internal rounding mode. + /// **See also:** Documentation for [std::llrint](https://en.cppreference.com/w/cpp/numeric/math/rint). + /// \param arg half expression to round + /// \return nearest integer using default rounding mode + /// \exception FE_INVALID if value is not representable as `long long` + /// \exception FE_INEXACT if value had to be rounded + inline long long llrint(half arg) { return detail::half2int(arg.data_); } +#endif + + /// \} + /// \anchor float + /// \name Floating point manipulation + /// \{ + + /// Decompress floating-point number. + /// **See also:** Documentation for [std::frexp](https://en.cppreference.com/w/cpp/numeric/math/frexp). + /// \param arg number to decompress + /// \param exp address to store exponent at + /// \return significant in range [0.5, 1) + /// \exception FE_INVALID for signaling NaN + inline half frexp(half arg, int *exp) + { + *exp = 0; + unsigned int abs = arg.data_ & 0x7FFF; + if(abs >= 0x7C00 || !abs) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + for(; abs<0x400; abs<<=1,--*exp) ; + *exp += (abs>>10) - 14; + return half(detail::binary, (arg.data_&0x8000)|0x3800|(abs&0x3FF)); + } + + /// Multiply by power of two. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::scalbln](https://en.cppreference.com/w/cpp/numeric/math/scalbn). + /// \param arg number to modify + /// \param exp power of two to multiply with + /// \return \a arg multplied by 2 raised to \a exp + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half scalbln(half arg, long exp) + { + unsigned int abs = arg.data_ & 0x7FFF, sign = arg.data_ & 0x8000; + if(abs >= 0x7C00 || !abs) + return (abs>0x7C00) ? half(detail::binary, detail::signal(arg.data_)) : arg; + for(; abs<0x400; abs<<=1,--exp) ; + exp += abs >> 10; + if(exp > 30) + return half(detail::binary, detail::overflow(sign)); + else if(exp < -10) + return half(detail::binary, detail::underflow(sign)); + else if(exp > 0) + return half(detail::binary, sign|(exp<<10)|(abs&0x3FF)); + unsigned int m = (abs&0x3FF) | 0x400; + return half(detail::binary, detail::rounded(sign|(m>>(1-exp)), (m>>-exp)&1, (m&((1<<-exp)-1))!=0)); + } + + /// Multiply by power of two. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::scalbn](https://en.cppreference.com/w/cpp/numeric/math/scalbn). + /// \param arg number to modify + /// \param exp power of two to multiply with + /// \return \a arg multplied by 2 raised to \a exp + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half scalbn(half arg, int exp) { return scalbln(arg, exp); } + + /// Multiply by power of two. + /// This function is exact to rounding for all rounding modes. + /// + /// **See also:** Documentation for [std::ldexp](https://en.cppreference.com/w/cpp/numeric/math/ldexp). + /// \param arg number to modify + /// \param exp power of two to multiply with + /// \return \a arg multplied by 2 raised to \a exp + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + inline half ldexp(half arg, int exp) { return scalbln(arg, exp); } + + /// Extract integer and fractional parts. + /// **See also:** Documentation for [std::modf](https://en.cppreference.com/w/cpp/numeric/math/modf). + /// \param arg number to decompress + /// \param iptr address to store integer part at + /// \return fractional part + /// \exception FE_INVALID for signaling NaN + inline half modf(half arg, half *iptr) + { + unsigned int abs = arg.data_ & 0x7FFF; + if(abs > 0x7C00) + { + arg = half(detail::binary, detail::signal(arg.data_)); + return *iptr = arg, arg; + } + if(abs >= 0x6400) + return *iptr = arg, half(detail::binary, arg.data_&0x8000); + if(abs < 0x3C00) + return iptr->data_ = arg.data_ & 0x8000, arg; + unsigned int exp = abs >> 10, mask = (1<<(25-exp)) - 1, m = arg.data_ & mask; + iptr->data_ = arg.data_ & ~mask; + if(!m) + return half(detail::binary, arg.data_&0x8000); + for(; m<0x400; m<<=1,--exp) ; + return half(detail::binary, (arg.data_&0x8000)|(exp<<10)|(m&0x3FF)); + } + + /// Extract exponent. + /// **See also:** Documentation for [std::ilogb](https://en.cppreference.com/w/cpp/numeric/math/ilogb). + /// \param arg number to query + /// \return floating-point exponent + /// \retval FP_ILOGB0 for zero + /// \retval FP_ILOGBNAN for NaN + /// \retval INT_MAX for infinity + /// \exception FE_INVALID for 0 or infinite values + inline int ilogb(half arg) + { + int abs = arg.data_ & 0x7FFF, exp; + if(!abs || abs >= 0x7C00) + { + detail::raise(FE_INVALID); + return !abs ? FP_ILOGB0 : (abs==0x7C00) ? INT_MAX : FP_ILOGBNAN; + } + for(exp=(abs>>10)-15; abs<0x200; abs<<=1,--exp) ; + return exp; + } + + /// Extract exponent. + /// **See also:** Documentation for [std::logb](https://en.cppreference.com/w/cpp/numeric/math/logb). + /// \param arg number to query + /// \return floating-point exponent + /// \exception FE_INVALID for signaling NaN + /// \exception FE_DIVBYZERO for 0 + inline half logb(half arg) + { + int abs = arg.data_ & 0x7FFF, exp; + if(!abs) + return half(detail::binary, detail::pole(0x8000)); + if(abs >= 0x7C00) + return half(detail::binary, (abs==0x7C00) ? 0x7C00 : detail::signal(arg.data_)); + for(exp=(abs>>10)-15; abs<0x200; abs<<=1,--exp) ; + unsigned int value = static_cast(exp<0) << 15; + if(exp) + { + unsigned int m = std::abs(exp) << 6; + for(exp=18; m<0x400; m<<=1,--exp) ; + value |= (exp<<10) + m; + } + return half(detail::binary, value); + } + + /// Next representable value. + /// **See also:** Documentation for [std::nextafter](https://en.cppreference.com/w/cpp/numeric/math/nextafter). + /// \param from value to compute next representable value for + /// \param to direction towards which to compute next value + /// \return next representable value after \a from in direction towards \a to + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW for infinite result from finite argument + /// \exception FE_UNDERFLOW for subnormal result + inline half nextafter(half from, half to) + { + int fabs = from.data_ & 0x7FFF, tabs = to.data_ & 0x7FFF; + if(fabs > 0x7C00 || tabs > 0x7C00) + return half(detail::binary, detail::signal(from.data_, to.data_)); + if(from.data_ == to.data_ || !(fabs|tabs)) + return to; + if(!fabs) + { + detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT); + return half(detail::binary, (to.data_&0x8000)+1); + } + unsigned int out = from.data_ + (((from.data_>>15)^static_cast( + (from.data_^(0x8000|(0x8000-(from.data_>>15))))<(to.data_^(0x8000|(0x8000-(to.data_>>15))))))<<1) - 1; + detail::raise(FE_OVERFLOW, fabs<0x7C00 && (out&0x7C00)==0x7C00); + detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT && (out&0x7C00)<0x400); + return half(detail::binary, out); + } + + /// Next representable value. + /// **See also:** Documentation for [std::nexttoward](https://en.cppreference.com/w/cpp/numeric/math/nexttoward). + /// \param from value to compute next representable value for + /// \param to direction towards which to compute next value + /// \return next representable value after \a from in direction towards \a to + /// \exception FE_INVALID for signaling NaN + /// \exception FE_OVERFLOW for infinite result from finite argument + /// \exception FE_UNDERFLOW for subnormal result + inline half nexttoward(half from, long double to) + { + int fabs = from.data_ & 0x7FFF; + if(fabs > 0x7C00) + return half(detail::binary, detail::signal(from.data_)); + long double lfrom = static_cast(from); + if(detail::builtin_isnan(to) || lfrom == to) + return half(static_cast(to)); + if(!fabs) + { + detail::raise(FE_UNDERFLOW, !HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT); + return half(detail::binary, (static_cast(detail::builtin_signbit(to))<<15)+1); + } + unsigned int out = from.data_ + (((from.data_>>15)^static_cast(lfrom 0x7C00; } + + /// Check if normal number. + /// **See also:** Documentation for [std::isnormal](https://en.cppreference.com/w/cpp/numeric/math/isnormal). + /// \param arg number to check + /// \retval true if normal number + /// \retval false if either subnormal, zero, infinity or NaN + inline HALF_CONSTEXPR bool isnormal(half arg) { return ((arg.data_&0x7C00)!=0) & ((arg.data_&0x7C00)!=0x7C00); } + + /// Check sign. + /// **See also:** Documentation for [std::signbit](https://en.cppreference.com/w/cpp/numeric/math/signbit). + /// \param arg number to check + /// \retval true for negative number + /// \retval false for positive number + inline HALF_CONSTEXPR bool signbit(half arg) { return (arg.data_&0x8000) != 0; } + + /// \} + /// \anchor compfunc + /// \name Comparison + /// \{ + + /// Quiet comparison for greater than. + /// **See also:** Documentation for [std::isgreater](https://en.cppreference.com/w/cpp/numeric/math/isgreater). + /// \param x first operand + /// \param y second operand + /// \retval true if \a x greater than \a y + /// \retval false else + inline HALF_CONSTEXPR bool isgreater(half x, half y) + { + return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) > ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y); + } + + /// Quiet comparison for greater equal. + /// **See also:** Documentation for [std::isgreaterequal](https://en.cppreference.com/w/cpp/numeric/math/isgreaterequal). + /// \param x first operand + /// \param y second operand + /// \retval true if \a x greater equal \a y + /// \retval false else + inline HALF_CONSTEXPR bool isgreaterequal(half x, half y) + { + return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) >= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y); + } + + /// Quiet comparison for less than. + /// **See also:** Documentation for [std::isless](https://en.cppreference.com/w/cpp/numeric/math/isless). + /// \param x first operand + /// \param y second operand + /// \retval true if \a x less than \a y + /// \retval false else + inline HALF_CONSTEXPR bool isless(half x, half y) + { + return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) < ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y); + } + + /// Quiet comparison for less equal. + /// **See also:** Documentation for [std::islessequal](https://en.cppreference.com/w/cpp/numeric/math/islessequal). + /// \param x first operand + /// \param y second operand + /// \retval true if \a x less equal \a y + /// \retval false else + inline HALF_CONSTEXPR bool islessequal(half x, half y) + { + return ((x.data_^(0x8000|(0x8000-(x.data_>>15))))+(x.data_>>15)) <= ((y.data_^(0x8000|(0x8000-(y.data_>>15))))+(y.data_>>15)) && !isnan(x) && !isnan(y); + } + + /// Quiet comarison for less or greater. + /// **See also:** Documentation for [std::islessgreater](https://en.cppreference.com/w/cpp/numeric/math/islessgreater). + /// \param x first operand + /// \param y second operand + /// \retval true if either less or greater + /// \retval false else + inline HALF_CONSTEXPR bool islessgreater(half x, half y) + { + return x.data_!=y.data_ && ((x.data_|y.data_)&0x7FFF) && !isnan(x) && !isnan(y); + } + + /// Quiet check if unordered. + /// **See also:** Documentation for [std::isunordered](https://en.cppreference.com/w/cpp/numeric/math/isunordered). + /// \param x first operand + /// \param y second operand + /// \retval true if unordered (one or two NaN operands) + /// \retval false else + inline HALF_CONSTEXPR bool isunordered(half x, half y) { return isnan(x) || isnan(y); } + + /// \} + /// \anchor casting + /// \name Casting + /// \{ + + /// Cast to or from half-precision floating-point number. + /// This casts between [half](\ref half_float::half) and any built-in arithmetic type. The values are converted + /// directly using the default rounding mode, without any roundtrip over `float` that a `static_cast` would otherwise do. + /// + /// Using this cast with neither of the two types being a [half](\ref half_float::half) or with any of the two types + /// not being a built-in arithmetic type (apart from [half](\ref half_float::half), of course) results in a compiler + /// error and casting between [half](\ref half_float::half)s returns the argument unmodified. + /// \tparam T destination type (half or built-in arithmetic type) + /// \tparam U source type (half or built-in arithmetic type) + /// \param arg value to cast + /// \return \a arg converted to destination type + /// \exception FE_INVALID if \a T is integer type and result is not representable as \a T + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + template T half_cast(U arg) { return detail::half_caster::cast(arg); } + + /// Cast to or from half-precision floating-point number. + /// This casts between [half](\ref half_float::half) and any built-in arithmetic type. The values are converted + /// directly using the specified rounding mode, without any roundtrip over `float` that a `static_cast` would otherwise do. + /// + /// Using this cast with neither of the two types being a [half](\ref half_float::half) or with any of the two types + /// not being a built-in arithmetic type (apart from [half](\ref half_float::half), of course) results in a compiler + /// error and casting between [half](\ref half_float::half)s returns the argument unmodified. + /// \tparam T destination type (half or built-in arithmetic type) + /// \tparam R rounding mode to use. + /// \tparam U source type (half or built-in arithmetic type) + /// \param arg value to cast + /// \return \a arg converted to destination type + /// \exception FE_INVALID if \a T is integer type and result is not representable as \a T + /// \exception FE_OVERFLOW, ...UNDERFLOW, ...INEXACT according to rounding + template T half_cast(U arg) { return detail::half_caster::cast(arg); } + /// \} + + /// \} + /// \anchor errors + /// \name Error handling + /// \{ + + /// Clear exception flags. + /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled, + /// but in that case manual flag management is the only way to raise flags. + /// + /// **See also:** Documentation for [std::feclearexcept](https://en.cppreference.com/w/cpp/numeric/fenv/feclearexcept). + /// \param excepts OR of exceptions to clear + /// \retval 0 all selected flags cleared successfully + inline int feclearexcept(int excepts) { detail::errflags() &= ~excepts; return 0; } + + /// Test exception flags. + /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled, + /// but in that case manual flag management is the only way to raise flags. + /// + /// **See also:** Documentation for [std::fetestexcept](https://en.cppreference.com/w/cpp/numeric/fenv/fetestexcept). + /// \param excepts OR of exceptions to test + /// \return OR of selected exceptions if raised + inline int fetestexcept(int excepts) { return detail::errflags() & excepts; } + + /// Raise exception flags. + /// This raises the specified floating point exceptions and also invokes any additional automatic exception handling as + /// configured with the [HALF_ERRHANDLIG_...](\ref HALF_ERRHANDLING_ERRNO) preprocessor symbols. + /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled, + /// but in that case manual flag management is the only way to raise flags. + /// + /// **See also:** Documentation for [std::feraiseexcept](https://en.cppreference.com/w/cpp/numeric/fenv/feraiseexcept). + /// \param excepts OR of exceptions to raise + /// \retval 0 all selected exceptions raised successfully + inline int feraiseexcept(int excepts) { detail::errflags() |= excepts; detail::raise(excepts); return 0; } + + /// Save exception flags. + /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled, + /// but in that case manual flag management is the only way to raise flags. + /// + /// **See also:** Documentation for [std::fegetexceptflag](https://en.cppreference.com/w/cpp/numeric/fenv/feexceptflag). + /// \param flagp adress to store flag state at + /// \param excepts OR of flags to save + /// \retval 0 for success + inline int fegetexceptflag(int *flagp, int excepts) { *flagp = detail::errflags() & excepts; return 0; } + + /// Restore exception flags. + /// This only copies the specified exception state (including unset flags) without incurring any additional exception handling. + /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled, + /// but in that case manual flag management is the only way to raise flags. + /// + /// **See also:** Documentation for [std::fesetexceptflag](https://en.cppreference.com/w/cpp/numeric/fenv/feexceptflag). + /// \param flagp adress to take flag state from + /// \param excepts OR of flags to restore + /// \retval 0 for success + inline int fesetexceptflag(const int *flagp, int excepts) { detail::errflags() = (detail::errflags()|(*flagp&excepts)) & (*flagp|~excepts); return 0; } + + /// Throw C++ exceptions based on set exception flags. + /// This function manually throws a corresponding C++ exception if one of the specified flags is set, + /// no matter if automatic throwing (via [HALF_ERRHANDLING_THROW_...](\ref HALF_ERRHANDLING_THROW_INVALID)) is enabled or not. + /// This function works even if [automatic exception flag handling](\ref HALF_ERRHANDLING_FLAGS) is disabled, + /// but in that case manual flag management is the only way to raise flags. + /// \param excepts OR of exceptions to test + /// \param msg error message to use for exception description + /// \throw std::domain_error if `FE_INVALID` or `FE_DIVBYZERO` is selected and set + /// \throw std::overflow_error if `FE_OVERFLOW` is selected and set + /// \throw std::underflow_error if `FE_UNDERFLOW` is selected and set + /// \throw std::range_error if `FE_INEXACT` is selected and set + inline void fethrowexcept(int excepts, const char *msg = "") + { + excepts &= detail::errflags(); + if(excepts & (FE_INVALID|FE_DIVBYZERO)) + throw std::domain_error(msg); + if(excepts & FE_OVERFLOW) + throw std::overflow_error(msg); + if(excepts & FE_UNDERFLOW) + throw std::underflow_error(msg); + if(excepts & FE_INEXACT) + throw std::range_error(msg); + } + /// \} +} + + +#undef HALF_UNUSED_NOERR +#undef HALF_CONSTEXPR +#undef HALF_CONSTEXPR_CONST +#undef HALF_CONSTEXPR_NOERR +#undef HALF_NOEXCEPT +#undef HALF_NOTHROW +#undef HALF_THREAD_LOCAL +#undef HALF_TWOS_COMPLEMENT_INT +#ifdef HALF_POP_WARNINGS + #pragma warning(pop) + #undef HALF_POP_WARNINGS +#endif + +#endif diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index f6f01790..240227f0 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -10,11 +10,11 @@ shared float g_scratchDepths[8][8]; // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl layout (set = 1, binding = 0) uniform sampler2D depthImage; -layout (r32f, set = 1, binding = 1) uniform image2D outDepth0; -layout (r32f, set = 1, binding = 2) uniform image2D outDepth1; -layout (r32f, set = 1, binding = 3) uniform image2D outDepth2; -layout (r32f, set = 1, binding = 4) uniform image2D outDepth3; -layout (r32f, set = 1, binding = 5) uniform image2D outDepth4; +layout (r16f, set = 1, binding = 1) uniform image2D outDepth0; +layout (r16f, set = 1, binding = 2) uniform image2D outDepth1; +layout (r16f, set = 1, binding = 3) uniform image2D outDepth2; +layout (r16f, set = 1, binding = 4) uniform image2D outDepth3; +layout (r16f, set = 1, binding = 5) uniform image2D outDepth4; layout (push_constant) uniform PushConstants { vec2 ndcToViewMult; diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index 421a7cac..cf81f709 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -21,6 +21,7 @@ #include "src/core/game_object/renderable.h" #include "src/core/scene/serializer.h" #include "src/util/file.h" +#include "src/util/math_utils.h" namespace will_engine { @@ -316,7 +317,8 @@ void ImguiWrapper::imguiInterface(Engine* engine) }; vk_helpers::saveImageR32F(*engine->resourceManager, *engine->immediate, engine->depthImage, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_DEPTH_BIT, path.string().c_str(), depthNormalize); + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_DEPTH_BIT, path.string().c_str(), + depthNormalize); } else { fmt::print(" Failed to find/create image save path directory"); @@ -335,7 +337,8 @@ void ImguiWrapper::imguiInterface(Engine* engine) return pixel; }; vk_helpers::savePacked32Bit(*engine->resourceManager, *engine->immediate, engine->normalRenderTarget, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), unpackFunc); + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), + unpackFunc); } else { fmt::print(" Failed to save normal render target"); @@ -351,7 +354,8 @@ void ImguiWrapper::imguiInterface(Engine* engine) return pixel; }; vk_helpers::savePacked32Bit(*engine->resourceManager, *engine->immediate, engine->albedoRenderTarget, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), unpackFunc); + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), + unpackFunc); } else { fmt::print(" Failed to save albedo render target"); @@ -367,7 +371,8 @@ void ImguiWrapper::imguiInterface(Engine* engine) return pixel; }; vk_helpers::savePacked32Bit(*engine->resourceManager, *engine->immediate, engine->pbrRenderTarget, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), unpackFunc); + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), + unpackFunc); } else { fmt::print(" Failed to save pbr render target"); @@ -705,8 +710,9 @@ void ImguiWrapper::imguiInterface(Engine* engine) Texture* randomTexture = engine->assetManager->getAnyTexture(); currentlySelectedTexture = randomTexture->getTextureResource(); - currentlySelectedTextureImguiId = ImGui_ImplVulkan_AddTexture(engine->resourceManager->getDefaultSamplerLinear(), currentlySelectedTexture->getTexture().imageView, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); + currentlySelectedTextureImguiId = ImGui_ImplVulkan_AddTexture( + engine->resourceManager->getDefaultSamplerLinear(), currentlySelectedTexture->getTexture().imageView, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); } } @@ -723,7 +729,8 @@ void ImguiWrapper::imguiInterface(Engine* engine) for (Texture* texture : engine->assetManager->getAllTextures()) { bool isSelected = (currentlySelectedTexture->getId() == texture->getId()); - std::string label = fmt::format("[{}] Texture ID - {}", texture->isTextureResourceLoaded() ? "LOADED" : "NOT LOADED", texture->getId()); + std::string label = fmt::format("[{}] Texture ID - {}", texture->isTextureResourceLoaded() ? "LOADED" : "NOT LOADED", + texture->getId()); if (ImGui::Selectable(label.c_str(), isSelected)) { if (currentlySelectedTextureImguiId != VK_NULL_HANDLE) { @@ -734,8 +741,9 @@ void ImguiWrapper::imguiInterface(Engine* engine) } currentlySelectedTexture = texture->getTextureResource(); - currentlySelectedTextureImguiId = ImGui_ImplVulkan_AddTexture(engine->resourceManager->getDefaultSamplerLinear(), currentlySelectedTexture->getTexture().imageView, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); + currentlySelectedTextureImguiId = ImGui_ImplVulkan_AddTexture( + engine->resourceManager->getDefaultSamplerLinear(), currentlySelectedTexture->getTexture().imageView, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); } } @@ -879,15 +887,17 @@ void ImguiWrapper::imguiInterface(Engine* engine) if (file::getOrCreateDirectory(file::imagesSavePath)) { const std::filesystem::path path = file::imagesSavePath / "gtao_depth.png"; - auto depthNormalize = [](const float depth) { - return depth / 1000.f; + auto depthNormalize = [](const uint16_t depth) { + // Equivalent + float manualDepth = math::halfToFloat(depth); + float libraryDepth = half_float::detail::half2float(depth); + return libraryDepth / 1000.f; }; - vk_helpers::saveImageR32F( + vk_helpers::saveImageR16F( *engine->resourceManager, *engine->immediate, engine->ambientOcclusionPipeline->depthPrefilterImage, - //engine->depthImage, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT, path.string().c_str(), diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index fc22b271..7ee4589e 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -351,7 +351,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V //vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::clearColorImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL); // Depth Prefilter - {; + { vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipeline); vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); @@ -377,7 +377,6 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); // vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - // // // Ambient Occlusion // { // vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); @@ -398,12 +397,11 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V // const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); // const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); // vkCmdDispatch(cmd, x, y, 1); - // vkCmdEndRendering(cmd); // } // // // vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - // + vkCmdEndDebugUtilsLabelEXT(cmd); } diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index f45c6a72..60ae2cbb 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -42,7 +42,7 @@ class GroundTruthAmbientOcclusionPipeline VkSampler depthPrefilterSampler{VK_NULL_HANDLE}; // 16 vs 32. look at cost later. - VkFormat depthPrefilterFormat{VK_FORMAT_R32_SFLOAT}; + VkFormat depthPrefilterFormat{VK_FORMAT_R16_SFLOAT}; AllocatedImage depthPrefilterImage{VK_NULL_HANDLE}; std::array depthPrefilterImageViews{}; diff --git a/src/renderer/vk_helpers.cpp b/src/renderer/vk_helpers.cpp index 73e5e9e0..834dfb50 100644 --- a/src/renderer/vk_helpers.cpp +++ b/src/renderer/vk_helpers.cpp @@ -9,7 +9,7 @@ #include #include #include "volk/volk.h" -#include "extern/half/half/half.hpp" +#include "extern/half/half.hpp" #include "immediate_submitter.h" #include "resource_manager.h" @@ -777,47 +777,49 @@ void will_engine::vk_helpers::saveImageR32F(const ResourceManager& resourceManag } void will_engine::vk_helpers::saveImageR16F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, - VkImageAspectFlags aspectFlag, const char* savePath, const std::function& valueTransform) + VkImageAspectFlags aspectFlag, const char* savePath, const std::function& valueTransform, int32_t mipLevel) { - using half_float::half; - const size_t dataSize = image.imageExtent.width * image.imageExtent.height * 1 * sizeof(half); + size_t newXSize = image.imageExtent.width / static_cast(std::pow(2, mipLevel)); + size_t newYSize = image.imageExtent.height / static_cast(std::pow(2, mipLevel)); + const size_t texelCount = newXSize * newYSize; + const size_t dataSize = texelCount * 1 * sizeof(uint16_t); AllocatedBuffer receivingBuffer = resourceManager.createReceivingBuffer(dataSize); - immediate.submit([&](VkCommandBuffer cmd) { + immediate.submit([&, mipLevel](VkCommandBuffer cmd) { VkBufferImageCopy bufferCopyRegion{}; bufferCopyRegion.imageSubresource.aspectMask = aspectFlag; - bufferCopyRegion.imageSubresource.mipLevel = 0; + bufferCopyRegion.imageSubresource.mipLevel = mipLevel; bufferCopyRegion.imageSubresource.baseArrayLayer = 0; bufferCopyRegion.imageSubresource.layerCount = 1; - bufferCopyRegion.imageExtent = image.imageExtent; + bufferCopyRegion.imageExtent = {static_cast(newXSize), static_cast(newYSize), 1u}; bufferCopyRegion.bufferOffset = 0; bufferCopyRegion.bufferRowLength = 0; bufferCopyRegion.bufferImageHeight = 0; - vk_helpers::transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, aspectFlag); + transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, aspectFlag); vkCmdCopyImageToBuffer(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, receivingBuffer.buffer, 1, &bufferCopyRegion); - vk_helpers::transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, aspectFlag); + transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, aspectFlag); }); void* data = receivingBuffer.info.pMappedData; - const auto imageData = static_cast(data); + const auto imageData = static_cast(data); - const auto byteImageData = new uint8_t[image.imageExtent.width * image.imageExtent.height * 4]; + + const auto byteImageData = new uint8_t[texelCount * 4]; const auto powEight = static_cast(pow(2, 8) - 1); - for (size_t i = 0; i < image.imageExtent.width * image.imageExtent.height; ++i) { - half originalData = imageData[i]; - float floatData = half_float::detail::half2float(originalData); - const float halfValue = valueTransform(originalData); - const auto value = static_cast(halfValue * powEight); + for (size_t i = 0; i < texelCount; ++i) { + const uint16_t rvalue = imageData[i]; + const float floatValue = valueTransform(rvalue); + const auto value = static_cast(floatValue * powEight); byteImageData[i * 4 + 0] = value; byteImageData[i * 4 + 1] = value; byteImageData[i * 4 + 2] = value; byteImageData[i * 4 + 3] = 255; } - stbi_write_png(savePath, image.imageExtent.width, image.imageExtent.height, 4, byteImageData, image.imageExtent.width * 4); + stbi_write_png(savePath, static_cast(newXSize), static_cast(newYSize), 4, byteImageData, static_cast(newXSize) * 4); delete[] byteImageData; resourceManager.destroyBuffer(receivingBuffer); diff --git a/src/renderer/vk_helpers.h b/src/renderer/vk_helpers.h index 3a5358ce..ad8c6455 100644 --- a/src/renderer/vk_helpers.h +++ b/src/renderer/vk_helpers.h @@ -12,7 +12,7 @@ #include #include -#include +#include #include "vk_types.h" @@ -55,9 +55,11 @@ namespace vk_helpers VkRenderingAttachmentInfo attachmentInfo(VkImageView view, const VkClearValue* clear, VkImageLayout layout /*= VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL*/); - VkRenderingInfo renderingInfo(VkExtent2D renderExtent, const VkRenderingAttachmentInfo* colorAttachment, const VkRenderingAttachmentInfo* depthAttachment); + VkRenderingInfo renderingInfo(VkExtent2D renderExtent, const VkRenderingAttachmentInfo* colorAttachment, + const VkRenderingAttachmentInfo* depthAttachment); - VkSubmitInfo2 submitInfo(const VkCommandBufferSubmitInfo* cmd, const VkSemaphoreSubmitInfo* signalSemaphoreInfo, const VkSemaphoreSubmitInfo* waitSemaphoreInfo); + VkSubmitInfo2 submitInfo(const VkCommandBufferSubmitInfo* cmd, const VkSemaphoreSubmitInfo* signalSemaphoreInfo, + const VkSemaphoreSubmitInfo* waitSemaphoreInfo); VkPresentInfoKHR presentInfo(); @@ -78,7 +80,8 @@ namespace vk_helpers */ VkDeviceSize getAlignedSize(VkDeviceSize value, VkDeviceSize alignment); - void clearColorImage(VkCommandBuffer cmd, VkImage image, VkImageLayout srcLayout, VkImageLayout dstLayout, VkClearColorValue clearColor = {0.0f, 0.0f, 0.0f, 1.0f}); + void clearColorImage(VkCommandBuffer cmd, VkImage image, VkImageLayout srcLayout, VkImageLayout dstLayout, + VkClearColorValue clearColor = {0.0f, 0.0f, 0.0f, 1.0f}); void transitionImage(VkCommandBuffer cmd, VkImage image, VkImageLayout currentLayout, VkImageLayout targetLayout, VkImageAspectFlags aspectMask); @@ -98,26 +101,30 @@ namespace vk_helpers VkPipelineShaderStageCreateInfo pipelineShaderStageCreateInfo(VkShaderStageFlagBits stage, VkShaderModule shaderModule, const char* entry = "main"); - void saveImageRGBA32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, bool overrideAlpha = true); + void saveImageRGBA32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, bool overrideAlpha = true); - void saveImageRGBA16SFLOAT(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, bool overrideAlpha = true); + void saveImageRGBA16SFLOAT(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, bool overrideAlpha = true); - void savePacked32Bit(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, const std::function& unpackingFunction); + void savePacked32Bit(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, + const std::function& unpackingFunction); - void savePacked64Bit(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, const std::function& unpackingFunction); + void savePacked64Bit(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, + const std::function& unpackingFunction); /** * Save the Allocated image as a grayscaled image. The image must be a format with only 1 channel (e.g. R32 or D32) */ - void saveImageR32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, const std::function& valueTransform, int32_t mipLevel = 0); + void saveImageR32F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, + const std::function& valueTransform, int32_t mipLevel = 0); - void saveImageR16F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, - const char* savePath, const std::function& valueTransform); + void saveImageR16F(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, + const std::function& valueTransform, int32_t mipLevel = 0); void saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha = true); diff --git a/src/util/math_utils.h b/src/util/math_utils.h index ab4f8572..d4667f67 100644 --- a/src/util/math_utils.h +++ b/src/util/math_utils.h @@ -25,6 +25,32 @@ inline void decomposeMatrix(const glm::mat4& matrix, glm::vec3& position, glm::q ); rotation = quat_cast(rotMat); } + +inline uint32_t as_uint(const float x) +{ + return *(uint32_t*) &x; +} + +inline float as_float(uint32_t x) +{ + return *(float*) &x; +} + +/** + * https://stackoverflow.com/a/60047308 + * @param x + * @return + */ +inline float halfToFloat(const uint16_t x) +{ + // IEEE-754 16-bit floating-point format (without infinity): 1-5-10, exp-15, +-131008.0, +-6.1035156E-5, +-5.9604645E-8, 3.311 digits + const uint32_t e = (x & 0x7C00) >> 10; // exponent + const uint32_t m = (x & 0x03FF) << 13; // mantissa + const uint32_t v = as_uint((float) m) >> 23; // evil log2 bit hack to count leading zeros in denormalized format + return as_float( + (x & 0x8000) << 16 | (e != 0) * ((e + 112) << 23 | m) | ((e == 0) & (m != 0)) * ( + (v - 37) << 23 | ((m << (150 - v)) & 0x007FE000))); // sign : normalized : denormalized +} } #endif //MATH_UTILS_H From 26e8586c928046ca53e8e19a3df6c6b851ef2d00 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Fri, 28 Mar 2025 21:05:06 +0700 Subject: [PATCH 15/27] Half License. GTAO main pass setup. --- licenses/half/ChangeLog.txt | 222 ++++++++++++ licenses/half/{LICENCE => LICENSE.txt} | 2 +- licenses/half/README.md | 13 - licenses/half/README.txt | 317 ++++++++++++++++++ .../ground_truth/gtao_main_pass.comp | 28 ++ src/renderer/imgui_wrapper.cpp | 18 + .../ground_truth_ambient_occlusion.cpp | 58 ++-- .../deferred_resolve/deferred_resolve.cpp | 42 +-- src/renderer/vk_helpers.cpp | 47 +++ src/renderer/vk_helpers.h | 3 + 10 files changed, 687 insertions(+), 63 deletions(-) create mode 100644 licenses/half/ChangeLog.txt rename licenses/half/{LICENCE => LICENSE.txt} (96%) delete mode 100644 licenses/half/README.md create mode 100644 licenses/half/README.txt diff --git a/licenses/half/ChangeLog.txt b/licenses/half/ChangeLog.txt new file mode 100644 index 00000000..766603c9 --- /dev/null +++ b/licenses/half/ChangeLog.txt @@ -0,0 +1,222 @@ +Release Notes +============= + +2.2.1 release (2025-03-02): +--------------------------- + +- Fixed error when converting minimum negative integer value to half. +- Added more type conversions to remove warnings on some 64-bit platforms. +- Removed potential naming conflicts of internal functions with C standard + library functions. + + +2.2.0 release (2021-06-12): +--------------------------- + +- Added `rsqrt` function for inverse square root. +- Improved performance of `pow` function. +- Fixed bug that forgot to include `` for F16C intrinsics. + + +2.1.0 release (2019-08-05): +--------------------------- + +- Added detection of IEEE floating-point exceptions to operators and functions. +- Added configuration options for automatic exception handling. +- Added functions for explicitly managing floating-point exception flags. +- Improved accuracy of `pow` and `atan2` functions. + + +2.0.0 release (2019-07-23): +--------------------------- + +- Made internal implementation independent from built-in floating point + facilities for increased reliability and IEEE-conformance. +- Changed default rounding mode to rounding to nearest. +- Always round ties to even when rounding to nearest. +- Extended `constexpr` support to comparison and classification functions. +- Added support for F16C compiler intrinsics for conversions. +- Enabled C++11 feature detection for Intel compilers. + + +1.12.0 release (2017-03-06): +---------------------------- + +- Changed behaviour of `half_cast` to perform conversions to/from `double` + and `long double` directly according to specified rounding mode, without an + intermediate `float` conversion. +- Added `noexcept` specifiers to constructors. +- Fixed minor portability problem with `logb` and `ilogb`. +- Tested for *VC++ 2015*. + + +1.11.0 release (2013-11-16): +---------------------------- + +- Made tie-breaking behaviour in round to nearest configurable by + `HALF_ROUND_TIES_TO_EVEN` macro. +- Completed support for all C++11 mathematical functions even if single- + precision versions from `` are unsupported. +- Fixed inability to disable support for C++11 mathematical functions on + *VC++ 2013*. + + +1.10.0 release (2013-11-09): +---------------------------- + +- Made default rounding mode configurable by `HALF_ROUND_STYLE` macro. +- Added support for non-IEEE single-precision implementations. +- Added `HALF_ENABLE_CPP11_TYPE_TRAITS` preprocessor flag for checking + support for C++11 type traits and TMP features. +- Restricted `half_cast` to support built-in arithmetic types only. +- Changed behaviour of `half_cast` to respect rounding mode when casting + to/from integer types. + + +1.9.2 release (2013-11-01): +--------------------------- + +- Tested for *gcc 4.8*. +- Tested and fixed for *VC++ 2013*. +- Removed unnecessary warnings in *MSVC*. + + +1.9.1 release (2013-08-08): +--------------------------- + +- Fixed problems with older gcc and MSVC versions. +- Small fix to non-C++11 implementations of `remainder` and `remquo`. + + +1.9.0 release (2013-08-07): +--------------------------- + +- Changed behaviour of `nearbyint`, `rint`, `lrint` and `llrint` to use + rounding mode of half-precision implementation (which is + truncating/indeterminate) instead of single-precision rounding mode. +- Added support for more C++11 mathematical functions even if single- + precision versions from `` are unsupported, in particular + `remainder`, `remquo` and `cbrt`. +- Minor implementation changes. + + +1.8.1 release (2013-01-22): +--------------------------- + +- Fixed bug resulting in multiple definitions of the `nanh` function due to + a missing `inline` specification. + + +1.8.0 release (2013-01-19): +--------------------------- + +- Added support for more C++11 mathematical functions even if single- + precision versions from `` are unsupported, in particular + exponential and logarithm functions, hyperbolic area functions and the + hypotenuse function. +- Made `fma` function use default implementation if single-precision version + from `` is not faster and thus `FP_FAST_FMAH` to be defined always. +- Fixed overload resolution issues when invoking certain mathematical + functions by unqualified calls. + + +1.7.0 release (2012-10-26): +--------------------------- + +- Added support for C++11 `noexcept` specifiers. +- Changed C++11 `long long` to be supported on *VC++ 2003* and up. + + +1.6.1 release (2012-09-13): +--------------------------- + +- Made `fma` and `fdim` functions available even if corresponding + single-precision functions are not. + + +1.6.0 release (2012-09-12): +--------------------------- + +- Added `HALF_ENABLE_CPP11_LONG_LONG` to control support for `long long` + integers and corresponding mathematical functions. +- Fixed C++98 compatibility on non-VC compilers. + + +1.5.1 release (2012-08-17): +--------------------------- + +- Recorrected `std::numeric_limits::round_style` to always return + `std::round_indeterminate`, due to overflow-handling deviating from + correct round-toward-zero behaviour. + + +1.5.0 release (2012-08-16): +--------------------------- + +- Added `half_cast` for explicitly casting between half and any type + convertible to/from `float` and allowing the explicit specification of + the rounding mode to use. + + +1.4.0 release (2012-08-12): +--------------------------- + +- Added support for C++11 generalized constant expressions (`constexpr`). + + +1.3.1 release (2012-08-11): +--------------------------- + +- Fixed requirement for `std::signbit` and `std::isnan` (even if C++11 + `` functions disabled) on non-VC compilers. + + +1.3.0 release (2012-08-10): +--------------------------- + +- Made requirement for `` and `static_assert` optional and thus + made the library C++98-compatible. +- Made support for C++11 features user-overridable through explicit + definition of corresponding preprocessor symbols to either 0 or 1. +- Renamed `HALF_ENABLE_HASH` to `HALF_ENABLE_CPP11_HASH` in correspondence + with other C++11 preprocessor symbols. + + +1.2.0 release (2012-08-07): +--------------------------- + +- Added proper preprocessor definitions for `HUGE_VALH` and `FP_FAST_FMAH` + in correspondence with their single-precision counterparts from ``. +- Fixed internal preprocessor macros to be properly undefined after use. + + +1.1.2 release (2012-08-07): +--------------------------- + +- Revised `std::numeric_limits::round_style` to return + `std::round_toward_zero` if the `float` version also does and + `std::round_indeterminate` otherwise. +- Fixed `std::numeric_limits::round_error` to reflect worst-case round + toward zero behaviour. + + +1.1.1 release (2012-08-06): +--------------------------- + +- Fixed `std::numeric_limits::min` to return smallest positive normal + number, instead of subnormal number. +- Fixed `std::numeric_limits::round_style` to return + `std::round_indeterminate` due to mixture of separately rounded + single-precision arithmetics with truncating single-to-half conversions. + + +1.1.0 release (2012-08-06): +--------------------------- + +- Added half-precision literals. + + +1.0.0 release (2012-08-05): +--------------------------- + +- First release. diff --git a/licenses/half/LICENCE b/licenses/half/LICENSE.txt similarity index 96% rename from licenses/half/LICENCE rename to licenses/half/LICENSE.txt index abee50b1..8579da1b 100644 --- a/licenses/half/LICENCE +++ b/licenses/half/LICENSE.txt @@ -1,6 +1,6 @@ The MIT License -Copyright (c) 2012-2017 Christian Rau +Copyright (c) 2012-2025 Christian Rau Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal diff --git a/licenses/half/README.md b/licenses/half/README.md deleted file mode 100644 index 52306d39..00000000 --- a/licenses/half/README.md +++ /dev/null @@ -1,13 +0,0 @@ -# Upstream - -This library was developed by CHRISTIAN RAU. Upstream repository is a https://sourceforge.net/projects/half/ . -Only half.hpp header file has been imported from upstream. - -# License - -Original upstream's MIT License. See LICENSE file in the repository root. - -# Update - -This repository should be kept up-to-date with the upstream. - diff --git a/licenses/half/README.txt b/licenses/half/README.txt new file mode 100644 index 00000000..6a97fb42 --- /dev/null +++ b/licenses/half/README.txt @@ -0,0 +1,317 @@ +HALF-PRECISION FLOATING-POINT LIBRARY (Version 2.2.1) +----------------------------------------------------- + +This is a C++ header-only library to provide an IEEE 754 conformant 16-bit +half-precision floating-point type along with corresponding arithmetic +operators, type conversions and common mathematical functions. It aims for both +efficiency and ease of use, trying to accurately mimic the behaviour of the +built-in floating-point types at the best performance possible. + + +INSTALLATION AND REQUIREMENTS +----------------------------- + +Conveniently, the library consists of just a single header file containing all +the functionality, which can be directly included by your projects, without the +neccessity to build anything or link to anything. + +Whereas this library is fully C++98-compatible, it can profit from certain +C++11 features. Support for those features is checked automatically at compile +(or rather preprocessing) time, but can be explicitly enabled or disabled by +predefining the corresponding preprocessor symbols to either 1 or 0 yourself +before including half.hpp. This is useful when the automatic detection fails +(for more exotic implementations) or when a feature should be explicitly +disabled: + + - 'long long' integer type for mathematical functions returning 'long long' + results (enabled for VC++ 2003 and icc 11.1 and newer, gcc and clang, + overridable with 'HALF_ENABLE_CPP11_LONG_LONG'). + + - Static assertions for extended compile-time checks (enabled for VC++ 2010, + gcc 4.3, clang 2.9, icc 11.1 and newer, overridable with + 'HALF_ENABLE_CPP11_STATIC_ASSERT'). + + - Generalized constant expressions (enabled for VC++ 2015, gcc 4.6, clang 3.1, + icc 14.0 and newer, overridable with 'HALF_ENABLE_CPP11_CONSTEXPR'). + + - noexcept exception specifications (enabled for VC++ 2015, gcc 4.6, + clang 3.0, icc 14.0 and newer, overridable with 'HALF_ENABLE_CPP11_NOEXCEPT'). + + - User-defined literals for half-precision literals to work (enabled for + VC++ 2015, gcc 4.7, clang 3.1, icc 15.0 and newer, overridable with + 'HALF_ENABLE_CPP11_USER_LITERALS'). + + - Thread-local storage for per-thread floating-point exception flags (enabled + for VC++ 2015, gcc 4.8, clang 3.3, icc 15.0 and newer, overridable with + 'HALF_ENABLE_CPP11_THREAD_LOCAL'). + + - Type traits and template meta-programming features from + (enabled for VC++ 2010, libstdc++ 4.3, libc++ and newer, overridable with + 'HALF_ENABLE_CPP11_TYPE_TRAITS'). + + - Special integer types from (enabled for VC++ 2010, libstdc++ 4.3, + libc++ and newer, overridable with 'HALF_ENABLE_CPP11_CSTDINT'). + + - Certain C++11 single-precision mathematical functions from for + floating-point classification during conversions from higher precision types + (enabled for VC++ 2013, libstdc++ 4.3, libc++ and newer, overridable with + 'HALF_ENABLE_CPP11_CMATH'). + + - Floating-point environment control from for possible exception + propagation to the built-in floating-point platform (enabled for VC++ 2013, + libstdc++ 4.3, libc++ and newer, overridable with 'HALF_ENABLE_CPP11_CFENV'). + + - Hash functor 'std::hash' from (enabled for VC++ 2010, + libstdc++ 4.3, libc++ and newer, overridable with 'HALF_ENABLE_CPP11_HASH'). + +The library has been tested successfully with Visual C++ 2005-2015, gcc 4-8 +and clang 3-8 on 32- and 64-bit x86 systems. Please contact me if you have any +problems, suggestions or even just success testing it on other platforms. + + +DOCUMENTATION +------------- + +What follows are some general words about the usage of the library and its +implementation. For a complete documentation of its interface consult the +corresponding website http://half.sourceforge.net. You may also generate the +complete developer documentation from the library's only include file's doxygen +comments, but this is more relevant to developers rather than mere users. + +BASIC USAGE + +To make use of the library just include its only header file half.hpp, which +defines all half-precision functionality inside the 'half_float' namespace. The +actual 16-bit half-precision data type is represented by the 'half' type, which +uses the standard IEEE representation with 1 sign bit, 5 exponent bits and 11 +mantissa bits (including the hidden bit) and supports all types of special +values, like subnormal values, infinity and NaNs. This type behaves like the +built-in floating-point types as much as possible, supporting the usual +arithmetic, comparison and streaming operators, which makes its use pretty +straight-forward: + + using half_float::half; + half a(3.4), b(5); + half c = a * b; + c += 3; + if(c > a) + std::cout << c << std::endl; + +Additionally the 'half_float' namespace also defines half-precision versions +for all mathematical functions of the C++ standard library, which can be used +directly through ADL: + + half a(-3.14159); + half s = sin(abs(a)); + long l = lround(s); + +You may also specify explicit half-precision literals, since the library +provides a user-defined literal inside the 'half_float::literal' namespace, +which you just need to import (assuming support for C++11 user-defined literals): + + using namespace half_float::literal; + half x = 1.0_h; + +Furthermore the library provides proper specializations for +'std::numeric_limits', defining various implementation properties, and +'std::hash' for hashing half-precision numbers (assuming support for C++11 +'std::hash'). Similar to the corresponding preprocessor symbols from +the library also defines the 'HUGE_VALH' constant and maybe the 'FP_FAST_FMAH' +symbol. + +CONVERSIONS AND ROUNDING + +The half is explicitly constructible/convertible from a single-precision float +argument. Thus it is also explicitly constructible/convertible from any type +implicitly convertible to float, but constructing it from types like double or +int will involve the usual warnings arising when implicitly converting those to +float because of the lost precision. On the one hand those warnings are +intentional, because converting those types to half neccessarily also reduces +precision. But on the other hand they are raised for explicit conversions from +those types, when the user knows what they are doing. So if those warnings keep +bugging you, then you won't get around first explicitly converting to float +before converting to half, or use the 'half_cast' described below. In addition +you can also directly assign float values to halfs. + +In contrast to the float-to-half conversion, which reduces precision, the +conversion from half to float (and thus to any other type implicitly +convertible from float) is implicit, because all values represetable with +half-precision are also representable with single-precision. This way the +half-to-float conversion behaves similar to the builtin float-to-double +conversion and all arithmetic expressions involving both half-precision and +single-precision arguments will be of single-precision type. This way you can +also directly use the mathematical functions of the C++ standard library, +though in this case you will invoke the single-precision versions which will +also return single-precision values, which is (even if maybe performing the +exact same computation, see below) not as conceptually clean when working in a +half-precision environment. + +The default rounding mode for conversions between half and more precise types +as well as for rounding results of arithmetic operations and mathematical +functions rounds to the nearest representable value. But by predefining the +'HALF_ROUND_STYLE' preprocessor symbol this default can be overridden with one +of the other standard rounding modes using their respective constants or the +equivalent values of 'std::float_round_style' (it can even be synchronized with +the built-in single-precision implementation by defining it to +'std::numeric_limits::round_style'): + + - 'std::round_indeterminate' (-1) for the fastest rounding. + + - 'std::round_toward_zero' (0) for rounding toward zero. + + - 'std::round_to_nearest' (1) for rounding to the nearest value (default). + + - 'std::round_toward_infinity' (2) for rounding toward positive infinity. + + - 'std::round_toward_neg_infinity' (3) for rounding toward negative infinity. + +In addition to changing the overall default rounding mode one can also use the +'half_cast'. This converts between half and any built-in arithmetic type using +a configurable rounding mode (or the default rounding mode if none is +specified). In addition to a configurable rounding mode, 'half_cast' has +another big difference to a mere 'static_cast': Any conversions are performed +directly using the given rounding mode, without any intermediate conversion +to/from 'float'. This is especially relevant for conversions to integer types, +which don't necessarily truncate anymore. But also for conversions from +'double' or 'long double' this may produce more precise results than a +pre-conversion to 'float' using the single-precision implementation's current +rounding mode would. + + half a = half_cast(4.2); + half b = half_cast::round_style>(4.2f); + assert( half_cast( 0.7_h ) == 1 ); + assert( half_cast( 4097 ) == 4096.0_h ); + assert( half_cast( 4097 ) == 4100.0_h ); + assert( half_cast( std::numeric_limits::min() ) > 0.0_h ); + +ACCURACY AND PERFORMANCE + +From version 2.0 onward the library is implemented without employing the +underlying floating-point implementation of the system (except for conversions, +of course), providing an entirely self-contained half-precision implementation +with results independent from the system's existing single- or double-precision +implementation and its rounding behaviour. + +As to accuracy, many of the operators and functions provided by this library +are exact to rounding for all rounding modes, i.e. the error to the exact +result is at most 0.5 ULP (unit in the last place) for rounding to nearest and +less than 1 ULP for all other rounding modes. This holds for all the operations +required by the IEEE 754 standard and many more. Specifically the following +functions might exhibit a deviation from the correctly rounded exact result by +1 ULP for a select few input values: 'expm1', 'log1p', 'pow', 'atan2', 'erf', +'erfc', 'lgamma', 'tgamma' (for more details see the documentation of the +individual functions). All other functions and operators are always exact to +rounding or independent of the rounding mode altogether. + +The increased IEEE-conformance and cleanliness of this implementation comes +with a certain performance cost compared to doing computations and mathematical +functions in hardware-accelerated single-precision. On average and depending on +the platform, the arithemtic operators are about 75% as fast and the +mathematical functions about 33-50% as fast as performing the corresponding +operations in single-precision and converting between the inputs and outputs. +However, directly computing with half-precision values is a rather rare +use-case and usually using actual 'float' values for all computations and +temproraries and using 'half's only for storage is the recommended way. But +nevertheless the goal of this library was to provide a complete and +conceptually clean IEEE-confromant half-precision implementation and in the few +cases when you do need to compute directly in half-precision you do so for a +reason and want accurate results. + +If necessary, this internal implementation can be overridden by predefining the +'HALF_ARITHMETIC_TYPE' preprocessor symbol to one of the built-in +floating-point types ('float', 'double' or 'long double'), which will cause the +library to use this type for computing arithmetic operations and mathematical +functions (if available). However, due to using the platform's floating-point +implementation (and its rounding behaviour) internally, this might cause +results to deviate from the specified half-precision rounding mode. It will of +course also inhibit the automatic exception detection described below. + +The conversion operations between half-precision and single-precision types can +also make use of the F16C extension for x86 processors by using the +corresponding compiler intrinsics from . Support for this is +checked at compile-time by looking for the '__F16C__' macro which at least gcc +and clang define based on the target platform. It can also be enabled manually +by predefining the 'HALF_ENABLE_F16C_INTRINSICS' preprocessor symbol to 1, or 0 +for explicitly disabling it. However, this will directly use the corresponding +intrinsics for conversion without checking if they are available at runtime +(possibly crashing if they are not), so make sure they are supported on the +target platform before enabling this. + +EXCEPTION HANDLING + +The half-precision implementation supports all 5 required floating-point +exceptions from the IEEE standard to indicate erroneous inputs or inexact +results during operations. These are represented by exception flags which +actually use the same values as the corresponding 'FE_...' flags defined in +C++11's header if supported, specifically: + + - 'FE_INVALID' for invalid inputs to an operation. + - 'FE_DIVBYZERO' for finite inputs producing infinite results. + - 'FE_OVERFLOW' if a result is too large to represent finitely. + - 'FE_UNDERFLOW' for a subnormal or zero result after rounding. + - 'FE_INEXACT' if a result needed rounding to be representable. + - 'FE_ALL_EXCEPT' as a convenient OR of all possible exception flags. + +The internal exception flag state will start with all flags cleared and is +maintained per thread if C++11 thread-local storage is supported, otherwise it +will be maintained globally and will theoretically NOT be thread-safe (while +practically being as thread-safe as a simple integer variable can be). These +flags can be managed explicitly using the library's error handling functions, +which again try to mimic the built-in functions for handling floating-point +exceptions from . You can clear them with 'feclearexcept' (which is the +only way a flag can be cleared), test them with 'fetestexcept', explicitly +raise errors with 'feraiseexcept' and save and restore their state using +'fegetexceptflag' and 'fesetexceptflag'. You can also throw corresponding C++ +exceptions based on the current flag state using 'fethrowexcept'. + +However, any automatic exception detection and handling during half-precision +operations and functions is DISABLED by default, since it comes with a minor +performance overhead due to runtime checks, and reacting to IEEE floating-point +exceptions is rarely ever needed in application code. But the library fully +supports IEEE-conformant detection of floating-point exceptions and various +ways for handling them, which can be enabled by pre-defining the corresponding +preprocessor symbols to 1. They can be enabled individually or all at once and +they will be processed in the order they are listed here: + + - 'HALF_ERRHANDLING_FLAGS' sets the internal exception flags described above + whenever the corresponding exception occurs. + - 'HALF_ERRHANDLING_ERRNO' sets the value of 'errno' from similar to + the behaviour of the built-in floating-point types when 'MATH_ERRNO' is used. + - 'HALF_ERRHANDLING_FENV' will propagate exceptions to the built-in + floating-point implementation using 'std::feraiseexcept' if support for + C++11 floating-point control is enabled. However, this does not synchronize + exceptions: neither will clearing propagate nor will it work in reverse. + - 'HALF_ERRHANDLING_THROW_...' can be defined to a string literal which will + be used as description message for a C++ exception that is thrown whenever + a 'FE_...' exception occurs, similar to the behaviour of 'fethrowexcept'. + +If any of the above error handling is activated, non-quiet operations on +half-precision values will also raise a 'FE_INVALID' exception whenever +they encounter a signaling NaN value, in addition to transforming the value +into a quiet NaN. If error handling is disabled, signaling NaNs will be +treated like quiet NaNs (while still getting explicitly quieted if propagated +to the result). There can also be additional treatment of overflow and +underflow errors after they have been processed as above, which is ENABLED by +default (but of course only takes effect if any other exception handling is +activated) unless overridden by pre-defining the corresponding preprocessor +symbol to 0: + + - 'HALF_ERRHANDLING_OVERFLOW_TO_INEXACT' will cause overflow errors to also + raise a 'FE_INEXACT' exception. + - 'HALF_ERRHANDLING_UNDERFLOW_TO_INEXACT' will cause underflow errors to also + raise a 'FE_INEXACT' exception. This will also slightly change the + behaviour of the underflow exception, which will ONLY be raised if the + result is actually inexact due to underflow. If this is disabled, underflow + exceptions will be raised for ANY (possibly exact) subnormal result. + + +CREDITS AND CONTACT +------------------- + +This library is developed by CHRISTIAN RAU and released under the MIT License +(see LICENSE.txt). If you have any questions or problems with it, feel free to +contact me at rauy@users.sourceforge.net. + +Additional credit goes to JEROEN VAN DER ZIJP for his paper on "Fast Half Float +Conversions", whose algorithms have been used in the library for converting +between half-precision and single-precision values. diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 6f9db5be..7f7f844b 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -2,12 +2,40 @@ #include "scene.glsl" +layout(local_size_x = 16, local_size_y = 16) in; + // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl layout (set = 1, binding = 0) uniform sampler2D prefilteredDepth; layout (set = 1, binding = 1) uniform sampler2D normalBuffer; layout (r8, set = 1, binding = 2) uniform image2D aoOutput; +layout (push_constant) uniform PushConstants { + vec2 ndcToViewMult; + vec2 ndcToViewAdd; + + float depthLinearizeMult; + float depthLinearizeAdd; + + float radius; + float falloff; + float radiusMultiplier; + + float strength; + + int numDirections; + int numSteps; + + float temporalWeight; + float spatialFilterRadius; +} pushConstants; + void main() { + const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + if (screenPos.x % 2 == 0 && screenPos.y % 2 == 0){ + imageStore(aoOutput, screenPos, vec4(1.0f)); + } else { + imageStore(aoOutput, screenPos, vec4(0.5f)); + } } diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index cf81f709..a7abdf74 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -909,6 +909,24 @@ void ImguiWrapper::imguiInterface(Engine* engine) fmt::print(" Failed to find/create image save path directory"); } } + + if (ImGui::Button("Save raw ao image")) { + if (file::getOrCreateDirectory(file::imagesSavePath)) { + const std::filesystem::path path = file::imagesSavePath / "raw_ao_image.png"; + + vk_helpers::saveImageR8UNORM( + *engine->resourceManager, + *engine->immediate, + engine->ambientOcclusionPipeline->ambientOcclusionImage, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + path.string().c_str(), + 0 + ); + } + else { + fmt::print(" Failed to find/create image save path directory"); + } + } } ImGui::End(); diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 7ee4589e..ca68f6e9 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -105,6 +105,8 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth VkImageUsageFlags usage{}; usage |= VK_IMAGE_USAGE_STORAGE_BIT; usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); ambientOcclusionImage = resourceManager.createImage(imgInfo); @@ -360,12 +362,10 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V bindingInfos[1] = depthPrefilterDescriptorBuffer.getDescriptorBufferBindingInfo(); vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - uint32_t index0 = 0; - uint32_t index1 = 1; - VkDeviceSize offset0 = drawInfo.sceneDataOffset; + constexpr std::array indices{0,1}; + const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 1, &index0, &offset0); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 1, 1, &index1, &ZERO_DEVICE_SIZE); + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 8.0f)); auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 8.0f)); @@ -376,31 +376,29 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V } vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - // vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - // // Ambient Occlusion - // { - // vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); - // vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &drawInfo.pushConstants); - // - // VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; - // bindingInfos[0] = drawInfo.sceneDataBinding; - // bindingInfos[1] = ambientOcclusionDescriptorBuffer.getDescriptorBufferBindingInfo(); - // vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - // - // constexpr VkDeviceSize zeroOffset{0}; - // constexpr uint32_t sceneDataIndex{0}; - // constexpr uint32_t descriptorIndex{1}; - // - // vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 0, 1, &sceneDataIndex, &drawInfo.sceneDataOffset); - // vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipelineLayout, 1, 1, &descriptorIndex, &zeroOffset); - // - // const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); - // const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); - // vkCmdDispatch(cmd, x, y, 1); - // } - // - // - // vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + // Ambient Occlusion + { + vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); + vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); + + VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; + bindingInfos[0] = drawInfo.sceneDataBinding; + bindingInfos[1] = ambientOcclusionDescriptorBuffer.getDescriptorBufferBindingInfo(); + vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); + + constexpr std::array indices{0,1}; + const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; + + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); + + const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + vkCmdDispatch(cmd, x, y, 1); + } + + + vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vkCmdEndDebugUtilsLabelEXT(cmd); diff --git a/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp b/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp index 9edfd8b9..dc6f3c96 100644 --- a/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp +++ b/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp @@ -4,12 +4,16 @@ #include "deferred_resolve.h" +#include + #include "volk/volk.h" #include "src/renderer/renderer_constants.h" #include "src/renderer/resource_manager.h" -will_engine::deferred_resolve::DeferredResolvePipeline::DeferredResolvePipeline(ResourceManager& resourceManager, VkDescriptorSetLayout environmentIBLLayout, - VkDescriptorSetLayout cascadeUniformLayout, VkDescriptorSetLayout cascadeSamplerLayout) +will_engine::deferred_resolve::DeferredResolvePipeline::DeferredResolvePipeline(ResourceManager& resourceManager, + VkDescriptorSetLayout environmentIBLLayout, + VkDescriptorSetLayout cascadeUniformLayout, + VkDescriptorSetLayout cascadeSamplerLayout) : resourceManager(resourceManager) { VkPushConstantRange pushConstants = {}; @@ -123,23 +127,23 @@ void will_engine::deferred_resolve::DeferredResolvePipeline::draw(VkCommandBuffe bindingInfos[4] = drawInfo.cascadeSamplerBinding; vkCmdBindDescriptorBuffersEXT(cmd, 5, bindingInfos); - constexpr VkDeviceSize zeroOffset{0}; - constexpr uint32_t sceneDataIndex{0}; - constexpr uint32_t renderTargetsIndex{1}; - constexpr uint32_t environmentIndex{2}; - constexpr uint32_t cascadedShadowMapUniformIndex{3}; - constexpr uint32_t cascadedShadowMapSamplerIndex{4}; - - const VkDeviceSize sceneDataOffset{drawInfo.sceneDataOffset}; - const VkDeviceSize environmentOffset{drawInfo.environmentIBLOffset}; - const VkDeviceSize cascadedShadowMapUniformOffset{drawInfo.cascadeUniformOffset}; - constexpr VkDeviceSize cascadedShadowMapSamplerOffset{0}; - - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 0, 1, &sceneDataIndex, &sceneDataOffset); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 1, 1, &renderTargetsIndex, &zeroOffset); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 2, 1, &environmentIndex, &environmentOffset); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 3, 1, &cascadedShadowMapUniformIndex, &cascadedShadowMapUniformOffset); - vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 4, 1, &cascadedShadowMapSamplerIndex, &cascadedShadowMapSamplerOffset); + constexpr std::array indices{ + 0, + 1, + 2, + 3, + 4 + }; + const std::array offsets{ + drawInfo.sceneDataOffset, + ZERO_DEVICE_SIZE, + drawInfo.environmentIBLOffset, + drawInfo.cascadeUniformOffset, + ZERO_DEVICE_SIZE + }; + + + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 0, 5, indices.data(), offsets.data()); const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); diff --git a/src/renderer/vk_helpers.cpp b/src/renderer/vk_helpers.cpp index 834dfb50..f593018c 100644 --- a/src/renderer/vk_helpers.cpp +++ b/src/renderer/vk_helpers.cpp @@ -6,6 +6,7 @@ #include #include +#include #include #include #include "volk/volk.h" @@ -825,6 +826,52 @@ void will_engine::vk_helpers::saveImageR16F(const ResourceManager& resourceManag resourceManager.destroyBuffer(receivingBuffer); } +void will_engine::vk_helpers::saveImageR8UNORM(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, + const AllocatedImage& image, VkImageLayout imageLayout, const char* savePath, int32_t mipLevel) +{ + const size_t width = image.imageExtent.width / static_cast(std::pow(2, mipLevel)); + const size_t height = image.imageExtent.height / static_cast(std::pow(2, mipLevel)); + const size_t texelCount = width * height; + const size_t dataSize = texelCount * 1 * sizeof(uint8_t); + + AllocatedBuffer receivingBuffer = resourceManager.createReceivingBuffer(dataSize); + + immediate.submit([&, mipLevel](VkCommandBuffer cmd) { + VkBufferImageCopy bufferCopyRegion{}; + bufferCopyRegion.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; + bufferCopyRegion.imageSubresource.mipLevel = mipLevel; + bufferCopyRegion.imageSubresource.baseArrayLayer = 0; + bufferCopyRegion.imageSubresource.layerCount = 1; + bufferCopyRegion.imageExtent = {static_cast(width), static_cast(height), 1u}; + bufferCopyRegion.bufferOffset = 0; + bufferCopyRegion.bufferRowLength = 0; + bufferCopyRegion.bufferImageHeight = 0; + + transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + + vkCmdCopyImageToBuffer(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, receivingBuffer.buffer, 1, &bufferCopyRegion); + + transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, VK_IMAGE_ASPECT_COLOR_BIT); + }); + + void* data = receivingBuffer.info.pMappedData; + const auto imageData = static_cast(data); + + const auto byteImageData = new uint8_t[texelCount * 4]; + for (size_t i = 0; i < texelCount; ++i) { + const auto value = imageData[i]; + byteImageData[i * 4 + 0] = value; + byteImageData[i * 4 + 1] = value; + byteImageData[i * 4 + 2] = value; + byteImageData[i * 4 + 3] = 255; + } + + stbi_write_png(savePath, static_cast(width), static_cast(height), 4, byteImageData, static_cast(width) * 4); + + delete[] byteImageData; + resourceManager.destroyBuffer(receivingBuffer); +} + void will_engine::vk_helpers::saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha) { const auto byteImageData = new uint8_t[width * height * 4]; diff --git a/src/renderer/vk_helpers.h b/src/renderer/vk_helpers.h index ad8c6455..819116e3 100644 --- a/src/renderer/vk_helpers.h +++ b/src/renderer/vk_helpers.h @@ -126,6 +126,9 @@ namespace vk_helpers VkImageLayout imageLayout, VkImageAspectFlags aspectFlag, const char* savePath, const std::function& valueTransform, int32_t mipLevel = 0); + void saveImageR8UNORM(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, const char* savePath, int32_t mipLevel = 0); + void saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha = true); void saveHeightmap(const std::vector& heightData, int width, int height, const std::filesystem::path& filename); From a1e1814b5ca232de04c9e2d6de988b0492b76a60 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Fri, 28 Mar 2025 23:26:45 +0700 Subject: [PATCH 16/27] GTAO main pass scaffolding --- .../ground_truth/gtao_main_pass.comp | 41 ++++++++++++++++--- 1 file changed, 36 insertions(+), 5 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 7f7f844b..4c0f0869 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -30,12 +30,43 @@ layout (push_constant) uniform PushConstants { float spatialFilterRadius; } pushConstants; +vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rightZ, const float topZ, const float bottomZ) +{ + vec4 edgesLRTB = vec4(leftZ, rightZ, topZ, bottomZ) - vec4(centerZ); + + float slopeLR = (edgesLRTB.y - edgesLRTB.x) * 0.5; + float slopeTB = (edgesLRTB.w - edgesLRTB.z) * 0.5; + vec4 edgesLRTBSlopeAdjusted = edgesLRTB + vec4(slopeLR, -slopeLR, slopeTB, -slopeTB); + edgesLRTB = min(abs(edgesLRTB), abs(edgesLRTBSlopeAdjusted)); + + return clamp((1.25 - edgesLRTB / (centerZ * 0.011)), 0, 1);; +} + void main() { + // todo: a debug output image that will be used to visualize every step in a separate output image shared by every stage! const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; + + float viewSpaceZM = textureLod(prefilteredDepth, uv + vec2(0.0, 0.0) * sceneData.texelSize, 0).r; + float viewSpaceZL = textureLod(prefilteredDepth, uv + vec2(-1.0, 0.0) * sceneData.texelSize, 0).r; + float viewSpaceZR = textureLod(prefilteredDepth, uv + vec2(1.0, 0.0) * sceneData.texelSize, 0).r; + float viewSpaceZT = textureLod(prefilteredDepth, uv + vec2(0.0, 1.0) * sceneData.texelSize, 0).r; + float viewSpaceZB = textureLod(prefilteredDepth, uv + vec2(0.0, -1.0) * sceneData.texelSize, 0).r; + + vec4 edges = calculateDepthEdges(viewSpaceZM, viewSpaceZL, viewSpaceZR, viewSpaceZT, viewSpaceZB); + float minEdge = min(min(edges.x, edges.y), min(edges.z, edges.w)); + imageStore(aoOutput, screenPos, vec4(minEdge)); + + // Get view space normal by sampling normal buffer and converting from world to view (code not relevant) + //vec3 viewspaceNormal = (lpfloat3)XeGTAO_CalculateNormal(edgesLRTB, CENTER, LEFT, RIGHT, TOP, BOTTOM); + + // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used + viewspaceZM = viewspaceZM * 0.99920f; + + // Get the viewspace fragment position of the center pixel (i.e. reconstruct from depth, but in view space instead of world space) + // const vec3 pixCenterPos = XeGTAO_ComputeViewspacePosition( normalizedScreenPos, viewspaceZ, consts ); + // const vec3 viewVec = normalize(-pixCenterPos); - if (screenPos.x % 2 == 0 && screenPos.y % 2 == 0){ - imageStore(aoOutput, screenPos, vec4(1.0f)); - } else { - imageStore(aoOutput, screenPos, vec4(0.5f)); - } + // prevents normals that are facing away from the view vector - xeGTAO struggles with extreme cases, but in Vanilla it seems rare so it's disabled by default + // viewspaceNormal = normalize( viewspaceNormal + max( 0, -dot( viewspaceNormal, viewVec ) ) * viewVec ); } From b8aeb008474433b57d2dfff01dd7707081b1c9f5 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sat, 29 Mar 2025 17:09:52 +0700 Subject: [PATCH 17/27] View Space normals, position (view vec), depth. Better GTAO debugging. --- .../ground_truth/gtao_depth_prefilter.comp | 36 +++++---- .../ground_truth/gtao_main_pass.comp | 66 +++++++++++----- shaders/deferredResolve.comp | 11 +-- src/core/engine.cpp | 1 + src/renderer/imgui_wrapper.cpp | 56 ++++++-------- src/renderer/imgui_wrapper.h | 4 +- .../ambient_occlusion_types.h | 30 ++++---- .../ground_truth_ambient_occlusion.cpp | 77 +++++++++++++++---- .../ground_truth_ambient_occlusion.h | 4 + src/renderer/vk_helpers.cpp | 47 +++++++++++ src/renderer/vk_helpers.h | 3 + 11 files changed, 228 insertions(+), 107 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 240227f0..3ceb6ae3 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -15,25 +15,25 @@ layout (r16f, set = 1, binding = 2) uniform image2D outDepth1; layout (r16f, set = 1, binding = 3) uniform image2D outDepth2; layout (r16f, set = 1, binding = 4) uniform image2D outDepth3; layout (r16f, set = 1, binding = 5) uniform image2D outDepth4; +layout (rgba8, set = 1, binding = 6) uniform image2D debugImage; layout (push_constant) uniform PushConstants { - vec2 ndcToViewMult; - vec2 ndcToViewAdd; - float depthLinearizeMult; float depthLinearizeAdd; - float radius; - float falloff; - float radiusMultiplier; - - float strength; + float projectionParamX; + float projectionParamY; - int numDirections; - int numSteps; + float effectRadius; + float effectFalloffRange; + float denoiseBlurBeta; - float temporalWeight; - float spatialFilterRadius; + float radiusMultiplier; + float sampleDistributionPower; + float thinOccluderCompensation; + float finalValuePower; + float depthMipSamplingOffset; + float noiseIndex; } pushConstants; float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float depthLinearizeAdd) { @@ -78,6 +78,10 @@ void main() { const uvec2 baseCoord = gl_GlobalInvocationID.xy; const ivec2 screenPos = ivec2(baseCoord.xy) * 2;// We process 2x2 pixels in MIP 0 + if (screenPos.x > sceneData.renderTargetSize.x || screenPos.y > sceneData.renderTargetSize.y) { + return; + } + vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; float rDepth0 = texture(depthImage, uv + vec2(0.0, 0.0) * sceneData.texelSize).r; @@ -97,7 +101,7 @@ void main() { // MIP 1 - float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.effectRadius, pushConstants.radiusMultiplier, pushConstants.effectFalloffRange); imageStore(outDepth1, ivec2(baseCoord), vec4(dm1)); g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm1; @@ -111,7 +115,7 @@ void main() { float inBL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+1]; float inBR = g_scratchDepths[groupThreadID.x+1][groupThreadID.y+1]; - float dm2 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + float dm2 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.effectRadius, pushConstants.radiusMultiplier, pushConstants.effectFalloffRange); imageStore(outDepth2, ivec2(baseCoord/2u), vec4(dm2)); g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm2; } @@ -126,7 +130,7 @@ void main() { float inBL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+2]; float inBR = g_scratchDepths[groupThreadID.x+2][groupThreadID.y+2]; - float dm3 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + float dm3 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.effectRadius, pushConstants.radiusMultiplier, pushConstants.effectFalloffRange); imageStore(outDepth3, ivec2(baseCoord/4u), vec4(dm3)); g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm3; } @@ -141,7 +145,7 @@ void main() { float inBL = g_scratchDepths[groupThreadID.x+0][groupThreadID.y+4]; float inBR = g_scratchDepths[groupThreadID.x+4][groupThreadID.y+4]; - float dm4 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.radius, pushConstants.radiusMultiplier, pushConstants.falloff); + float dm4 = depthMipFilter(inTL, inTR, inBL, inBR, pushConstants.effectRadius, pushConstants.radiusMultiplier, pushConstants.effectFalloffRange); imageStore(outDepth4, ivec2(baseCoord/8u), vec4(dm4)); //g_scratchDepths[groupThreadID.x][groupThreadID.y] = dm4; } diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 4c0f0869..27542611 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -9,25 +9,25 @@ layout(local_size_x = 16, local_size_y = 16) in; layout (set = 1, binding = 0) uniform sampler2D prefilteredDepth; layout (set = 1, binding = 1) uniform sampler2D normalBuffer; layout (r8, set = 1, binding = 2) uniform image2D aoOutput; +layout (rgba8, set = 1, binding = 3) uniform image2D debugImage; layout (push_constant) uniform PushConstants { - vec2 ndcToViewMult; - vec2 ndcToViewAdd; - float depthLinearizeMult; float depthLinearizeAdd; - float radius; - float falloff; - float radiusMultiplier; - - float strength; + float projectionParamX; + float projectionParamY; - int numDirections; - int numSteps; + float effectRadius; + float effectFalloffRange; + float denoiseBlurBeta; - float temporalWeight; - float spatialFilterRadius; + float radiusMultiplier; + float sampleDistributionPower; + float thinOccluderCompensation; + float finalValuePower; + float depthMipSamplingOffset; + float noiseIndex; } pushConstants; vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rightZ, const float topZ, const float bottomZ) @@ -42,9 +42,19 @@ vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rig return clamp((1.25 - edgesLRTB / (centerZ * 0.011)), 0, 1);; } +vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth, float xProj, float yProj) { + vec2 ndc = uv * 2.0 - 1.0; + vec2 projectionParams = vec2(1.0 / xProj, 1.0 / yProj); + return vec3(ndc * viewDepth * projectionParams, -viewDepth); +} + void main() { - // todo: a debug output image that will be used to visualize every step in a separate output image shared by every stage! const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + + if (screenPos.x > sceneData.renderTargetSize.x || screenPos.y > sceneData.renderTargetSize.y) { + return; + } + vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; float viewSpaceZM = textureLod(prefilteredDepth, uv + vec2(0.0, 0.0) * sceneData.texelSize, 0).r; @@ -55,18 +65,36 @@ void main() { vec4 edges = calculateDepthEdges(viewSpaceZM, viewSpaceZL, viewSpaceZR, viewSpaceZT, viewSpaceZB); float minEdge = min(min(edges.x, edges.y), min(edges.z, edges.w)); - imageStore(aoOutput, screenPos, vec4(minEdge)); + // imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); // Get view space normal by sampling normal buffer and converting from world to view (code not relevant) - //vec3 viewspaceNormal = (lpfloat3)XeGTAO_CalculateNormal(edgesLRTB, CENTER, LEFT, RIGHT, TOP, BOTTOM); + vec3 worldNormal = texture(normalBuffer, uv).rgb; + vec3 viewNormal = mat3(sceneData.view) * worldNormal; + // imageStore(debugImage, screenPos, vec4(viewNormal * 0.5f + 0.5f, 1.0f)); // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used - viewspaceZM = viewspaceZM * 0.99920f; + viewSpaceZM = viewSpaceZM * 0.99920f; + - // Get the viewspace fragment position of the center pixel (i.e. reconstruct from depth, but in view space instead of world space) - // const vec3 pixCenterPos = XeGTAO_ComputeViewspacePosition( normalizedScreenPos, viewspaceZ, consts ); - // const vec3 viewVec = normalize(-pixCenterPos); + vec3 vPos = reconstructViewSpacePosition(uv, viewSpaceZM, pushConstants.projectionParamX, pushConstants.projectionParamY); + vec3 viewVec = normalize(-vPos); + imageStore(debugImage, screenPos, vec4(viewVec, 1.0f)); + // Per Intel // prevents normals that are facing away from the view vector - xeGTAO struggles with extreme cases, but in Vanilla it seems rare so it's disabled by default // viewspaceNormal = normalize( viewspaceNormal + max( 0, -dot( viewspaceNormal, viewVec ) ) * viewVec ); + + + const float effectRadius = pushConstants.effectRadius * pushConstants.radiusMultiplier; + const float sampleDistributionPower = pushConstants.sampleDistributionPower; + const float thinOccluderCompensation = pushConstants.thinOccluderCompensation; + const float falloffRange = pushConstants.effectFalloffRange * effectRadius; + + const float falloffFrom = effectRadius * (1 - pushConstants.effectFalloffRange); + + // fadeout precompute optimisation + const float falloffMul = 1.0 / (falloffRange); + const float falloffAdd = falloffFrom / (falloffRange) + 1.0; + + vec3 bentNormal = viewNormal; } diff --git a/shaders/deferredResolve.comp b/shaders/deferredResolve.comp index a6de0812..3d7f2f3d 100644 --- a/shaders/deferredResolve.comp +++ b/shaders/deferredResolve.comp @@ -50,12 +50,12 @@ layout (push_constant) uniform PushConstants { float farPlane; } pushConstants; -vec3 reconstructPosition(ivec2 texCoord, vec2 texelSize, float depth) { +vec3 reconstructPosition(vec2 uv, float ndcDepth) { // Get normalized device coordinates - vec2 ndc = (vec2(texCoord) + 0.5) * texelSize * 2.0 - 1.0; + vec2 ndc = uv * 2.0 - 1.0; // Reconstruct view-space position - vec4 positionVS = sceneData.invProjection * vec4(ndc, depth, 1.0); + vec4 positionVS = sceneData.invProjection * vec4(ndc, ndcDepth, 1.0); positionVS /= positionVS.w; // Transform to world-space @@ -87,7 +87,7 @@ void main() { float depth = texture(depthBuffer, uv).r; vec3 normal = texture(normalRenderTarget, uv).rgb; vec4 pbrData = texture(pbrRenderTarget, uv); - vec3 position = reconstructPosition(screenPos, sceneData.texelSize, depth); + vec3 position = reconstructPosition(uv, depth); float roughness = pbrData.g; float metallic = pbrData.r; @@ -186,7 +186,8 @@ void main() { imageStore(outputImage, screenPos, vec4(vec3(dot(N, L)), 1.0)); break; case 9: - // imageStore(outputImage, screenPos, vec4(vec3(ao), 1.0f)); + vec4 viewPos = sceneData.view * vec4(position, 1.0f); + imageStore(outputImage, screenPos, vec4(-viewPos.xyz / 100.0f, 1.0f)); break; } } \ No newline at end of file diff --git a/src/core/engine.cpp b/src/core/engine.cpp index fc8ca46b..2299ad81 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -568,6 +568,7 @@ void Engine::draw(float deltaTime) ambient_occlusion::GTAODrawInfo gtaoDrawInfo{ camera, {}, + frameNumber, sceneDataDescriptorBuffer.getDescriptorBufferBindingInfo(), sceneDataDescriptorBuffer.getDescriptorBufferSize() * currentFrameOverlap }; diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index a7abdf74..1a5e17ae 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -207,7 +207,7 @@ void ImguiWrapper::imguiInterface(Engine* engine) if (ImGui::BeginTabItem("Pipelines")) { ImGui::Text("Deferred Debug"); - const char* deferredDebugOptions[]{"None", "Depth", "Velocity", "Albedo", "Normal", "PBR", "Shadows", "Cascade Level", "nDotL"}; + const char* deferredDebugOptions[]{"None", "Depth", "Velocity", "Albedo", "Normal", "PBR", "Shadows", "Cascade Level", "nDotL", "AO"}; ImGui::Combo("Deferred Debug", &engine->deferredDebug, deferredDebugOptions, IM_ARRAYSIZE(deferredDebugOptions)); ImGui::EndTabItem(); @@ -881,51 +881,43 @@ void ImguiWrapper::imguiInterface(Engine* engine) constexpr uint32_t minMip = 0; constexpr uint32_t maxMip = 4; - ImGui::SliderScalar("GTAO level", ImGuiDataType_S32, >aoMip, &minMip, &maxMip); + if (aoDebugTextureImguiId == VK_NULL_HANDLE) { + if (engine->ambientOcclusionPipeline->debugImage.image != VK_NULL_HANDLE) { + aoDebugTextureImguiId = ImGui_ImplVulkan_AddTexture(engine->resourceManager->getDefaultSamplerLinear(), + engine->ambientOcclusionPipeline->debugImage.imageView, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); + } + } - if (ImGui::Button("Save GTAO depth image")) { - if (file::getOrCreateDirectory(file::imagesSavePath)) { - const std::filesystem::path path = file::imagesSavePath / "gtao_depth.png"; - auto depthNormalize = [](const uint16_t depth) { - // Equivalent - float manualDepth = math::halfToFloat(depth); - float libraryDepth = half_float::detail::half2float(depth); - return libraryDepth / 1000.f; - }; + if (aoDebugTextureImguiId == VK_NULL_HANDLE) { + ImGui::Text("Issue."); + } + else { + float maxSize = ImGui::GetContentRegionAvail().x; + maxSize = glm::min(maxSize, 1024.0f); - vk_helpers::saveImageR16F( - *engine->resourceManager, - *engine->immediate, - engine->ambientOcclusionPipeline->depthPrefilterImage, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, - VK_IMAGE_ASPECT_COLOR_BIT, - path.string().c_str(), - depthNormalize, - gtaoMip - ); - } - else { - fmt::print(" Failed to find/create image save path directory"); - } + VkExtent3D imageExtent = engine->ambientOcclusionPipeline->debugImage.imageExtent; + float width = std::min(maxSize, static_cast(imageExtent.width)); + float aspectRatio = static_cast(imageExtent.width) / static_cast(imageExtent.height); + float height = width / aspectRatio; + + ImGui::Image(reinterpret_cast(aoDebugTextureImguiId), ImVec2(width, height)); } - if (ImGui::Button("Save raw ao image")) { + if (ImGui::Button("Save GTAO Debug Image")) { if (file::getOrCreateDirectory(file::imagesSavePath)) { - const std::filesystem::path path = file::imagesSavePath / "raw_ao_image.png"; + const std::filesystem::path path = file::imagesSavePath / "gtao_debug.png"; - vk_helpers::saveImageR8UNORM( + vk_helpers::saveImageR8G8B8A8UNORM( *engine->resourceManager, *engine->immediate, - engine->ambientOcclusionPipeline->ambientOcclusionImage, + engine->ambientOcclusionPipeline->debugImage, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, path.string().c_str(), 0 ); } - else { - fmt::print(" Failed to find/create image save path directory"); - } } } ImGui::End(); diff --git a/src/renderer/imgui_wrapper.h b/src/renderer/imgui_wrapper.h index 17b6f844..c77f836d 100644 --- a/src/renderer/imgui_wrapper.h +++ b/src/renderer/imgui_wrapper.h @@ -80,7 +80,9 @@ class ImguiWrapper terrain::TerrainProperties terrainProperties{}; std::array terrainTextures; - VkDescriptorSet currentlySelectedTextureImguiId = VK_NULL_HANDLE; + VkDescriptorSet currentlySelectedTextureImguiId{VK_NULL_HANDLE}; + + VkDescriptorSet aoDebugTextureImguiId{VK_NULL_HANDLE}; }; } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 55338e44..30cb29ef 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -11,37 +11,33 @@ namespace will_engine::ambient_occlusion { static constexpr int32_t DEPTH_PREFILTER_MIP_COUNT = 5; +static constexpr int32_t GTAO_DENOISE_PASSES = 1; struct GTAOPushConstants { - glm::vec2 ndcToViewMult; - glm::vec2 ndcToViewAdd; - - // Depth prefilter parameters float depthLinearizeMult; float depthLinearizeAdd; - // Defaults follow Intel's implementation - float radius = 0.5f; - float falloff = 0.615f; - float radiusMultiplier = 1.457f; - - // AO parameters - float strength; + float projectionParamX; + float projectionParamY; - // Sampling parameters - uint32_t numDirections; - uint32_t numSteps; + float effectRadius = 0.5f; + float effectFalloffRange = 0.615f; + float denoiseBlurBeta = (GTAO_DENOISE_PASSES == 0) ? (1e4f) : (1.2f); - // Temporal/filter parameters - float temporalWeight; - float spatialFilterRadius; + float radiusMultiplier = 1.457f; + float sampleDistributionPower = 2.0f; + float thinOccluderCompensation = 0.0f; + float finalValuePower = 2.2f; + float depthMipSamplingOffset = 3.30f; + float noiseIndex; }; struct GTAODrawInfo { Camera* camera{nullptr}; GTAOPushConstants pushConstants{}; + int32_t currentFrame{}; VkDescriptorBufferBindingInfoEXT sceneDataBinding{}; VkDeviceSize sceneDataOffset{0}; }; diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index ca68f6e9..48ddb4db 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -10,8 +10,21 @@ #include "src/renderer/vk_descriptors.h" #include "src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h" -will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruthAmbientOcclusionPipeline(ResourceManager& resourceManager) : resourceManager(resourceManager) +will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruthAmbientOcclusionPipeline( + ResourceManager& resourceManager) : resourceManager(resourceManager) { + // Debug + { + VkImageUsageFlags usage{}; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; + usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; + + VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(debugFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + debugImage = resourceManager.createImage(imgInfo); + } + // Depth Pre-filtering { DescriptorLayoutBuilder layoutBuilder; @@ -21,8 +34,10 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 2 layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 3 layoutBuilder.addBinding(5, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao depth mip 4 + layoutBuilder.addBinding(6, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // debug image - depthPrefilterSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + depthPrefilterSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, + VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); VkPushConstantRange pushConstants{}; pushConstants.offset = 0; @@ -77,8 +92,10 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // pre-filtered depth layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT normal buffer layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao output + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // debug image - ambientOcclusionSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + ambientOcclusionSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, + VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); VkPushConstantRange pushConstants{}; pushConstants.offset = 0; @@ -156,7 +173,8 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT normal buffer layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // filtered ao - spatialFilteringSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + spatialFilteringSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, + VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); VkPushConstantRange pushConstants{}; pushConstants.offset = 0; @@ -197,7 +215,8 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT depth buffer layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // final output - temporalAccumulationSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); + temporalAccumulationSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, + VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); VkPushConstantRange pushConstants{}; pushConstants.offset = 0; @@ -239,6 +258,9 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTruthAmbientOcclusionPipeline() { + // Debug Resources + resourceManager.destroyImage(debugImage); + // Depth Prefilter Resources resourceManager.destroyDescriptorSetLayout(depthPrefilterSetLayout); resourceManager.destroyPipelineLayout(depthPrefilterPipelineLayout); @@ -288,7 +310,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTrut void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(const VkImageView& depthImageView) { std::vector imageDescriptors{}; - imageDescriptors.reserve(1 + DEPTH_PREFILTER_MIP_COUNT); + imageDescriptors.reserve(1 + DEPTH_PREFILTER_MIP_COUNT + 1); imageDescriptors.push_back( { @@ -307,13 +329,19 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupD imageDescriptors.push_back(imageData); } + imageDescriptors.push_back({ + VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, + {VK_NULL_HANDLE, debugImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + false + }); + resourceManager.setupDescriptorBufferSampler(depthPrefilterDescriptorBuffer, imageDescriptors, 0); } void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView) { std::vector imageDescriptors{}; - imageDescriptors.reserve(2); + imageDescriptors.reserve(4); imageDescriptors.push_back( { @@ -333,6 +361,11 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupA {VK_NULL_HANDLE, ambientOcclusionImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, false }); + imageDescriptors.push_back({ + VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, + {VK_NULL_HANDLE, debugImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + false + }); resourceManager.setupDescriptorBufferSampler(ambientOcclusionDescriptorBuffer, imageDescriptors, 0); } @@ -349,8 +382,13 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V push.depthLinearizeMult = -projMatrix[3][2]; push.depthLinearizeAdd = projMatrix[2][2]; + push.projectionParamX = projMatrix[0][0]; + push.projectionParamY = projMatrix[1][1]; + + push.noiseIndex = GTAO_DENOISE_PASSES > 0 ? drawInfo.currentFrame % 64 : 0; + + vk_helpers::transitionImage(cmd, debugImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - //vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::clearColorImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL); // Depth Prefilter { @@ -362,20 +400,21 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V bindingInfos[1] = depthPrefilterDescriptorBuffer.getDescriptorBufferBindingInfo(); vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - constexpr std::array indices{0,1}; + constexpr std::array indices{0, 1}; const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); - auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 8.0f)); - auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 8.0f)); - // divided by 2 because depth prepass operates on 2x2 (still input4 -> output4) - x = x / 2; - y = y / 2; + auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + // shader only operates on 8,8 work groups, mip 0 will operate on 2x2 texels + // x = x / 2 + 1; + // y = y / 2 + 1; vkCmdDispatch(cmd, x, y, 1); } - vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, depthPrefilterImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); // Ambient Occlusion { @@ -387,7 +426,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V bindingInfos[1] = ambientOcclusionDescriptorBuffer.getDescriptorBufferBindingInfo(); vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); - constexpr std::array indices{0,1}; + constexpr std::array indices{0, 1}; const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipelineLayout, 0, 2, indices.data(), offsets.data()); @@ -398,8 +437,12 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V } - vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + VK_IMAGE_ASPECT_COLOR_BIT); + + vk_helpers::transitionImage(cmd, debugImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + VK_IMAGE_ASPECT_COLOR_BIT); vkCmdEndDebugUtilsLabelEXT(cmd); } diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index 60ae2cbb..c243313f 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -83,6 +83,10 @@ class GroundTruthAmbientOcclusionPipeline private: // Output AllocatedImage ambientOcclusionOutputImage{VK_NULL_HANDLE}; +private: // Debug + VkFormat debugFormat{VK_FORMAT_R8G8B8A8_UNORM}; + AllocatedImage debugImage{VK_NULL_HANDLE}; + private: ResourceManager& resourceManager; diff --git a/src/renderer/vk_helpers.cpp b/src/renderer/vk_helpers.cpp index f593018c..f88ed89c 100644 --- a/src/renderer/vk_helpers.cpp +++ b/src/renderer/vk_helpers.cpp @@ -872,6 +872,53 @@ void will_engine::vk_helpers::saveImageR8UNORM(const ResourceManager& resourceMa resourceManager.destroyBuffer(receivingBuffer); } +void will_engine::vk_helpers::saveImageR8G8B8A8UNORM(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, + const AllocatedImage& image, VkImageLayout imageLayout, const char* savePath, int32_t mipLevel) +{ + const size_t width = image.imageExtent.width / static_cast(std::pow(2, mipLevel)); + const size_t height = image.imageExtent.height / static_cast(std::pow(2, mipLevel)); + const size_t texelCount = width * height; + const size_t dataSize = texelCount * 4 * sizeof(uint8_t); + + AllocatedBuffer receivingBuffer = resourceManager.createReceivingBuffer(dataSize); + + immediate.submit([&, mipLevel](VkCommandBuffer cmd) { + VkBufferImageCopy bufferCopyRegion{}; + bufferCopyRegion.imageSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT; + bufferCopyRegion.imageSubresource.mipLevel = mipLevel; + bufferCopyRegion.imageSubresource.baseArrayLayer = 0; + bufferCopyRegion.imageSubresource.layerCount = 1; + bufferCopyRegion.imageExtent = {static_cast(width), static_cast(height), 1u}; + bufferCopyRegion.bufferOffset = 0; + bufferCopyRegion.bufferRowLength = 0; + bufferCopyRegion.bufferImageHeight = 0; + + transitionImage(cmd, image.image, imageLayout, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + + vkCmdCopyImageToBuffer(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, receivingBuffer.buffer, 1, &bufferCopyRegion); + + transitionImage(cmd, image.image, VK_IMAGE_LAYOUT_GENERAL, imageLayout, VK_IMAGE_ASPECT_COLOR_BIT); + }); + + void* data = receivingBuffer.info.pMappedData; + const auto imageData = static_cast(data); + + const auto byteImageData = new uint8_t[texelCount * 4]; + for (size_t i = 0; i < texelCount; ++i) { + const auto value = imageData[i]; + byteImageData[i * 4 + 0] = imageData[i * 4 + 0]; + byteImageData[i * 4 + 1] = imageData[i * 4 + 1]; + byteImageData[i * 4 + 2] = imageData[i * 4 + 2]; + // overwrite alpha to 1.0f + byteImageData[i * 4 + 3] = 255; + } + + stbi_write_png(savePath, static_cast(width), static_cast(height), 4, byteImageData, static_cast(width) * 4); + + delete[] byteImageData; + resourceManager.destroyBuffer(receivingBuffer); +} + void will_engine::vk_helpers::saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha) { const auto byteImageData = new uint8_t[width * height * 4]; diff --git a/src/renderer/vk_helpers.h b/src/renderer/vk_helpers.h index 819116e3..47eedfde 100644 --- a/src/renderer/vk_helpers.h +++ b/src/renderer/vk_helpers.h @@ -129,6 +129,9 @@ namespace vk_helpers void saveImageR8UNORM(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, VkImageLayout imageLayout, const char* savePath, int32_t mipLevel = 0); + void saveImageR8G8B8A8UNORM(const ResourceManager& resourceManager, const ImmediateSubmitter& immediate, const AllocatedImage& image, + VkImageLayout imageLayout, const char* savePath, int32_t mipLevel = 0); + void saveImage(const std::vector& imageData, int width, int height, std::filesystem::path filename, bool overrideAlpha = true); void saveHeightmap(const std::vector& heightData, int width, int height, const std::filesystem::path& filename); From 582d4da1fe1b7a872238787bfddc76fc3813c02a Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sat, 29 Mar 2025 20:57:39 +0700 Subject: [PATCH 18/27] GTAO main pass progress. Early exit for distant pixels. --- .../ground_truth/gtao_depth_prefilter.comp | 12 +- .../ground_truth/gtao_main_pass.comp | 105 ++++++++++++++++-- shaders/deferredResolve.comp | 5 +- shaders/environment/environment.vert | 2 +- src/core/camera/free_camera.cpp | 9 +- src/core/engine.cpp | 2 +- .../ambient_occlusion_types.h | 12 +- .../ground_truth_ambient_occlusion.cpp | 15 ++- .../post_process/post_process_types.h | 4 +- 9 files changed, 137 insertions(+), 29 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 3ceb6ae3..ec9bb0bc 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -18,12 +18,16 @@ layout (r16f, set = 1, binding = 5) uniform image2D outDepth4; layout (rgba8, set = 1, binding = 6) uniform image2D debugImage; layout (push_constant) uniform PushConstants { + vec2 cameraTanHalfFOV; + + vec2 ndcToViewMul; + vec2 ndcToViewAdd; + + vec2 ndcToViewMul_x_PixelSize; + float depthLinearizeMult; float depthLinearizeAdd; - float projectionParamX; - float projectionParamY; - float effectRadius; float effectFalloffRange; float denoiseBlurBeta; @@ -33,7 +37,7 @@ layout (push_constant) uniform PushConstants { float thinOccluderCompensation; float finalValuePower; float depthMipSamplingOffset; - float noiseIndex; + uint noiseIndex; } pushConstants; float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float depthLinearizeAdd) { diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 27542611..2917bfd5 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -12,12 +12,16 @@ layout (r8, set = 1, binding = 2) uniform image2D aoOutput; layout (rgba8, set = 1, binding = 3) uniform image2D debugImage; layout (push_constant) uniform PushConstants { + vec2 cameraTanHalfFOV; + + vec2 ndcToViewMul; + vec2 ndcToViewAdd; + + vec2 ndcToViewMul_x_PixelSize; + float depthLinearizeMult; float depthLinearizeAdd; - float projectionParamX; - float projectionParamY; - float effectRadius; float effectFalloffRange; float denoiseBlurBeta; @@ -27,9 +31,44 @@ layout (push_constant) uniform PushConstants { float thinOccluderCompensation; float finalValuePower; float depthMipSamplingOffset; - float noiseIndex; + uint noiseIndex; } pushConstants; +#define XE_HILBERT_LEVEL 6u +#define XE_HILBERT_WIDTH (1u << XE_HILBERT_LEVEL) +#define XE_HILBERT_AREA (XE_HILBERT_WIDTH * XE_HILBERT_WIDTH) + +uint hilbertIndex(uint posX, uint posY) +{ + uint index = 0u; + for (uint curLevel = XE_HILBERT_WIDTH/2u; curLevel > 0u; curLevel /= 2u) + { + uint regionX = (posX & curLevel) > 0u ? 1u : 0u; + uint regionY = (posY & curLevel) > 0u ? 1u : 0u; + index += curLevel * curLevel * ((3u * regionX) ^ regionY); + if (regionY == 0u) + { + if (regionX == 1u) + { + posX = uint(XE_HILBERT_WIDTH - 1u) - posX; + posY = uint(XE_HILBERT_WIDTH - 1u) - posY; + } + + uint temp = posX; + posX = posY; + posY = temp; + } + } + return index; +} + +vec2 spatioTemporalNoise(ivec2 pixCoord, uint temporalIndex)// without TAA, temporalIndex is always 0 +{ + uint index = hilbertIndex(uint(pixCoord.x), uint(pixCoord.y)); + index += 288u * (temporalIndex % 64u); + return vec2(fract(0.5 + index * vec2(0.75487766624669276005, 0.5698402909980532659114))); +} + vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rightZ, const float topZ, const float bottomZ) { vec4 edgesLRTB = vec4(leftZ, rightZ, topZ, bottomZ) - vec4(centerZ); @@ -42,10 +81,29 @@ vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rig return clamp((1.25 - edgesLRTB / (centerZ * 0.011)), 0, 1);; } -vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth, float xProj, float yProj) { +vec3 cheapReconstructViewSpacePosition(vec2 uv, const float viewspaceDepth, vec2 ndcToViewMul, vec2 ndcToViewAdd) +{ + vec3 ret; + ret.xy = (ndcToViewMul * uv.xy + ndcToViewAdd) * viewspaceDepth; + ret.z = -viewspaceDepth; + return ret; +} + + +vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth, float depthLinearizeAdd, float depthLinearizeMult) { + float ndcDepth = pushConstants.depthLinearizeAdd - (pushConstants.depthLinearizeMult / viewDepth); + uv.y = 1 - uv.y; vec2 ndc = uv * 2.0 - 1.0; - vec2 projectionParams = vec2(1.0 / xProj, 1.0 / yProj); - return vec3(ndc * viewDepth * projectionParams, -viewDepth); + vec4 positionVS = sceneData.invProjection * vec4(ndc, ndcDepth, 1.0); + + positionVS /= positionVS.w; + return positionVS.xyz; +} + +void outputWorkingTerm(ivec2 screenPos, float visibility, vec3 bentNormal, image2D outputImage){ + const float XE_GTAO_OCCLUSION_TERM_SCALE = 1.5f; + visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); + imageStore(outputImage, screenPos, vec4(visibility + 0.5f / 255.0f)); } void main() { @@ -75,10 +133,14 @@ void main() { // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used viewSpaceZM = viewSpaceZM * 0.99920f; + vec3 vPos = cheapReconstructViewSpacePosition(uv, viewSpaceZM, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); - vec3 vPos = reconstructViewSpacePosition(uv, viewSpaceZM, pushConstants.projectionParamX, pushConstants.projectionParamY); vec3 viewVec = normalize(-vPos); - imageStore(debugImage, screenPos, vec4(viewVec, 1.0f)); + //imageStore(debugImage, screenPos, vec4(viewVec, 1.0f)); + + // debug world pos + vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0)).xyz; + imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); // Per Intel // prevents normals that are facing away from the view vector - xeGTAO struggles with extreme cases, but in Vanilla it seems rare so it's disabled by default @@ -96,5 +158,30 @@ void main() { const float falloffMul = 1.0 / (falloffRange); const float falloffAdd = falloffFrom / (falloffRange) + 1.0; + float visibility = 0; + // set bent normal to 0 if generating and outputting to buffer for use in deferred resolve vec3 bentNormal = viewNormal; + + // NOISE + vec2 noise = spatioTemporalNoise(screenPos, pushConstants.noiseIndex); + float noiseSlice = noise.x; + float noiseSample = noise.y; + + const float pixelTooCloseThreshold = 1.3; + const vec2 pixelDirRBViewspaceSizeAtCenterZ = viewSpaceZM.xx * pushConstants.ndcToViewMul_x_PixelSize; + + float screenspaceRadius = effectRadius / pixelDirRBViewspaceSizeAtCenterZ.x; + visibility += clamp((10 - screenspaceRadius)/100, 0, 1) * 0.5; + + if(screenspaceRadius < pixelTooCloseThreshold) + { + visibility = 1; + const float XE_GTAO_OCCLUSION_TERM_SCALE = 1.5f; + visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); + imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); + // if outputting bent normals, need to write `viewNormal's` value to the buffer (i.e. no change to trajectory of normal) + return; + } else { + imageStore(debugImage, screenPos, vec4(vec3(1.0f, 1.0f, 0.0f), 1.0f)); + } } diff --git a/shaders/deferredResolve.comp b/shaders/deferredResolve.comp index 3d7f2f3d..e04d7ee6 100644 --- a/shaders/deferredResolve.comp +++ b/shaders/deferredResolve.comp @@ -186,8 +186,9 @@ void main() { imageStore(outputImage, screenPos, vec4(vec3(dot(N, L)), 1.0)); break; case 9: - vec4 viewPos = sceneData.view * vec4(position, 1.0f); - imageStore(outputImage, screenPos, vec4(-viewPos.xyz / 100.0f, 1.0f)); + // vec4 viewPos = sceneData.view * vec4(position, 1.0f); + // imageStore(outputImage, screenPos, vec4(-viewPos.xyz / 100.0f, 1.0f)); + imageStore(outputImage, screenPos, vec4(position.xyz / 1000.0f, 1.0f)); break; } } \ No newline at end of file diff --git a/shaders/environment/environment.vert b/shaders/environment/environment.vert index 6ea6da69..8b8d9aec 100644 --- a/shaders/environment/environment.vert +++ b/shaders/environment/environment.vert @@ -9,7 +9,7 @@ layout (location = 2) out vec4 outPrevMvpPosition; // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl void main() { - const vec3 vertices[3] = vec3[3](vec3(-1, -1, 0.0002), vec3(3, -1, 0.0002), vec3(-1, 3, 0.0002)); + const vec3 vertices[3] = vec3[3](vec3(-1, -1, 0.000001), vec3(3, -1, 0.000001), vec3(-1, 3, 0.000001)); vec4 currClipPos = vec4(vertices[gl_VertexIndex], 1); diff --git a/src/core/camera/free_camera.cpp b/src/core/camera/free_camera.cpp index 8d84ac26..9f6c8c71 100644 --- a/src/core/camera/free_camera.cpp +++ b/src/core/camera/free_camera.cpp @@ -18,6 +18,7 @@ void will_engine::FreeCamera::update(const float deltaTime) } glm::vec3 velocity{0.f}; + float verticalVelocity{0.f}; if (input.isKeyDown(SDLK_D)) { velocity.x += 1.0f; @@ -26,10 +27,10 @@ void will_engine::FreeCamera::update(const float deltaTime) velocity.x -= 1.0f; } if (input.isKeyDown(SDLK_LCTRL)) { - velocity.y -= 1.0f; + verticalVelocity -= 1.0f; } if (input.isKeyDown(SDLK_SPACE)) { - velocity.y += 1.0f; + verticalVelocity += 1.0f; } // I guess vulkan is negative Z forward?! if (input.isKeyDown(SDLK_W)) { @@ -54,6 +55,7 @@ void will_engine::FreeCamera::update(const float deltaTime) const auto currentSpeed = static_cast(glm::pow(10, scale)); velocity *= deltaTime * currentSpeed; + verticalVelocity *= deltaTime * currentSpeed; const float yaw = glm::radians(-input.getMouseXDelta() / 10.0f); const float pitch = glm::radians(-input.getMouseYDelta() / 10.0f); @@ -74,7 +76,8 @@ void will_engine::FreeCamera::update(const float deltaTime) transform.setRotation(newRotation); const glm::mat4 rotationMatrix = getRotationMatrixWS(); - const auto finalVelocity = glm::vec3(rotationMatrix * glm::vec4(velocity, 0.f)); + auto finalVelocity = glm::vec3(rotationMatrix * glm::vec4(velocity, 0.f)); + finalVelocity += glm::vec3(0.0f, verticalVelocity, 0.0f); transform.translate(finalVelocity); updateViewMatrix(); diff --git a/src/core/engine.cpp b/src/core/engine.cpp index 2299ad81..a19eee4a 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -608,7 +608,7 @@ void Engine::draw(float deltaTime) vk_helpers::transitionImage(cmd, taaResolveTarget.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, postProcessOutputBuffer.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - postProcessPipeline->draw(cmd, post_process::PostProcessType::ALL); + postProcessPipeline->draw(cmd, post_process::PostProcessType::None); vk_helpers::transitionImage(cmd, postProcessOutputBuffer.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, swapchainImages[swapchainImageIndex], VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 30cb29ef..6985b044 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -15,12 +15,16 @@ static constexpr int32_t GTAO_DENOISE_PASSES = 1; struct GTAOPushConstants { + glm::vec2 cameraTanHalfFOV; + + glm::vec2 ndcToViewMul; + glm::vec2 ndcToViewAdd; + + glm::vec2 ndcToViewMul_x_PixelSize; + float depthLinearizeMult; float depthLinearizeAdd; - float projectionParamX; - float projectionParamY; - float effectRadius = 0.5f; float effectFalloffRange = 0.615f; float denoiseBlurBeta = (GTAO_DENOISE_PASSES == 0) ? (1e4f) : (1.2f); @@ -30,7 +34,7 @@ struct GTAOPushConstants float thinOccluderCompensation = 0.0f; float finalValuePower = 2.2f; float depthMipSamplingOffset = 3.30f; - float noiseIndex; + uint32_t noiseIndex; }; struct GTAODrawInfo diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 48ddb4db..e2dcc0dd 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -381,9 +381,18 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V glm::mat4 projMatrix = drawInfo.camera->getProjMatrix(); push.depthLinearizeMult = -projMatrix[3][2]; push.depthLinearizeAdd = projMatrix[2][2]; + if (push.depthLinearizeMult * push.depthLinearizeAdd < 0) { + push.depthLinearizeAdd = -push.depthLinearizeAdd; + } + + float tanHalfFOVY = 1.0f / projMatrix[1][1]; + float tanHalfFOVX = 1.0F / projMatrix[0][0]; + push.cameraTanHalfFOV = {tanHalfFOVX, tanHalfFOVY}; + push.ndcToViewMul = {push.cameraTanHalfFOV.x * 2.0f, push.cameraTanHalfFOV.y * -2.0f}; + push.ndcToViewAdd = {push.cameraTanHalfFOV.x * -1.0f, push.cameraTanHalfFOV.y * 1.0f}; + constexpr glm::vec2 texelSize = {1.0f / RENDER_EXTENT_WIDTH, 1.0f / RENDER_EXTENT_HEIGHT}; + push.ndcToViewMul_x_PixelSize = {push.ndcToViewMul.x * texelSize.x, push.ndcToViewMul.y * texelSize.y}; - push.projectionParamX = projMatrix[0][0]; - push.projectionParamY = projMatrix[1][1]; push.noiseIndex = GTAO_DENOISE_PASSES > 0 ? drawInfo.currentFrame % 64 : 0; @@ -442,7 +451,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vk_helpers::transitionImage(cmd, debugImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, - VK_IMAGE_ASPECT_COLOR_BIT); + VK_IMAGE_ASPECT_COLOR_BIT); vkCmdEndDebugUtilsLabelEXT(cmd); } diff --git a/src/renderer/post_process/post_process_types.h b/src/renderer/post_process/post_process_types.h index ad01f53a..c00a95ea 100644 --- a/src/renderer/post_process/post_process_types.h +++ b/src/renderer/post_process/post_process_types.h @@ -11,10 +11,10 @@ namespace post_process { enum class PostProcessType : uint32_t { - None = 0, + None = 0x00000000, Tonemapping = 1 << 0, Sharpening = 1 << 1, - ALL = Tonemapping | Sharpening + ALL = 0xFFFFFFFF }; } From b09f345b5940e6223711bcbe1e06970528c092b9 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sat, 29 Mar 2025 23:30:24 +0700 Subject: [PATCH 19/27] GTAO main pass first try. --- .../ground_truth/gtao_main_pass.comp | 252 +++++++++++++++--- src/core/engine.cpp | 2 +- src/core/engine.h | 2 +- .../ambient_occlusion_types.h | 2 +- .../lighting/shadows/shadow_constants.cpp | 8 +- .../lighting/shadows/shadow_constants.h | 4 +- 6 files changed, 230 insertions(+), 40 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 2917bfd5..3542ac86 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -38,6 +38,38 @@ layout (push_constant) uniform PushConstants { #define XE_HILBERT_WIDTH (1u << XE_HILBERT_LEVEL) #define XE_HILBERT_AREA (XE_HILBERT_WIDTH * XE_HILBERT_WIDTH) +#define XE_GTAO_PI (3.1415926535897932384626433832795) +#define XE_GTAO_PI_HALF (1.5707963267948966192313216916398) + +#define XE_GTAO_SLICE_COUNT_LOW 1.0f +#define XE_GTAO_SLICE_COUNT_MEDIUM 2.0f +#define XE_GTAO_SLICE_COUNT_HIGH 3.0f +#define XE_GTAO_SLICE_COUNT_ULTRA 9.0f + +#define XE_GTAO_STEPS_PER_SLICE_COUNT_LOW 2.0f +#define XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM 2.0f +#define XE_GTAO_STEPS_PER_SLICE_COUNT_HIGH 3.0f +#define XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA 3.0f + +#define XE_GTAO_OCCLUSION_TERM_SCALE 1.5f + +// http://h14s.p5r.org/2012/09/0x5f3759df.html, [Drobot2014a] Low Level Optimizations for GCN, https://blog.selfshadow.com/publications/s2016-shading-course/activision/s2016_pbs_activision_occlusion.pdf slide 63 +float XeGTAO_FastSqrt(float x) +{ + return uintBitsToFloat(0x1fbd1df5 + (floatBitsToUint(x) >> 1)); +} + +// input [-1, 1] and output [0, PI], from https://seblagarde.wordpress.com/2014/12/01/inverse-trigonometric-functions-gpu-optimization-for-amd-gcn-architecture/ +float XeGTAO_FastACos(float inX) +{ + const float PI = 3.141593; + const float HALF_PI = 1.570796; + float x = abs(inX); + float res = -0.156583 * x + HALF_PI; + res *= XeGTAO_FastSqrt(1.0 - x); + return (inX >= 0) ? res : PI - res; +} + uint hilbertIndex(uint posX, uint posY) { uint index = 0u; @@ -90,7 +122,7 @@ vec3 cheapReconstructViewSpacePosition(vec2 uv, const float viewspaceDepth, vec2 } -vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth, float depthLinearizeAdd, float depthLinearizeMult) { +vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth) { float ndcDepth = pushConstants.depthLinearizeAdd - (pushConstants.depthLinearizeMult / viewDepth); uv.y = 1 - uv.y; vec2 ndc = uv * 2.0 - 1.0; @@ -101,11 +133,19 @@ vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth, float depthLinearize } void outputWorkingTerm(ivec2 screenPos, float visibility, vec3 bentNormal, image2D outputImage){ - const float XE_GTAO_OCCLUSION_TERM_SCALE = 1.5f; visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); imageStore(outputImage, screenPos, vec4(visibility + 0.5f / 255.0f)); } +mat3 adjugate(mat4 m) { + return mat3( + cross(m[1].xyz, m[2].xyz), + cross(m[2].xyz, m[0].xyz), + cross(m[0].xyz, m[1].xyz) + ); + +} + void main() { const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); @@ -127,24 +167,23 @@ void main() { // Get view space normal by sampling normal buffer and converting from world to view (code not relevant) vec3 worldNormal = texture(normalBuffer, uv).rgb; - vec3 viewNormal = mat3(sceneData.view) * worldNormal; - // imageStore(debugImage, screenPos, vec4(viewNormal * 0.5f + 0.5f, 1.0f)); + vec3 viewNormal = adjugate(sceneData.view) * worldNormal; // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used - viewSpaceZM = viewSpaceZM * 0.99920f; + viewSpaceZM = viewSpaceZM * 0.998f; vec3 vPos = cheapReconstructViewSpacePosition(uv, viewSpaceZM, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); - + //vec3 vPos = reconstructViewSpacePosition(uv, viewSpaceZM); vec3 viewVec = normalize(-vPos); - //imageStore(debugImage, screenPos, vec4(viewVec, 1.0f)); // debug world pos - vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0)).xyz; - imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); +// vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0)).xyz; +// imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); +// return; // Per Intel // prevents normals that are facing away from the view vector - xeGTAO struggles with extreme cases, but in Vanilla it seems rare so it's disabled by default - // viewspaceNormal = normalize( viewspaceNormal + max( 0, -dot( viewspaceNormal, viewVec ) ) * viewVec ); + viewNormal = normalize(viewNormal + max(0, -dot(viewNormal, viewVec)) * viewVec); const float effectRadius = pushConstants.effectRadius * pushConstants.radiusMultiplier; @@ -155,33 +194,184 @@ void main() { const float falloffFrom = effectRadius * (1 - pushConstants.effectFalloffRange); // fadeout precompute optimisation - const float falloffMul = 1.0 / (falloffRange); + const float falloffMul = -1.0 / (falloffRange); const float falloffAdd = falloffFrom / (falloffRange) + 1.0; float visibility = 0; - // set bent normal to 0 if generating and outputting to buffer for use in deferred resolve - vec3 bentNormal = viewNormal; + vec3 bentNormal = vec3(0.0f); + + { + // NOISE + vec2 noise = spatioTemporalNoise(screenPos, pushConstants.noiseIndex); + float noiseSlice = noise.x; + float noiseSample = noise.y; - // NOISE - vec2 noise = spatioTemporalNoise(screenPos, pushConstants.noiseIndex); - float noiseSlice = noise.x; - float noiseSample = noise.y; + const float pixelTooCloseThreshold = 1.3; + const vec2 pixelDirRBViewspaceSizeAtCenterZ = viewSpaceZM.xx * pushConstants.ndcToViewMul_x_PixelSize; - const float pixelTooCloseThreshold = 1.3; - const vec2 pixelDirRBViewspaceSizeAtCenterZ = viewSpaceZM.xx * pushConstants.ndcToViewMul_x_PixelSize; + float screenspaceRadius = effectRadius / pixelDirRBViewspaceSizeAtCenterZ.x; - float screenspaceRadius = effectRadius / pixelDirRBViewspaceSizeAtCenterZ.x; - visibility += clamp((10 - screenspaceRadius)/100, 0, 1) * 0.5; + visibility += clamp((10 - screenspaceRadius)/100, 0, 1) * 0.5; - if(screenspaceRadius < pixelTooCloseThreshold) - { - visibility = 1; - const float XE_GTAO_OCCLUSION_TERM_SCALE = 1.5f; - visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); - imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); - // if outputting bent normals, need to write `viewNormal's` value to the buffer (i.e. no change to trajectory of normal) - return; - } else { - imageStore(debugImage, screenPos, vec4(vec3(1.0f, 1.0f, 0.0f), 1.0f)); + if (screenspaceRadius < pixelTooCloseThreshold) + { + visibility = 1; + visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); + imageStore(debugImage, screenPos, vec4(vec3(visibility, 0, visibility), 1.0f)); + // todo: look at how this will return 1/1.5 instead of 1 for visibility always? + // todo: (bent normal) need to write `viewNormal` value to the buffer (i.e. no change to trajectory of normal) + return; + } + + const float minS = pixelTooCloseThreshold / screenspaceRadius; + + float sliceCount = XE_GTAO_SLICE_COUNT_ULTRA; + float stepsPerSlice = XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA; + + for (float slice = 0; slice < sliceCount; slice++){ + float sliceK = (slice+noiseSlice) / sliceCount; + // lines 5, 6 from the paper + float phi = sliceK * XE_GTAO_PI; + float cosPhi = cos(phi); + float sinPhi = sin(phi); + vec2 omega = vec2(cosPhi, sinPhi);//lpfloat2 on omega causes issues with big radii + + // convert to screen units (pixels) for later use + omega *= screenspaceRadius; + + // line 8 from the paper + const vec3 directionVec = vec3(cosPhi, sinPhi, 0); + + // line 9 from the paper + const vec3 orthoDirectionVec = directionVec - (dot(directionVec, viewVec) * viewVec); + + // line 10 from the paper + //axisVec is orthogonal to directionVec and viewVec, used to define projectedNormal + const vec3 axisVec = normalize(cross(orthoDirectionVec, viewVec)); + + // alternative line 9 from the paper + // float3 orthoDirectionVec = cross( viewVec, axisVec ); + + // line 11 from the paper + vec3 projectedNormalVec = viewNormal - axisVec * dot(viewNormal, axisVec); + + // line 13 from the paper + float signNorm = sign(dot(orthoDirectionVec, projectedNormalVec)); + + // line 14 from the paper + float projectedNormalVecLength = length(projectedNormalVec); + float cosNorm = clamp(dot(projectedNormalVec, viewVec) / projectedNormalVecLength, 0, 1); + + // line 15 from the paper + float n = signNorm * XeGTAO_FastACos(cosNorm); + + // this is a lower weight target; not using -1 as in the original paper because it is under horizon, so a 'weight' has different meaning based on the normal + const float lowHorizonCos0 = cos(n+XE_GTAO_PI_HALF); + const float lowHorizonCos1 = cos(n-XE_GTAO_PI_HALF); + + // lines 17, 18 from the paper, manually unrolled the 'side' loop + float horizonCos0 = lowHorizonCos0;//-1; + float horizonCos1 = lowHorizonCos1;//-1; + + for (float step = 0; step < stepsPerSlice; step++) { + // R1 sequence (http://extremelearning.com.au/unreasonable-effectiveness-of-quasirandom-sequences/) + const float stepBaseNoise = (slice + step * stepsPerSlice) * 0.6180339887498948482;// <- this should unroll + float stepNoise = fract(noiseSample + stepBaseNoise); + + // approx line 20 from the paper, with added noise + float s = (step+stepNoise) / (stepsPerSlice);// + (lpfloat2)1e-6f); + + // additional distribution modifier + s = pow(s, sampleDistributionPower); + + // avoid sampling center pixel + s += minS; + + // approx lines 21-22 from the paper, unrolled + vec2 sampleOffset = s * omega; + + float sampleOffsetLength = length(sampleOffset); + + const int XE_GTAO_DEPTH_MIP_LEVELS = 5; + // note: when sampling, using point_point_point or point_point_linear sampler works, but linear_linear_linear will cause unwanted interpolation between neighbouring depth values on the same MIP level! + const float mipLevel = clamp(log2(sampleOffsetLength) - pushConstants.depthMipSamplingOffset, 0, XE_GTAO_DEPTH_MIP_LEVELS); + + // Snap to pixel center (more correct direction math, avoids artifacts due to sampling pos not matching depth texel center - messes up slope - but adds other + // artifacts due to them being pushed off the slice). Also use full precision for high res cases. + sampleOffset = round(sampleOffset) * sceneData.texelSize; + + vec2 sampleScreenPos0 = uv + sampleOffset; + float SZ0 = textureLod(prefilteredDepth, sampleScreenPos0, mipLevel).r; + vec3 samplePos0 = cheapReconstructViewSpacePosition(sampleScreenPos0, SZ0, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); + //vec3 samplePos0 = reconstructViewSpacePosition(sampleScreenPos0, SZ0); + + vec2 sampleScreenPos1 = uv - sampleOffset; + float SZ1 = textureLod(prefilteredDepth, sampleScreenPos1, mipLevel).r; + vec3 samplePos1 = cheapReconstructViewSpacePosition(sampleScreenPos1, SZ1, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); + //vec3 samplePos1 = reconstructViewSpacePosition(sampleScreenPos1, SZ1); + + vec3 sampleDelta0 = (samplePos0 - vec3(vPos));// using lpfloat for sampleDelta causes precision issues + vec3 sampleDelta1 = (samplePos1 - vec3(vPos));// using lpfloat for sampleDelta causes precision issues + float sampleDist0 = length(sampleDelta0); + float sampleDist1 = length(sampleDelta1); + + // approx lines 23, 24 from the paper, unrolled + vec3 sampleHorizonVec0 = (sampleDelta0 / sampleDist0); + vec3 sampleHorizonVec1 = (sampleDelta1 / sampleDist1); + + + // this is our own thickness heuristic that relies on sooner discarding samples behind the center + float falloffBase0 = length(vec3(sampleDelta0.x, sampleDelta0.y, sampleDelta0.z * (1+thinOccluderCompensation))); + float falloffBase1 = length(vec3(sampleDelta1.x, sampleDelta1.y, sampleDelta1.z * (1+thinOccluderCompensation))); + float weight0 = clamp(falloffBase0 * falloffMul + falloffAdd, 0, 1); + float weight1 = clamp(falloffBase1 * falloffMul + falloffAdd, 0, 1); + + // sample horizon cos + float shc0 = dot(sampleHorizonVec0, viewVec); + float shc1 = dot(sampleHorizonVec1, viewVec); + + // discard unwanted samples + shc0 = mix(lowHorizonCos0, shc0, weight0);// this would be more correct but too expensive: cos(lerp( acos(lowHorizonCos0), acos(shc0), weight0 )); + shc1 = mix(lowHorizonCos1, shc1, weight1);// this would be more correct but too expensive: cos(lerp( acos(lowHorizonCos1), acos(shc1), weight1 )); + + + // this is a version where thicknessHeuristic is completely disabled + horizonCos0 = max(horizonCos0, shc0); + horizonCos1 = max(horizonCos1, shc1); + } + + #if 1// I can't figure out the slight overdarkening on high slopes, so I'm adding this fudge - in the training set, 0.05 is close (PSNR 21.34) to disabled (PSNR 21.45) + projectedNormalVecLength = mix(projectedNormalVecLength, 1, 0.05); + #endif + + // line ~27, unrolled + float h0 = -XeGTAO_FastACos(horizonCos1); + float h1 = XeGTAO_FastACos(horizonCos0); + float iarc0 = (cosNorm + 2 * h0 * sin(n) - cos(2 * h0 - n)) / 4; + float iarc1 = (cosNorm + 2 * h1 * sin(n) - cos(2 * h1 - n)) / 4; + float localVisibility = projectedNormalVecLength * (iarc0+iarc1); + visibility += localVisibility; + + // todo: uncomment and fix if outputting bent normals + // see "Algorithm 2 Extension that computes bent normals b." + // lpfloat t0 = (6*sin(h0-n)-sin(3*h0-n)+6*sin(h1-n)-sin(3*h1-n)+16*sin(n)-3*(sin(h0+n)+sin(h1+n)))/12; + // lpfloat t1 = (-cos(3 * h0-n)-cos(3 * h1-n) +8 * cos(n)-3 * (cos(h0+n) +cos(h1+n)))/12; + // lpfloat3 localBentNormal = lpfloat3(directionVec.x * (lpfloat)t0, directionVec.y * (lpfloat)t0, -lpfloat(t1) ); + // localBentNormal = (lpfloat3)mul(XeGTAO_RotFromToMatrix(lpfloat3(0, 0, -1), viewVec), localBentNormal ) * projectedNormalVecLength; + // bentNormal += localBentNormal; + } + + visibility /= sliceCount; + visibility = pow(visibility, pushConstants.finalValuePower); + visibility = max(0.03, visibility);// disallow total occlusion (which wouldn't make any sense anyhow since pixel is visible but also helps with packing bent normals) + + // todo (bent normal) + // bentNormal = normalize(bentNormal) ; } + + // todo (bent normal) + visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); + imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); + + //imageStore(debugImage, screenPos, vec4(viewNormal * 0.5f + 0.5f, 1.0f)); } diff --git a/src/core/engine.cpp b/src/core/engine.cpp index a19eee4a..eb3af9ea 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -608,7 +608,7 @@ void Engine::draw(float deltaTime) vk_helpers::transitionImage(cmd, taaResolveTarget.image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, postProcessOutputBuffer.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); - postProcessPipeline->draw(cmd, post_process::PostProcessType::None); + postProcessPipeline->draw(cmd, post_process::PostProcessType::Sharpening); vk_helpers::transitionImage(cmd, postProcessOutputBuffer.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, swapchainImages[swapchainImageIndex], VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); diff --git a/src/core/engine.h b/src/core/engine.h index 547d5b5c..f3308948 100644 --- a/src/core/engine.h +++ b/src/core/engine.h @@ -173,7 +173,7 @@ class Engine bool bEnableTaa{true}; float taaBlendValue{0.1f}; bool bEnableDebugFrustumCullDraw{false}; - int32_t csmPcf{5}; + int32_t csmPcf{1}; int32_t deferredDebug{0}; bool bDrawTerrainLines{false}; diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 6985b044..1b96cc2b 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -34,7 +34,7 @@ struct GTAOPushConstants float thinOccluderCompensation = 0.0f; float finalValuePower = 2.2f; float depthMipSamplingOffset = 3.30f; - uint32_t noiseIndex; + uint32_t noiseIndex{0}; }; struct GTAODrawInfo diff --git a/src/renderer/lighting/shadows/shadow_constants.cpp b/src/renderer/lighting/shadows/shadow_constants.cpp index f20008ea..353619f1 100644 --- a/src/renderer/lighting/shadows/shadow_constants.cpp +++ b/src/renderer/lighting/shadows/shadow_constants.cpp @@ -5,8 +5,8 @@ #include "shadow_constants.h" float will_engine::shadows::CASCADE_BIAS[SHADOW_CASCADE_COUNT][2] = { - {400.0f, 7.0f}, - {350.0f, 6.0f}, - {300.0f, 5.0f}, - {250.0f, 5.0f}, + {400.0f, 1.0f}, + {350.0f, 1.0f}, + {300.0f, 1.0f}, + {250.0f, 1.0f}, }; diff --git a/src/renderer/lighting/shadows/shadow_constants.h b/src/renderer/lighting/shadows/shadow_constants.h index 65018b5f..0d04db8b 100644 --- a/src/renderer/lighting/shadows/shadow_constants.h +++ b/src/renderer/lighting/shadows/shadow_constants.h @@ -9,11 +9,11 @@ namespace will_engine::shadows { -static constexpr float LAMBDA = 0.5f; +static constexpr float LAMBDA = 0.8f; static constexpr float OVERLAP = 1.05f; static constexpr uint32_t SHADOW_CASCADE_COUNT = 4; static constexpr float CASCADE_NEAR_PLANE = 0.1f; -static constexpr float CASCADE_FAR_PLANE = 200.0f; +static constexpr float CASCADE_FAR_PLANE = 1000.0f; extern float CASCADE_BIAS[SHADOW_CASCADE_COUNT][2]; static constexpr int32_t CASCADE_WIDTH{2048}; static constexpr int32_t CASCADE_HEIGHT{2048}; From 51bfdaf69c0fcccf367b12da3df2f295900445fd Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sun, 30 Mar 2025 16:39:19 +0700 Subject: [PATCH 20/27] GTAO spatial filter progress. --- .../ground_truth/gtao_depth_prefilter.comp | 45 +++--- .../ground_truth/gtao_main_pass.comp | 71 ++++++---- .../ground_truth/gtao_spatial_filter.comp | 38 +++++- shaders/include/gtao.glsl | 49 +++++++ src/core/engine.cpp | 5 +- src/core/engine.h | 1 + src/renderer/imgui_wrapper.cpp | 6 +- .../ambient_occlusion_types.h | 2 + .../ground_truth_ambient_occlusion.cpp | 128 +++++++++++++++--- .../ground_truth_ambient_occlusion.h | 13 +- src/renderer/renderer_constants.h | 1 + 11 files changed, 272 insertions(+), 87 deletions(-) create mode 100644 shaders/include/gtao.glsl diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index ec9bb0bc..0439a7a7 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -2,6 +2,7 @@ #extension GL_EXT_nonuniform_qualifier: enable #include "scene.glsl" +#include "gtao.glsl" layout(local_size_x = 8, local_size_y = 8) in; @@ -17,29 +18,6 @@ layout (r16f, set = 1, binding = 4) uniform image2D outDepth3; layout (r16f, set = 1, binding = 5) uniform image2D outDepth4; layout (rgba8, set = 1, binding = 6) uniform image2D debugImage; -layout (push_constant) uniform PushConstants { - vec2 cameraTanHalfFOV; - - vec2 ndcToViewMul; - vec2 ndcToViewAdd; - - vec2 ndcToViewMul_x_PixelSize; - - float depthLinearizeMult; - float depthLinearizeAdd; - - float effectRadius; - float effectFalloffRange; - float denoiseBlurBeta; - - float radiusMultiplier; - float sampleDistributionPower; - float thinOccluderCompensation; - float finalValuePower; - float depthMipSamplingOffset; - uint noiseIndex; -} pushConstants; - float screenToViewSpaceDepth(float screenDepth, float depthLinearizeMul, float depthLinearizeAdd) { // Optimization by XeGTAO // https://github.com/GameTechDev/XeGTAO/blob/a5b1686c7ea37788eeb3576b5be47f7c03db532c/Source/Rendering/Shaders/XeGTAO.hlsli#L112 @@ -88,10 +66,16 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - float rDepth0 = texture(depthImage, uv + vec2(0.0, 0.0) * sceneData.texelSize).r; - float rDepth1 = texture(depthImage, uv + vec2(1.0, 0.0) * sceneData.texelSize).r; - float rDepth2 = texture(depthImage, uv + vec2(0.0, 1.0) * sceneData.texelSize).r; - float rDepth3 = texture(depthImage, uv + vec2(1.0, 1.0) * sceneData.texelSize).r; + // todo: optimize with textureGather? + vec4 depths = textureGather(depthImage, uv); + float rDepth0 = depths.w; // top-left + float rDepth1 = depths.z; // top-right + float rDepth2 = depths.x; // bottom-left + float rDepth3 = depths.y; // bottom-right +// float rDepth0 = texture(depthImage, uv + vec2(0.0, 0.0) * sceneData.texelSize).r; +// float rDepth1 = texture(depthImage, uv + vec2(1.0, 0.0) * sceneData.texelSize).r; +// float rDepth2 = texture(depthImage, uv + vec2(0.0, 1.0) * sceneData.texelSize).r; +// float rDepth3 = texture(depthImage, uv + vec2(1.0, 1.0) * sceneData.texelSize).r; float depth0 = clampDepth(screenToViewSpaceDepth(rDepth0, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); float depth1 = clampDepth(screenToViewSpaceDepth(rDepth1, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); @@ -103,6 +87,13 @@ void main() { imageStore(outDepth0, screenPos + ivec2(0, 1), vec4(depth2)); imageStore(outDepth0, screenPos + ivec2(1, 1), vec4(depth3)); + if (pushConstants.debug == 1){ + imageStore(debugImage, screenPos + ivec2(0, 0), vec4(vec3(depth0 / 1000.0f), 1.0f)); + imageStore(debugImage, screenPos + ivec2(1, 0), vec4(vec3(depth1 / 1000.0f), 1.0f)); + imageStore(debugImage, screenPos + ivec2(0, 1), vec4(vec3(depth2 / 1000.0f), 1.0f)); + imageStore(debugImage, screenPos + ivec2(1, 1), vec4(vec3(depth3 / 1000.0f), 1.0f)); + return; + } // MIP 1 float dm1 = depthMipFilter(depth0, depth1, depth2, depth3, pushConstants.effectRadius, pushConstants.radiusMultiplier, pushConstants.effectFalloffRange); diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 3542ac86..867b91fb 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -1,6 +1,7 @@ #version 460 #include "scene.glsl" +#include "gtao.glsl" layout(local_size_x = 16, local_size_y = 16) in; @@ -9,30 +10,9 @@ layout(local_size_x = 16, local_size_y = 16) in; layout (set = 1, binding = 0) uniform sampler2D prefilteredDepth; layout (set = 1, binding = 1) uniform sampler2D normalBuffer; layout (r8, set = 1, binding = 2) uniform image2D aoOutput; -layout (rgba8, set = 1, binding = 3) uniform image2D debugImage; +layout (r8, set = 1, binding = 3) uniform image2D edgeDataOutput; +layout (rgba8, set = 1, binding = 4) uniform image2D debugImage; -layout (push_constant) uniform PushConstants { - vec2 cameraTanHalfFOV; - - vec2 ndcToViewMul; - vec2 ndcToViewAdd; - - vec2 ndcToViewMul_x_PixelSize; - - float depthLinearizeMult; - float depthLinearizeAdd; - - float effectRadius; - float effectFalloffRange; - float denoiseBlurBeta; - - float radiusMultiplier; - float sampleDistributionPower; - float thinOccluderCompensation; - float finalValuePower; - float depthMipSamplingOffset; - uint noiseIndex; -} pushConstants; #define XE_HILBERT_LEVEL 6u #define XE_HILBERT_WIDTH (1u << XE_HILBERT_LEVEL) @@ -161,14 +141,35 @@ void main() { float viewSpaceZT = textureLod(prefilteredDepth, uv + vec2(0.0, 1.0) * sceneData.texelSize, 0).r; float viewSpaceZB = textureLod(prefilteredDepth, uv + vec2(0.0, -1.0) * sceneData.texelSize, 0).r; + // Theoretically Equivalent, but visibly more noisy for some reason. + // vec4 leftGather = textureGather(prefilteredDepth, uv + vec2(-1.0, 0.0) * sceneData.texelSize, 0); + // float viewSpaceZL = leftGather.x; + // float viewSpaceZT = leftGather.w; + // float viewSpaceZM = leftGather.y; + // vec4 bottomather = textureGather(prefilteredDepth, uv + vec2(0.0, -1.0) * sceneData.texelSize, 0); + // float viewSpaceZB = bottomather.x; + // float viewSpaceZR = bottomather.w; + vec4 edges = calculateDepthEdges(viewSpaceZM, viewSpaceZL, viewSpaceZR, viewSpaceZT, viewSpaceZB); + float packedEdges = XeGTAO_PackEdges(edges); + imageStore(edgeDataOutput, screenPos, vec4(packedEdges)); + float minEdge = min(min(edges.x, edges.y), min(edges.z, edges.w)); - // imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); + + if (pushConstants.debug == 2){ + imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); + return; + } // Get view space normal by sampling normal buffer and converting from world to view (code not relevant) vec3 worldNormal = texture(normalBuffer, uv).rgb; vec3 viewNormal = adjugate(sceneData.view) * worldNormal; + if (pushConstants.debug == 2){ + imageStore(debugImage, screenPos, vec4(viewNormal, 1.0f)); + return; + } + // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used viewSpaceZM = viewSpaceZM * 0.998f; @@ -177,9 +178,11 @@ void main() { vec3 viewVec = normalize(-vPos); // debug world pos -// vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0)).xyz; -// imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); -// return; + if (pushConstants.debug == 3){ + vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0)).xyz; + imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); + return; + } // Per Intel // prevents normals that are facing away from the view vector - xeGTAO struggles with extreme cases, but in Vanilla it seems rare so it's disabled by default @@ -217,7 +220,12 @@ void main() { { visibility = 1; visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); - imageStore(debugImage, screenPos, vec4(vec3(visibility, 0, visibility), 1.0f)); + if (pushConstants.debug == 4){ + imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); + return; + } + + imageStore(aoOutput, screenPos, vec4(visibility)); // todo: look at how this will return 1/1.5 instead of 1 for visibility always? // todo: (bent normal) need to write `viewNormal` value to the buffer (i.e. no change to trajectory of normal) return; @@ -371,7 +379,10 @@ void main() { // todo (bent normal) visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); - imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); + if (pushConstants.debug == 4){ + imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); + return; + } - //imageStore(debugImage, screenPos, vec4(viewNormal * 0.5f + 0.5f, 1.0f)); + imageStore(aoOutput, screenPos, vec4(visibility)); } diff --git a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp index ee6acbe2..6bba912a 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp @@ -1,14 +1,46 @@ #version 460 #include "scene.glsl" +#include "gtao.glsl" + +layout(local_size_x = 16, local_size_y = 16) in; // layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl layout (set = 1, binding = 0) uniform sampler2D rawAO; -layout (set = 1, binding = 1) uniform sampler2D depthBuffer; -layout (set = 1, binding = 2) uniform sampler2D normalBuffer; -layout (r8, set = 1, binding = 3) uniform image2D filteredAO; +layout (set = 1, binding = 1) uniform sampler2D edgeData; +layout (r8, set = 1, binding = 2) uniform image2D filteredAO; +layout (rgba8, set = 1, binding = 3) uniform image2D debugImage; void main() { + const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + + if (screenPos.x > sceneData.renderTargetSize.x || screenPos.y > sceneData.renderTargetSize.y) { + return; + } + + vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; + + float blurAmount = pushConstants.denoiseBlurBeta / 5.0f; + float diagWeight = 0.85 * 0.5; + + // each dispatch operates on 2x1 pixels + + vec4 edges = XeGTAO_UnpackEdges(texture(edgeData, uv).r); + float minEdge = min(min(edges.x, edges.y), min(edges.z, edges.w)); + + if (pushConstants.debug == 5){ + imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); + return; + } + if (pushConstants.debug == 5) { + imageStore(debugImage, screenPos, vec4(edges.xyz, 1.0f)); + return; + if (screenPos.x % 2 == 0 && screenPos.y %2 == 0){ + imageStore(debugImage, screenPos, vec4(vec3(1.0f), 1.0f)); + } else { + imageStore(debugImage, screenPos, vec4(vec3(0.5f), 1.0f)); + } + } } diff --git a/shaders/include/gtao.glsl b/shaders/include/gtao.glsl new file mode 100644 index 00000000..c51b59b5 --- /dev/null +++ b/shaders/include/gtao.glsl @@ -0,0 +1,49 @@ +layout (push_constant) uniform PushConstants { + vec2 cameraTanHalfFOV; + + vec2 ndcToViewMul; + vec2 ndcToViewAdd; + + vec2 ndcToViewMul_x_PixelSize; + + float depthLinearizeMult; + float depthLinearizeAdd; + + float effectRadius; + float effectFalloffRange; + float denoiseBlurBeta; + + float radiusMultiplier; + float sampleDistributionPower; + float thinOccluderCompensation; + float finalValuePower; + float depthMipSamplingOffset; + uint noiseIndex; + + int debug; +} pushConstants; + + +// packing/unpacking for edges; 2 bits per edge mean 4 gradient values (0, 0.33, 0.66, 1) for smoother transitions! +float XeGTAO_PackEdges(vec4 edgesLRTB) +{ + // integer version: + // edgesLRTB = saturate(edgesLRTB) * 2.9.xxxx + 0.5.xxxx; + // return (((uint)edgesLRTB.x) << 6) + (((uint)edgesLRTB.y) << 4) + (((uint)edgesLRTB.z) << 2) + (((uint)edgesLRTB.w)); + // + // optimized, should be same as above + edgesLRTB = round(clamp(edgesLRTB, 0, 1) * 2.9); + return dot(edgesLRTB, vec4(64.0 / 255.0, 16.0 / 255.0, 4.0 / 255.0, 1.0 / 255.0)); +} + +vec4 XeGTAO_UnpackEdges(float _packedVal) +{ + int packedVal = int(_packedVal * 255.5f); + vec4 edgesLRTB; + edgesLRTB.x = float((packedVal >> 6) & 0x03) / 3.0;// there's really no need for mask (as it's an 8 bit input) but I'll leave it in so it doesn't cause any trouble in the future + edgesLRTB.y = float((packedVal >> 4) & 0x03) / 3.0; + edgesLRTB.z = float((packedVal >> 2) & 0x03) / 3.0; + edgesLRTB.w = float((packedVal >> 0) & 0x03) / 3.0; + + return clamp(edgesLRTB, 0, 1); +} \ No newline at end of file diff --git a/src/core/engine.cpp b/src/core/engine.cpp index eb3af9ea..c0137938 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -208,6 +208,7 @@ void Engine::initRenderer() ambientOcclusionPipeline->setupDepthPrefilterDescriptorBuffer(depthImage.imageView); ambientOcclusionPipeline->setupAmbientOcclusionDescriptorBuffer(normalRenderTarget.imageView); + ambientOcclusionPipeline->setupSpatialFilteringDescriptorBuffer(depthImage.imageView, normalRenderTarget.imageView); const deferred_resolve::DeferredResolveDescriptor deferredResolveDescriptor{ normalRenderTarget.imageView, @@ -565,9 +566,11 @@ void Engine::draw(float deltaTime) vk_helpers::transitionImage(cmd, depthImage.image, VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_DEPTH_BIT); vk_helpers::transitionImage(cmd, drawImage.image, VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); + ambient_occlusion::GTAOPushConstants gtaoPush{}; + gtaoPush.debug = gtaoDebug; ambient_occlusion::GTAODrawInfo gtaoDrawInfo{ camera, - {}, + gtaoPush, frameNumber, sceneDataDescriptorBuffer.getDescriptorBufferBindingInfo(), sceneDataDescriptorBuffer.getDescriptorBufferSize() * currentFrameOverlap diff --git a/src/core/engine.h b/src/core/engine.h index f3308948..da6d9894 100644 --- a/src/core/engine.h +++ b/src/core/engine.h @@ -175,6 +175,7 @@ class Engine bool bEnableDebugFrustumCullDraw{false}; int32_t csmPcf{1}; int32_t deferredDebug{0}; + int32_t gtaoDebug{5}; bool bDrawTerrainLines{false}; void hotReloadShaders() const; diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index 1a5e17ae..603fd10d 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -877,13 +877,11 @@ void ImguiWrapper::imguiInterface(Engine* engine) ImGui::End(); if (ImGui::Begin("Discardable Debug")) { - static int32_t gtaoMip; - constexpr uint32_t minMip = 0; - constexpr uint32_t maxMip = 4; + ImGui::InputInt("GTAO Debug", &engine->gtaoDebug); if (aoDebugTextureImguiId == VK_NULL_HANDLE) { if (engine->ambientOcclusionPipeline->debugImage.image != VK_NULL_HANDLE) { - aoDebugTextureImguiId = ImGui_ImplVulkan_AddTexture(engine->resourceManager->getDefaultSamplerLinear(), + aoDebugTextureImguiId = ImGui_ImplVulkan_AddTexture(engine->resourceManager->getDefaultSamplerNearest(), engine->ambientOcclusionPipeline->debugImage.imageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 1b96cc2b..5d3f0369 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -35,6 +35,8 @@ struct GTAOPushConstants float finalValuePower = 2.2f; float depthMipSamplingOffset = 3.30f; uint32_t noiseIndex{0}; + + int32_t debug{0}; }; struct GTAODrawInfo diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index e2dcc0dd..3c7fbf0e 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -64,7 +64,6 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth VkImageUsageFlags usage{}; usage |= VK_IMAGE_USAGE_STORAGE_BIT; usage |= VK_IMAGE_USAGE_SAMPLED_BIT; - usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(depthPrefilterFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); @@ -82,8 +81,17 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth VkSamplerCreateInfo samplerInfo = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO}; samplerInfo.magFilter = VK_FILTER_NEAREST; samplerInfo.minFilter = VK_FILTER_NEAREST; - - depthPrefilterSampler = resourceManager.createSampler(samplerInfo); + samplerInfo.mipmapMode = VK_SAMPLER_MIPMAP_MODE_NEAREST; + samplerInfo.addressModeU = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeV = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.addressModeW = VK_SAMPLER_ADDRESS_MODE_CLAMP_TO_EDGE; + samplerInfo.anisotropyEnable = VK_FALSE; + samplerInfo.maxAnisotropy = 1.0f; + samplerInfo.compareEnable = VK_FALSE; + samplerInfo.minLod = 0.0f; + samplerInfo.maxLod = 0.0f; + + depthSampler = resourceManager.createSampler(samplerInfo); } // AO Calculation @@ -92,7 +100,8 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // pre-filtered depth layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT normal buffer layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // ao output - layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // debug image + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // edge data output + layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // debug image ambientOcclusionSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); @@ -122,12 +131,18 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth VkImageUsageFlags usage{}; usage |= VK_IMAGE_USAGE_STORAGE_BIT; usage |= VK_IMAGE_USAGE_SAMPLED_BIT; - usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; - usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); ambientOcclusionImage = resourceManager.createImage(imgInfo); + usage = {}; + usage |= VK_IMAGE_USAGE_STORAGE_BIT; + usage |= VK_IMAGE_USAGE_SAMPLED_BIT; + + imgInfo = vk_helpers::imageCreateInfo(edgeDataFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); + edgeDataImage = resourceManager.createImage(imgInfo); + + // Depth Mip sampler { VkSamplerCreateInfo samplerInfo = {.sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO}; @@ -143,7 +158,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth samplerInfo.minLod = 0.0f; samplerInfo.maxLod = DEPTH_PREFILTER_MIP_COUNT - 1; - ambientOcclusionDepthSampler = resourceManager.createSampler(samplerInfo); + depthPrefilterSampler = resourceManager.createSampler(samplerInfo); } // Normals sampler @@ -161,7 +176,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth samplerInfo.minLod = 0.0f; samplerInfo.maxLod = 0.0f; - ambientOcclusionNormalsSampler = resourceManager.createSampler(samplerInfo); + normalsSampler = resourceManager.createSampler(samplerInfo); } } @@ -169,9 +184,9 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth { DescriptorLayoutBuilder layoutBuilder; layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // raw ao - layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT depth buffer - layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT normal buffer - layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // filtered ao + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // edge data + layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // filtered ao + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // debug image spatialFilteringSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); @@ -271,7 +286,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTrut } resourceManager.destroyImage(depthPrefilterImage); - resourceManager.destroySampler(depthPrefilterSampler); + resourceManager.destroySampler(depthSampler); resourceManager.destroyDescriptorBuffer(depthPrefilterDescriptorBuffer); @@ -280,9 +295,10 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTrut resourceManager.destroyPipelineLayout(ambientOcclusionPipelineLayout); resourceManager.destroyPipeline(ambientOcclusionPipeline); - resourceManager.destroySampler(ambientOcclusionDepthSampler); - resourceManager.destroySampler(ambientOcclusionNormalsSampler); + resourceManager.destroySampler(depthPrefilterSampler); + resourceManager.destroySampler(normalsSampler); resourceManager.destroyImage(ambientOcclusionImage); + resourceManager.destroyImage(edgeDataImage); resourceManager.destroyDescriptorBuffer(ambientOcclusionDescriptorBuffer); @@ -315,7 +331,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupD imageDescriptors.push_back( { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, - {resourceManager.getDefaultSamplerLinear(), depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + {depthSampler, depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, false } ); @@ -338,7 +354,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupD resourceManager.setupDescriptorBufferSampler(depthPrefilterDescriptorBuffer, imageDescriptors, 0); } -void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView) +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupAmbientOcclusionDescriptorBuffer(const VkImageView& normalsImageView) { std::vector imageDescriptors{}; imageDescriptors.reserve(4); @@ -346,13 +362,13 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupA imageDescriptors.push_back( { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, - {ambientOcclusionDepthSampler, depthPrefilterImage.imageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + {depthPrefilterSampler, depthPrefilterImage.imageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, false }); imageDescriptors.push_back( { VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, - {ambientOcclusionNormalsSampler, normalsImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + {normalsSampler, normalsImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, false }); imageDescriptors.push_back( @@ -361,6 +377,12 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupA {VK_NULL_HANDLE, ambientOcclusionImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, false }); + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, + {VK_NULL_HANDLE, edgeDataImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + false + }); imageDescriptors.push_back({ VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, {VK_NULL_HANDLE, debugImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, @@ -370,6 +392,50 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupA resourceManager.setupDescriptorBufferSampler(ambientOcclusionDescriptorBuffer, imageDescriptors, 0); } +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupSpatialFilteringDescriptorBuffer(const VkImageView& depthImageView, + const VkImageView& normalsImageView) +{ + std::vector imageDescriptors{}; + imageDescriptors.reserve(5); + + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + {depthSampler, ambientOcclusionImage.imageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + false + }); + // imageDescriptors.push_back( + // { + // VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + // {depthSampler, depthImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + // false + // }); + // imageDescriptors.push_back( + // { + // VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + // {normalsSampler, normalsImageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + // false + // }); + imageDescriptors.push_back( + { + VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, + {depthSampler, edgeDataImage.imageView, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL}, + false + }); + imageDescriptors.push_back({ + VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, + {VK_NULL_HANDLE, spatialFilteringImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + false + }); + imageDescriptors.push_back({ + VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, + {VK_NULL_HANDLE, debugImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + false + }); + + resourceManager.setupDescriptorBufferSampler(spatialFilteringDescriptorBuffer, imageDescriptors, 0); +} + void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const { VkDebugUtilsLabelEXT label{}; @@ -448,6 +514,32 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); + vk_helpers::transitionImage(cmd, spatialFilteringImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, + VK_IMAGE_ASPECT_COLOR_BIT); + // Spatial Filtering + { + vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, spatialFilteringPipeline); + vkCmdPushConstants(cmd, spatialFilteringPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); + + VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; + bindingInfos[0] = drawInfo.sceneDataBinding; + bindingInfos[1] = spatialFilteringDescriptorBuffer.getDescriptorBufferBindingInfo(); + vkCmdBindDescriptorBuffersEXT(cmd, 2, bindingInfos); + + constexpr std::array indices{0, 1}; + const std::array offsets{drawInfo.sceneDataOffset, ZERO_DEVICE_SIZE}; + + vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, spatialFilteringPipelineLayout, 0, 2, indices.data(), offsets.data()); + + const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); + vkCmdDispatch(cmd, x, y, 1); + } + + vk_helpers::transitionImage(cmd, spatialFilteringImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + VK_IMAGE_ASPECT_COLOR_BIT); + + vk_helpers::transitionImage(cmd, debugImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index c243313f..ffefe4c8 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -23,7 +23,9 @@ class GroundTruthAmbientOcclusionPipeline void setupDepthPrefilterDescriptorBuffer(const VkImageView& depthImageView); - void setupAmbientOcclusionDescriptorBuffer(VkImageView normalsImageView); + void setupAmbientOcclusionDescriptorBuffer(const VkImageView& normalsImageView); + + void setupSpatialFilteringDescriptorBuffer(const VkImageView& depthImageView, const VkImageView& normalsImageView); void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const; @@ -39,7 +41,7 @@ class GroundTruthAmbientOcclusionPipeline VkPipelineLayout depthPrefilterPipelineLayout{VK_NULL_HANDLE}; VkPipeline depthPrefilterPipeline{VK_NULL_HANDLE}; - VkSampler depthPrefilterSampler{VK_NULL_HANDLE}; + VkSampler depthSampler{VK_NULL_HANDLE}; // 16 vs 32. look at cost later. VkFormat depthPrefilterFormat{VK_FORMAT_R16_SFLOAT}; @@ -53,13 +55,16 @@ class GroundTruthAmbientOcclusionPipeline VkPipelineLayout ambientOcclusionPipelineLayout{VK_NULL_HANDLE}; VkPipeline ambientOcclusionPipeline{VK_NULL_HANDLE}; - VkSampler ambientOcclusionDepthSampler{VK_NULL_HANDLE}; - VkSampler ambientOcclusionNormalsSampler{VK_NULL_HANDLE}; + VkSampler depthPrefilterSampler{VK_NULL_HANDLE}; + VkSampler normalsSampler{VK_NULL_HANDLE}; // 8 is supposedly enough? VkFormat ambientOcclusionFormat{VK_FORMAT_R8_UNORM}; AllocatedImage ambientOcclusionImage{VK_NULL_HANDLE}; + VkFormat edgeDataFormat{VK_FORMAT_R8_UNORM}; + AllocatedImage edgeDataImage{VK_NULL_HANDLE}; + DescriptorBufferSampler ambientOcclusionDescriptorBuffer; private: // Spatial Filtering diff --git a/src/renderer/renderer_constants.h b/src/renderer/renderer_constants.h index 574825e0..54db7310 100644 --- a/src/renderer/renderer_constants.h +++ b/src/renderer/renderer_constants.h @@ -11,6 +11,7 @@ constexpr char ENGINE_NAME[] = "Will Engine"; constexpr bool USING_REVERSED_DEPTH_BUFFER = true; constexpr VkDeviceSize ZERO_DEVICE_SIZE = 0; constexpr VkExtent2D RENDER_EXTENTS{1920, 1080}; +//constexpr VkExtent2D RENDER_EXTENTS{3840, 2160}; constexpr float RENDER_EXTENT_WIDTH{RENDER_EXTENTS.width}; constexpr float RENDER_EXTENT_HEIGHT{RENDER_EXTENTS.height}; From c70f7f47be8e7ab8adeeef7b4582ed6f1da4b02a Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Sun, 30 Mar 2025 17:05:01 +0700 Subject: [PATCH 21/27] Fix texture gathers. --- .../ground_truth/gtao_depth_prefilter.comp | 2 +- .../ground_truth/gtao_main_pass.comp | 21 +++++++------------ 2 files changed, 8 insertions(+), 15 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index 0439a7a7..a096d70a 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -67,7 +67,7 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; // todo: optimize with textureGather? - vec4 depths = textureGather(depthImage, uv); + vec4 depths = textureGatherOffset(depthImage,screenPos * sceneData.texelSize, ivec2(1,1), 0); float rDepth0 = depths.w; // top-left float rDepth1 = depths.z; // top-right float rDepth2 = depths.x; // bottom-left diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 867b91fb..eabc759a 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -135,20 +135,13 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - float viewSpaceZM = textureLod(prefilteredDepth, uv + vec2(0.0, 0.0) * sceneData.texelSize, 0).r; - float viewSpaceZL = textureLod(prefilteredDepth, uv + vec2(-1.0, 0.0) * sceneData.texelSize, 0).r; - float viewSpaceZR = textureLod(prefilteredDepth, uv + vec2(1.0, 0.0) * sceneData.texelSize, 0).r; - float viewSpaceZT = textureLod(prefilteredDepth, uv + vec2(0.0, 1.0) * sceneData.texelSize, 0).r; - float viewSpaceZB = textureLod(prefilteredDepth, uv + vec2(0.0, -1.0) * sceneData.texelSize, 0).r; - - // Theoretically Equivalent, but visibly more noisy for some reason. - // vec4 leftGather = textureGather(prefilteredDepth, uv + vec2(-1.0, 0.0) * sceneData.texelSize, 0); - // float viewSpaceZL = leftGather.x; - // float viewSpaceZT = leftGather.w; - // float viewSpaceZM = leftGather.y; - // vec4 bottomather = textureGather(prefilteredDepth, uv + vec2(0.0, -1.0) * sceneData.texelSize, 0); - // float viewSpaceZB = bottomather.x; - // float viewSpaceZR = bottomather.w; + vec4 valuesUL = textureGather(prefilteredDepth, screenPos * sceneData.texelSize, 0); + vec4 valuesBR = textureGatherOffset(prefilteredDepth, screenPos * sceneData.texelSize, ivec2(1,1), 0); + float viewSpaceZM = valuesUL.y; + const float viewSpaceZL = valuesUL.x; + const float viewSpaceZR = valuesUL.z; + const float viewSpaceZT = valuesBR.z; + const float viewSpaceZB = valuesBR.x; vec4 edges = calculateDepthEdges(viewSpaceZM, viewSpaceZL, viewSpaceZR, viewSpaceZT, viewSpaceZB); float packedEdges = XeGTAO_PackEdges(edges); From b3b64c27f7254400f329d9a50d75f3261dec217b Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 31 Mar 2025 14:13:15 +0700 Subject: [PATCH 22/27] GTAO denoise pass. --- .../ground_truth/gtao_depth_prefilter.comp | 2 +- .../ground_truth/gtao_main_pass.comp | 6 +- .../ground_truth/gtao_spatial_filter.comp | 121 +++++++++++++++--- shaders/include/gtao.glsl | 3 +- .../ambient_occlusion_types.h | 1 + .../ground_truth_ambient_occlusion.cpp | 3 +- 6 files changed, 115 insertions(+), 21 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index a096d70a..f41ea7d2 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -67,7 +67,7 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; // todo: optimize with textureGather? - vec4 depths = textureGatherOffset(depthImage,screenPos * sceneData.texelSize, ivec2(1,1), 0); + vec4 depths = textureGatherOffset(depthImage, uv, ivec2(1,1), 0); float rDepth0 = depths.w; // top-left float rDepth1 = depths.z; // top-right float rDepth2 = depths.x; // bottom-left diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index eabc759a..f16bf251 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -135,8 +135,10 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - vec4 valuesUL = textureGather(prefilteredDepth, screenPos * sceneData.texelSize, 0); - vec4 valuesBR = textureGatherOffset(prefilteredDepth, screenPos * sceneData.texelSize, ivec2(1,1), 0); + //vec4 valuesUL = textureGather(prefilteredDepth, screenPos * sceneData.texelSize, 0); + //vec4 valuesBR = textureGatherOffset(prefilteredDepth, screenPos * sceneData.texelSize, ivec2(1,1), 0); + vec4 valuesUL = textureGather(prefilteredDepth, uv, 0); + vec4 valuesBR = textureGatherOffset(prefilteredDepth, uv, ivec2(1,1), 0); float viewSpaceZM = valuesUL.y; const float viewSpaceZL = valuesUL.x; const float viewSpaceZR = valuesUL.z; diff --git a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp index 6bba912a..4dcaf681 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp @@ -12,8 +12,20 @@ layout (set = 1, binding = 1) uniform sampler2D edgeData; layout (r8, set = 1, binding = 2) uniform image2D filteredAO; layout (rgba8, set = 1, binding = 3) uniform image2D debugImage; +#define XE_GTAO_OCCLUSION_TERM_SCALE 1.5f + +void XeGTAO_AddSample(float ssaoValue, float edgeValue, inout float sum, inout float sumWeight) +{ + float weight = edgeValue; + + sum += (weight * ssaoValue); + sumWeight += weight; +} + void main() { - const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + // each dispatch operates on 2x1 pixels + //const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); + const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy) * ivec2(2, 1); if (screenPos.x > sceneData.renderTargetSize.x || screenPos.y > sceneData.renderTargetSize.y) { return; @@ -21,26 +33,103 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - float blurAmount = pushConstants.denoiseBlurBeta / 5.0f; + // if final apply use below (see, using multiple denoise passes) + float blurAmount = pushConstants.isFinalDenoisePass == 1 ? pushConstants.denoiseBlurBeta : pushConstants.denoiseBlurBeta / 5.0f; + float diagWeight = 0.85 * 0.5; - // each dispatch operates on 2x1 pixels + float aoTerm[2];// pixel pixCoordBase and pixel pixCoordBase + int2( 1, 0 ) + vec4 edgesC_LRTB[2]; + float weightTL[2]; + float weightTR[2]; + float weightBL[2]; + float weightBR[2]; - vec4 edges = XeGTAO_UnpackEdges(texture(edgeData, uv).r); - float minEdge = min(min(edges.x, edges.y), min(edges.z, edges.w)); - if (pushConstants.debug == 5){ - imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); - return; - } - if (pushConstants.debug == 5) { - imageStore(debugImage, screenPos, vec4(edges.xyz, 1.0f)); - return; - if (screenPos.x % 2 == 0 && screenPos.y %2 == 0){ - imageStore(debugImage, screenPos, vec4(vec3(1.0f), 1.0f)); - } else { - imageStore(debugImage, screenPos, vec4(vec3(0.5f), 1.0f)); + vec4 edgesQ0 = textureGatherOffset(edgeData, uv, ivec2(0, 0), 0); + vec4 edgesQ1 = textureGatherOffset(edgeData, uv, ivec2(2, 0), 0); + vec4 edgesQ2 = textureGatherOffset(edgeData, uv, ivec2(1, 2), 0); + + vec4 visQ0 = textureGatherOffset(rawAO, uv, ivec2(0, 0), 0); + vec4 visQ1 = textureGatherOffset(rawAO, uv, ivec2(2, 0), 0); + vec4 visQ2 = textureGatherOffset(rawAO, uv, ivec2(0, 2), 0); + vec4 visQ3 = textureGatherOffset(rawAO, uv, ivec2(2, 2), 0); + + for (int side = 0; side < 2; side++) + { + const ivec2 sideScreenPos = ivec2(screenPos.x + side, screenPos.y); + + vec4 edgesL_LRTB = XeGTAO_UnpackEdges(side == 0 ? edgesQ0.x : edgesQ0.y); + vec4 edgesT_LRTB = XeGTAO_UnpackEdges(side == 0 ? edgesQ0.z : edgesQ1.w); + vec4 edgesR_LRTB = XeGTAO_UnpackEdges(side == 0 ? edgesQ1.x : edgesQ1.y); + vec4 edgesB_LRTB = XeGTAO_UnpackEdges(side == 0 ? edgesQ2.w : edgesQ2.z); + + edgesC_LRTB[side] = XeGTAO_UnpackEdges(side==0 ? edgesQ0.y : edgesQ1.x); + + // Edges aren't perfectly symmetrical: edge detection algorithm does not guarantee that a left edge on the right pixel will match the right edge on the left pixel (although + // they will match in majority of cases). This line further enforces the symmetricity, creating a slightly sharper blur. Works real nice with TAA. + edgesC_LRTB[side] *= vec4(edgesL_LRTB.y, edgesR_LRTB.x, edgesT_LRTB.w, edgesB_LRTB.z); + + // this allows some small amount of AO leaking from neighbours if there are 3 or 4 edges; this reduces both spatial and temporal aliasing + const float leak_threshold = 2.5; + const float leak_strength = 0.5; + float edginess = (clamp(4.0 - leak_threshold - dot(edgesC_LRTB[side], vec4(1.0, 1.0, 1.0, 1.0)), 0.0, 1.0) / (4-leak_threshold)) * leak_strength; + edgesC_LRTB[side] = clamp(edgesC_LRTB[side] + edginess, 0.0, 1.0); + + // for diagonals; used by first and second pass + weightTL[side] = diagWeight * (edgesC_LRTB[side].x * edgesL_LRTB.z + edgesC_LRTB[side].z * edgesT_LRTB.x); + weightTR[side] = diagWeight * (edgesC_LRTB[side].z * edgesT_LRTB.y + edgesC_LRTB[side].y * edgesR_LRTB.z); + weightBL[side] = diagWeight * (edgesC_LRTB[side].w * edgesB_LRTB.x + edgesC_LRTB[side].x * edgesL_LRTB.w); + weightBR[side] = diagWeight * (edgesC_LRTB[side].y * edgesR_LRTB.w + edgesC_LRTB[side].w * edgesB_LRTB.y); + + // first pass + float ssaoValue = side==0 ? visQ0[1] : visQ1[0]; + float ssaoValueL = side==0 ? visQ0[0] : visQ0[1]; + float ssaoValueT = side==0 ? visQ0[2] : visQ1[3]; + float ssaoValueR = side==0 ? visQ1[0] : visQ1[1]; + float ssaoValueB = side==0 ? visQ2[2] : visQ3[3]; + float ssaoValueTL = side==0 ? visQ0[3] : visQ0[2]; + float ssaoValueBR = side==0 ? visQ3[3] : visQ3[2]; + float ssaoValueTR = side==0 ? visQ1[3] : visQ1[2]; + float ssaoValueBL = side==0 ? visQ2[3] : visQ2[2]; + + float sumWeight = blurAmount; + float sum = ssaoValue * sumWeight; + + XeGTAO_AddSample(ssaoValueL, edgesC_LRTB[side].x, sum, sumWeight); + XeGTAO_AddSample(ssaoValueR, edgesC_LRTB[side].y, sum, sumWeight); + XeGTAO_AddSample(ssaoValueT, edgesC_LRTB[side].z, sum, sumWeight); + XeGTAO_AddSample(ssaoValueB, edgesC_LRTB[side].w, sum, sumWeight); + + XeGTAO_AddSample(ssaoValueTL, weightTL[side], sum, sumWeight); + XeGTAO_AddSample(ssaoValueTR, weightTR[side], sum, sumWeight); + XeGTAO_AddSample(ssaoValueBL, weightBL[side], sum, sumWeight); + XeGTAO_AddSample(ssaoValueBR, weightBR[side], sum, sumWeight); + + aoTerm[side] = sum / sumWeight; + + // use 1 instead of occ term scale for no-final + float outputValue = aoTerm[side] * (pushConstants.isFinalDenoisePass == 1 ? XE_GTAO_OCCLUSION_TERM_SCALE : 1); + + if (pushConstants.debug == 5){ + imageStore(debugImage, sideScreenPos, vec4(vec3(outputValue), 1.0f)); } } + + +// if (pushConstants.debug == 5){ +// imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); +// return; +// } +// +// if (pushConstants.debug == 5) { +// imageStore(debugImage, screenPos, vec4(edges.xyz, 1.0f)); +// return; +// if (screenPos.x % 2 == 0 && screenPos.y %2 == 0){ +// imageStore(debugImage, screenPos, vec4(vec3(1.0f), 1.0f)); +// } else { +// imageStore(debugImage, screenPos, vec4(vec3(0.5f), 1.0f)); +// } +// } } diff --git a/shaders/include/gtao.glsl b/shaders/include/gtao.glsl index c51b59b5..da3a9f5a 100644 --- a/shaders/include/gtao.glsl +++ b/shaders/include/gtao.glsl @@ -18,7 +18,8 @@ layout (push_constant) uniform PushConstants { float thinOccluderCompensation; float finalValuePower; float depthMipSamplingOffset; - uint noiseIndex; + int noiseIndex; + int isFinalDenoisePass; int debug; } pushConstants; diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 5d3f0369..d9afb53e 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -35,6 +35,7 @@ struct GTAOPushConstants float finalValuePower = 2.2f; float depthMipSamplingOffset = 3.30f; uint32_t noiseIndex{0}; + int32_t isFinalDenoisePass{1}; int32_t debug{0}; }; diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 3c7fbf0e..c0381ef2 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -531,7 +531,8 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, spatialFilteringPipelineLayout, 0, 2, indices.data(), offsets.data()); - const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / 16.0f)); + // each dispatch operates on 2x1 pixels + const auto x = static_cast(std::ceil(RENDER_EXTENT_WIDTH / (16.0f * 2.0f))); const auto y = static_cast(std::ceil(RENDER_EXTENT_HEIGHT / 16.0f)); vkCmdDispatch(cmd, x, y, 1); } From 7f3e2f8130eb5da2a3a8d011975e51ce3791863c Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 31 Mar 2025 16:32:35 +0700 Subject: [PATCH 23/27] Fix mesh visibility synchronization issues. --- shaders/visibility_pass.comp | 12 +++----- .../visibility_pass/visibility_pass.cpp | 3 ++ .../visibility_pass/visibility_pass.h | 2 ++ src/renderer/vk_helpers.cpp | 28 +++++++++++++++++++ src/renderer/vk_helpers.h | 3 ++ 5 files changed, 40 insertions(+), 8 deletions(-) diff --git a/shaders/visibility_pass.comp b/shaders/visibility_pass.comp index 03530ad2..56aecb1b 100644 --- a/shaders/visibility_pass.comp +++ b/shaders/visibility_pass.comp @@ -50,8 +50,8 @@ layout(set = 1, binding = 0) uniform bufferData } buffers; layout (push_constant) uniform PushConstants { - bool enabledFrustumCull; - bool shadowPass; + int enabledFrustumCull; + int shadowPass; } push; int checkIsVisible(mat4 mat, vec3 origin, float radius) @@ -86,12 +86,10 @@ void main() uint invocationId = gl_GlobalInvocationID.x; if (invocationId >= buffers.commandBufferCount) { return; } - - uint modelMatrixId = buffers.commandBuffer.commands[invocationId].firstInstance; Model models = buffers.modelMatrixBuffer.models[modelMatrixId]; - if (push.shadowPass){ + if (push.shadowPass == 1){ if (models.flags.y == 0){ buffers.commandBuffer.commands[invocationId].instanceCount = 0; return; @@ -109,9 +107,7 @@ void main() } - - - if (!push.enabledFrustumCull){ + if (push.enabledFrustumCull == 0){ buffers.commandBuffer.commands[invocationId].instanceCount = 1; return; } diff --git a/src/renderer/pipelines/visibility_pass/visibility_pass.cpp b/src/renderer/pipelines/visibility_pass/visibility_pass.cpp index 3b24b1f6..787f4955 100644 --- a/src/renderer/pipelines/visibility_pass/visibility_pass.cpp +++ b/src/renderer/pipelines/visibility_pass/visibility_pass.cpp @@ -75,11 +75,14 @@ void will_engine::visibility_pass::VisibilityPassPipeline::draw(VkCommandBuffer vkCmdSetDescriptorBufferOffsetsEXT(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, pipelineLayout, 1, 1, &addressesIndex, &addressesOffset); vkCmdDispatch(cmd, static_cast(std::ceil(static_cast(renderObject->getDrawIndirectCommandCount()) / 64.0f)), 1, 1); + + vk_helpers::synchronizeUniform(cmd, renderObject->getIndirectBuffer(drawInfo.currentFrameOverlap), VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT, VK_ACCESS_2_SHADER_WRITE_BIT, VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT, VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT); } vkCmdEndDebugUtilsLabelEXT(cmd); } + void will_engine::visibility_pass::VisibilityPassPipeline::createPipeline() { resourceManager.destroyPipeline(pipeline); diff --git a/src/renderer/pipelines/visibility_pass/visibility_pass.h b/src/renderer/pipelines/visibility_pass/visibility_pass.h index 1c7bf721..cb7695d6 100644 --- a/src/renderer/pipelines/visibility_pass/visibility_pass.h +++ b/src/renderer/pipelines/visibility_pass/visibility_pass.h @@ -39,6 +39,8 @@ class VisibilityPassPipeline void draw(VkCommandBuffer cmd, const VisibilityPassDrawInfo& drawInfo) const; + void indirectBufferSynchronize(); + void reloadShaders() { createPipeline(); } private: diff --git a/src/renderer/vk_helpers.cpp b/src/renderer/vk_helpers.cpp index f88ed89c..19ffca52 100644 --- a/src/renderer/vk_helpers.cpp +++ b/src/renderer/vk_helpers.cpp @@ -306,6 +306,34 @@ void will_engine::vk_helpers::transitionImage(VkCommandBuffer cmd, VkImage image vkCmdPipelineBarrier2(cmd, &depInfo); } +void will_engine::vk_helpers::synchronizeUniform(VkCommandBuffer cmd, const AllocatedBuffer& buffer, VkPipelineStageFlagBits2 srcPipelineStage, VkAccessFlagBits2 srcAccessBit , VkPipelineStageFlagBits2 dstPipelineStage, VkAccessFlagBits2 dstAccessBit) +{ + VkBufferMemoryBarrier2 bufferBarrier{}; + bufferBarrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER_2; + bufferBarrier.pNext = nullptr; + + bufferBarrier.srcStageMask = srcPipelineStage; + bufferBarrier.srcAccessMask = srcAccessBit; + + bufferBarrier.dstStageMask = dstPipelineStage; + bufferBarrier.dstAccessMask = dstAccessBit; + + bufferBarrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED; + bufferBarrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED; + bufferBarrier.buffer = buffer.buffer; + bufferBarrier.offset = 0; + bufferBarrier.size = VK_WHOLE_SIZE; + + VkDependencyInfo depInfo{}; + depInfo.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO; + depInfo.pNext = nullptr; + depInfo.dependencyFlags = 0; + depInfo.bufferMemoryBarrierCount = 1; + depInfo.pBufferMemoryBarriers = &bufferBarrier; + + vkCmdPipelineBarrier2(cmd, &depInfo); +} + void will_engine::vk_helpers::copyImageToImage(VkCommandBuffer cmd, VkImage source, VkImage destination, VkExtent2D srcSize, VkExtent2D dstSize) { VkImageBlit2 blitRegion{.sType = VK_STRUCTURE_TYPE_IMAGE_BLIT_2, .pNext = nullptr}; diff --git a/src/renderer/vk_helpers.h b/src/renderer/vk_helpers.h index 47eedfde..87c73eeb 100644 --- a/src/renderer/vk_helpers.h +++ b/src/renderer/vk_helpers.h @@ -87,6 +87,9 @@ namespace vk_helpers void transitionImage(VkCommandBuffer cmd, VkImage image, VkImageLayout currentLayout, VkImageAspectFlags aspectMask, VkImageLayout targetLayout); + void synchronizeUniform(VkCommandBuffer cmd, const AllocatedBuffer& buffer, VkPipelineStageFlagBits2 srcPipelineStage, VkAccessFlagBits2 srcAccessBit, VkPipelineStageFlagBits2 + dstPipelineStage, VkAccessFlagBits2 dstAccessBit); + void copyImageToImage(VkCommandBuffer cmd, VkImage source, VkImage destination, VkExtent2D srcSize, VkExtent2D dstSize); void copyDepthToDepth(VkCommandBuffer cmd, VkImage source, VkImage destination, VkExtent2D srcSize, VkExtent2D dstSize); From c48d17d558cbe47424f98770d1454bd4c7ef1404 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 31 Mar 2025 16:32:43 +0700 Subject: [PATCH 24/27] GTAO output into deferred resolve. --- .../ground_truth/gtao_main_pass.comp | 26 +++-- .../ground_truth/gtao_spatial_filter.comp | 30 +++--- shaders/deferredResolve.comp | 8 +- src/core/engine.cpp | 1 + src/core/engine.h | 2 +- .../components/mesh_renderer_component.cpp | 6 ++ .../ground_truth_ambient_occlusion.cpp | 94 +------------------ .../ground_truth_ambient_occlusion.h | 20 ++-- .../deferred_resolve/deferred_resolve.cpp | 6 ++ .../deferred_resolve/deferred_resolve.h | 1 + src/renderer/resource_manager.cpp | 13 +-- 11 files changed, 65 insertions(+), 142 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index f16bf251..8b75a447 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -93,23 +93,22 @@ vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rig return clamp((1.25 - edgesLRTB / (centerZ * 0.011)), 0, 1);; } -vec3 cheapReconstructViewSpacePosition(vec2 uv, const float viewspaceDepth, vec2 ndcToViewMul, vec2 ndcToViewAdd) +vec3 cheapReconstructViewSpacePosition(vec2 uv, float viewspaceDepth) { vec3 ret; - ret.xy = (ndcToViewMul * uv.xy + ndcToViewAdd) * viewspaceDepth; + ret.xy = (pushConstants.ndcToViewMul * uv.xy + pushConstants.ndcToViewAdd) * viewspaceDepth; ret.z = -viewspaceDepth; return ret; } -vec3 reconstructViewSpacePosition(vec2 uv, float viewDepth) { +vec4 reconstructViewSpacePosition(vec2 uv, float viewDepth) { float ndcDepth = pushConstants.depthLinearizeAdd - (pushConstants.depthLinearizeMult / viewDepth); - uv.y = 1 - uv.y; vec2 ndc = uv * 2.0 - 1.0; vec4 positionVS = sceneData.invProjection * vec4(ndc, ndcDepth, 1.0); positionVS /= positionVS.w; - return positionVS.xyz; + return positionVS; } void outputWorkingTerm(ivec2 screenPos, float visibility, vec3 bentNormal, image2D outputImage){ @@ -135,8 +134,6 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - //vec4 valuesUL = textureGather(prefilteredDepth, screenPos * sceneData.texelSize, 0); - //vec4 valuesBR = textureGatherOffset(prefilteredDepth, screenPos * sceneData.texelSize, ivec2(1,1), 0); vec4 valuesUL = textureGather(prefilteredDepth, uv, 0); vec4 valuesBR = textureGatherOffset(prefilteredDepth, uv, ivec2(1,1), 0); float viewSpaceZM = valuesUL.y; @@ -168,13 +165,14 @@ void main() { // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used viewSpaceZM = viewSpaceZM * 0.998f; - vec3 vPos = cheapReconstructViewSpacePosition(uv, viewSpaceZM, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); - //vec3 vPos = reconstructViewSpacePosition(uv, viewSpaceZM); + //vec3 vPos = cheapReconstructViewSpacePosition(uv, viewSpaceZM); + vec4 vPosAlt = reconstructViewSpacePosition(uv, viewSpaceZM); + vec3 vPos = vPosAlt.xyz; vec3 viewVec = normalize(-vPos); // debug world pos if (pushConstants.debug == 3){ - vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0)).xyz; + vec3 worldPos = (sceneData.invView * vPosAlt).xyz; imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); return; } @@ -305,13 +303,13 @@ void main() { vec2 sampleScreenPos0 = uv + sampleOffset; float SZ0 = textureLod(prefilteredDepth, sampleScreenPos0, mipLevel).r; - vec3 samplePos0 = cheapReconstructViewSpacePosition(sampleScreenPos0, SZ0, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); - //vec3 samplePos0 = reconstructViewSpacePosition(sampleScreenPos0, SZ0); + //vec3 samplePos0 = cheapReconstructViewSpacePosition(sampleScreenPos0, SZ0); + vec3 samplePos0 = reconstructViewSpacePosition(sampleScreenPos0, SZ0).xyz; vec2 sampleScreenPos1 = uv - sampleOffset; float SZ1 = textureLod(prefilteredDepth, sampleScreenPos1, mipLevel).r; - vec3 samplePos1 = cheapReconstructViewSpacePosition(sampleScreenPos1, SZ1, pushConstants.ndcToViewMul, pushConstants.ndcToViewAdd); - //vec3 samplePos1 = reconstructViewSpacePosition(sampleScreenPos1, SZ1); + //vec3 samplePos1 = cheapReconstructViewSpacePosition(sampleScreenPos1, SZ1); + vec3 samplePos1 = reconstructViewSpacePosition(sampleScreenPos1, SZ1).xyz; vec3 sampleDelta0 = (samplePos0 - vec3(vPos));// using lpfloat for sampleDelta causes precision issues vec3 sampleDelta1 = (samplePos1 - vec3(vPos));// using lpfloat for sampleDelta causes precision issues diff --git a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp index 4dcaf681..7ffe6fb6 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp @@ -115,21 +115,23 @@ void main() { if (pushConstants.debug == 5){ imageStore(debugImage, sideScreenPos, vec4(vec3(outputValue), 1.0f)); } + + imageStore(filteredAO, sideScreenPos, vec4(outputValue)); } -// if (pushConstants.debug == 5){ -// imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); -// return; -// } -// -// if (pushConstants.debug == 5) { -// imageStore(debugImage, screenPos, vec4(edges.xyz, 1.0f)); -// return; -// if (screenPos.x % 2 == 0 && screenPos.y %2 == 0){ -// imageStore(debugImage, screenPos, vec4(vec3(1.0f), 1.0f)); -// } else { -// imageStore(debugImage, screenPos, vec4(vec3(0.5f), 1.0f)); -// } -// } + // if (pushConstants.debug == 5){ + // imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); + // return; + // } + // + // if (pushConstants.debug == 5) { + // imageStore(debugImage, screenPos, vec4(edges.xyz, 1.0f)); + // return; + // if (screenPos.x % 2 == 0 && screenPos.y %2 == 0){ + // imageStore(debugImage, screenPos, vec4(vec3(1.0f), 1.0f)); + // } else { + // imageStore(debugImage, screenPos, vec4(vec3(0.5f), 1.0f)); + // } + // } } diff --git a/shaders/deferredResolve.comp b/shaders/deferredResolve.comp index e04d7ee6..e393bb23 100644 --- a/shaders/deferredResolve.comp +++ b/shaders/deferredResolve.comp @@ -16,8 +16,9 @@ layout (set = 1, binding = 1) uniform sampler2D albedoRenderTarget; layout (set = 1, binding = 2) uniform sampler2D pbrRenderTarget; layout (set = 1, binding = 3) uniform sampler2D depthBuffer; layout (set = 1, binding = 4) uniform sampler2D velocityBuffer; // velocity buffer is not actually used in this deferred resolve +layout (set = 1, binding = 5) uniform sampler2D aoBuffer; -layout (rgba16f, set = 1, binding = 5) uniform image2D outputImage; +layout (rgba16f, set = 1, binding = 6) uniform image2D outputImage; layout (set = 2, binding = 0) uniform samplerCube environmentDiffuseAndSpecular; layout (set = 2, binding = 1) uniform sampler2D lut; @@ -186,9 +187,8 @@ void main() { imageStore(outputImage, screenPos, vec4(vec3(dot(N, L)), 1.0)); break; case 9: - // vec4 viewPos = sceneData.view * vec4(position, 1.0f); - // imageStore(outputImage, screenPos, vec4(-viewPos.xyz / 100.0f, 1.0f)); - imageStore(outputImage, screenPos, vec4(position.xyz / 1000.0f, 1.0f)); + float ao = texture(aoBuffer, uv).r; + imageStore(outputImage, screenPos, vec4(vec3(ao), 1.0f)); break; } } \ No newline at end of file diff --git a/src/core/engine.cpp b/src/core/engine.cpp index c0137938..65288151 100644 --- a/src/core/engine.cpp +++ b/src/core/engine.cpp @@ -216,6 +216,7 @@ void Engine::initRenderer() pbrRenderTarget.imageView, depthImage.imageView, velocityRenderTarget.imageView, + ambientOcclusionPipeline->getAmbientOcclusionRenderTarget().imageView, drawImage.imageView, resourceManager->getDefaultSamplerLinear() }; diff --git a/src/core/engine.h b/src/core/engine.h index da6d9894..299ad7a3 100644 --- a/src/core/engine.h +++ b/src/core/engine.h @@ -175,7 +175,7 @@ class Engine bool bEnableDebugFrustumCullDraw{false}; int32_t csmPcf{1}; int32_t deferredDebug{0}; - int32_t gtaoDebug{5}; + int32_t gtaoDebug{4}; bool bDrawTerrainLines{false}; void hotReloadShaders() const; diff --git a/src/core/game_object/components/mesh_renderer_component.cpp b/src/core/game_object/components/mesh_renderer_component.cpp index a4d89616..997a3dde 100644 --- a/src/core/game_object/components/mesh_renderer_component.cpp +++ b/src/core/game_object/components/mesh_renderer_component.cpp @@ -100,8 +100,14 @@ void MeshRendererComponent::updateRenderImgui() } ImGui::Separator(); + const bool originalVis = bIsVisible; + const bool originalShadow = bIsShadowCaster; ImGui::Checkbox("Visible", &bIsVisible); ImGui::Checkbox("Cast Shadows", &bIsShadowCaster); + + if (bIsVisible != originalVis || bIsShadowCaster != originalShadow) { + dirty(); + } } } } diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index c0381ef2..26ad79fd 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -218,56 +218,7 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::GroundTruth usage |= VK_IMAGE_USAGE_SAMPLED_BIT; VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); - spatialFilteringImage = resourceManager.createImage(imgInfo); - } - - // Temporal Accumulation - { - DescriptorLayoutBuilder layoutBuilder; - layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // filtered ao - layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // final output history - layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT velocity buffer - layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // MRT depth buffer - layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // final output - - temporalAccumulationSetLayout = resourceManager.createDescriptorSetLayout(layoutBuilder, VK_SHADER_STAGE_COMPUTE_BIT, - VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); - - VkPushConstantRange pushConstants{}; - pushConstants.offset = 0; - pushConstants.size = sizeof(GTAOPushConstants); - pushConstants.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT; - - VkDescriptorSetLayout setLayouts[2]; - setLayouts[0] = resourceManager.getSceneDataLayout(); - setLayouts[1] = temporalAccumulationSetLayout; - - VkPipelineLayoutCreateInfo layoutInfo{}; - layoutInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO; - layoutInfo.pNext = nullptr; - layoutInfo.pSetLayouts = setLayouts; - layoutInfo.setLayoutCount = 2; - layoutInfo.pPushConstantRanges = &pushConstants; - layoutInfo.pushConstantRangeCount = 1; - - temporalAccumulationPipelineLayout = resourceManager.createPipelineLayout(layoutInfo); - createTemporalAccumulationPipeline(); - - temporalAccumulationDescriptorBuffer = resourceManager.createDescriptorBufferSampler(temporalAccumulationSetLayout, 1); - - - VkImageUsageFlags usage{}; - usage |= VK_IMAGE_USAGE_TRANSFER_DST_BIT; - usage |= VK_IMAGE_USAGE_SAMPLED_BIT; - VkImageCreateInfo imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); - historyOutputImage = resourceManager.createImage(imgInfo); - - usage = {}; - usage |= VK_IMAGE_USAGE_TRANSFER_SRC_BIT; - usage |= VK_IMAGE_USAGE_STORAGE_BIT; - usage |= VK_IMAGE_USAGE_SAMPLED_BIT; - imgInfo = vk_helpers::imageCreateInfo(ambientOcclusionFormat, usage, {RENDER_EXTENTS.width, RENDER_EXTENTS.height, 1}); - ambientOcclusionOutputImage = resourceManager.createImage(imgInfo); + denoisedFinalAO = resourceManager.createImage(imgInfo); } } @@ -307,20 +258,9 @@ will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::~GroundTrut resourceManager.destroyPipelineLayout(spatialFilteringPipelineLayout); resourceManager.destroyPipeline(spatialFilteringPipeline); - resourceManager.destroyImage(spatialFilteringImage); + resourceManager.destroyImage(denoisedFinalAO); resourceManager.destroyDescriptorBuffer(spatialFilteringDescriptorBuffer); - - - // Temporal Accumulation Resources - resourceManager.destroyDescriptorSetLayout(temporalAccumulationSetLayout); - resourceManager.destroyPipelineLayout(temporalAccumulationPipelineLayout); - resourceManager.destroyPipeline(temporalAccumulationPipeline); - - resourceManager.destroyImage(historyOutputImage); - resourceManager.destroyImage(ambientOcclusionOutputImage); - - resourceManager.destroyDescriptorBuffer(temporalAccumulationDescriptorBuffer); } void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupDepthPrefilterDescriptorBuffer(const VkImageView& depthImageView) @@ -424,7 +364,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupS }); imageDescriptors.push_back({ VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, - {VK_NULL_HANDLE, spatialFilteringImage.imageView, VK_IMAGE_LAYOUT_GENERAL}, + {VK_NULL_HANDLE, denoisedFinalAO.imageView, VK_IMAGE_LAYOUT_GENERAL}, false }); imageDescriptors.push_back({ @@ -514,7 +454,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vk_helpers::transitionImage(cmd, ambientOcclusionImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); - vk_helpers::transitionImage(cmd, spatialFilteringImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, + vk_helpers::transitionImage(cmd, denoisedFinalAO.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); // Spatial Filtering { @@ -537,7 +477,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V vkCmdDispatch(cmd, x, y, 1); } - vk_helpers::transitionImage(cmd, spatialFilteringImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + vk_helpers::transitionImage(cmd, denoisedFinalAO.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, VK_IMAGE_ASPECT_COLOR_BIT); @@ -554,7 +494,6 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::reload createDepthPrefilterPipeline(); createAmbientOcclusionPipeline(); createSpatialFilteringPipeline(); - createTemporalAccumulationPipeline(); } void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createDepthPrefilterPipeline() @@ -625,26 +564,3 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::create spatialFilteringPipeline = resourceManager.createComputePipeline(pipelineInfo); resourceManager.destroyShaderModule(computeShader); } - -void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::createTemporalAccumulationPipeline() -{ - resourceManager.destroyPipeline(temporalAccumulationPipeline); - VkShaderModule computeShader = resourceManager.createShaderModule("shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp"); - - VkPipelineShaderStageCreateInfo stageInfo{}; - stageInfo.sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO; - stageInfo.pNext = nullptr; - stageInfo.stage = VK_SHADER_STAGE_COMPUTE_BIT; - stageInfo.module = computeShader; - stageInfo.pName = "main"; - - VkComputePipelineCreateInfo pipelineInfo{}; - pipelineInfo.sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO; - pipelineInfo.pNext = nullptr; - pipelineInfo.layout = temporalAccumulationPipelineLayout; - pipelineInfo.stage = stageInfo; - pipelineInfo.flags = VK_PIPELINE_CREATE_DESCRIPTOR_BUFFER_BIT_EXT; - - temporalAccumulationPipeline = resourceManager.createComputePipeline(pipelineInfo); - resourceManager.destroyShaderModule(computeShader); -} diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index ffefe4c8..5964b1f3 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -30,11 +30,15 @@ class GroundTruthAmbientOcclusionPipeline void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const; void reloadShaders(); + + AllocatedImage getAmbientOcclusionRenderTarget() const { return denoisedFinalAO; } + private: void createDepthPrefilterPipeline(); + void createAmbientOcclusionPipeline(); + void createSpatialFilteringPipeline(); - void createTemporalAccumulationPipeline(); private: // Depth Pre-filter VkDescriptorSetLayout depthPrefilterSetLayout{VK_NULL_HANDLE}; @@ -72,22 +76,10 @@ class GroundTruthAmbientOcclusionPipeline VkPipelineLayout spatialFilteringPipelineLayout{VK_NULL_HANDLE}; VkPipeline spatialFilteringPipeline{VK_NULL_HANDLE}; - AllocatedImage spatialFilteringImage{VK_NULL_HANDLE}; + AllocatedImage denoisedFinalAO{VK_NULL_HANDLE}; DescriptorBufferSampler spatialFilteringDescriptorBuffer; -private: // Temporal Accumulation - VkDescriptorSetLayout temporalAccumulationSetLayout{VK_NULL_HANDLE}; - VkPipelineLayout temporalAccumulationPipelineLayout{VK_NULL_HANDLE}; - VkPipeline temporalAccumulationPipeline{VK_NULL_HANDLE}; - - AllocatedImage historyOutputImage{VK_NULL_HANDLE}; - - DescriptorBufferSampler temporalAccumulationDescriptorBuffer; - -private: // Output - AllocatedImage ambientOcclusionOutputImage{VK_NULL_HANDLE}; - private: // Debug VkFormat debugFormat{VK_FORMAT_R8G8B8A8_UNORM}; AllocatedImage debugImage{VK_NULL_HANDLE}; diff --git a/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp b/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp index dc6f3c96..faea01fe 100644 --- a/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp +++ b/src/renderer/pipelines/deferred_resolve/deferred_resolve.cpp @@ -80,6 +80,11 @@ void will_engine::deferred_resolve::DeferredResolvePipeline::setupDescriptorBuff velocityTarget.imageView = drawInfo.velocityTarget; velocityTarget.imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + VkDescriptorImageInfo aoTarget = {}; + aoTarget.sampler = drawInfo.sampler; + aoTarget.imageView = drawInfo.aoTarget; + aoTarget.imageLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL; + VkDescriptorImageInfo drawImageTarget = {}; drawImageTarget.imageView = drawInfo.outputTarget; drawImageTarget.imageLayout = VK_IMAGE_LAYOUT_GENERAL; @@ -89,6 +94,7 @@ void will_engine::deferred_resolve::DeferredResolvePipeline::setupDescriptorBuff renderTargetDescriptors.push_back({VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, pbrDataTarget, false}); renderTargetDescriptors.push_back({VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, depthImageTarget, false}); renderTargetDescriptors.push_back({VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, velocityTarget, false}); + renderTargetDescriptors.push_back({VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER, aoTarget, false}); renderTargetDescriptors.push_back({VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, drawImageTarget, false}); resourceManager.setupDescriptorBufferSampler(resolveDescriptorBuffer, renderTargetDescriptors, 0); diff --git a/src/renderer/pipelines/deferred_resolve/deferred_resolve.h b/src/renderer/pipelines/deferred_resolve/deferred_resolve.h index ce817a49..ada68554 100644 --- a/src/renderer/pipelines/deferred_resolve/deferred_resolve.h +++ b/src/renderer/pipelines/deferred_resolve/deferred_resolve.h @@ -19,6 +19,7 @@ struct DeferredResolveDescriptor VkImageView pbrTarget; VkImageView depthTarget; VkImageView velocityTarget; + VkImageView aoTarget; VkImageView outputTarget; VkSampler sampler; diff --git a/src/renderer/resource_manager.cpp b/src/renderer/resource_manager.cpp index de3c8bfc..47a072d1 100644 --- a/src/renderer/resource_manager.cpp +++ b/src/renderer/resource_manager.cpp @@ -113,12 +113,13 @@ will_engine::ResourceManager::ResourceManager(const VulkanContext& context, Imme // Render Targets { DescriptorLayoutBuilder layoutBuilder; - layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); - layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); - layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); - layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); - layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); - layoutBuilder.addBinding(5, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); + layoutBuilder.addBinding(0, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // Normals + layoutBuilder.addBinding(1, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // Albedo + layoutBuilder.addBinding(2, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // PBR + layoutBuilder.addBinding(3, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // Depth + layoutBuilder.addBinding(4, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // Velocity + layoutBuilder.addBinding(5, VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER); // AO + layoutBuilder.addBinding(6, VK_DESCRIPTOR_TYPE_STORAGE_IMAGE); // Output renderTargetsLayout = layoutBuilder.build(context.device, VK_SHADER_STAGE_COMPUTE_BIT, nullptr, VK_DESCRIPTOR_SET_LAYOUT_CREATE_DESCRIPTOR_BUFFER_BIT_EXT); } From 5cbe33915427bd5981ff26602565f3c7f171ac3c Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 31 Mar 2025 17:28:54 +0700 Subject: [PATCH 25/27] Clean up minor GTAO texture sampling issues. --- .../ground_truth/gtao_depth_prefilter.comp | 14 +++++--------- .../ground_truth/gtao_main_pass.comp | 17 ++++------------- .../ground_truth/gtao_spatial_filter.comp | 17 ++++++++--------- shaders/deferredResolve.comp | 7 ++++--- 4 files changed, 21 insertions(+), 34 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index f41ea7d2..ab682687 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -67,15 +67,11 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; // todo: optimize with textureGather? - vec4 depths = textureGatherOffset(depthImage, uv, ivec2(1,1), 0); - float rDepth0 = depths.w; // top-left - float rDepth1 = depths.z; // top-right - float rDepth2 = depths.x; // bottom-left - float rDepth3 = depths.y; // bottom-right -// float rDepth0 = texture(depthImage, uv + vec2(0.0, 0.0) * sceneData.texelSize).r; -// float rDepth1 = texture(depthImage, uv + vec2(1.0, 0.0) * sceneData.texelSize).r; -// float rDepth2 = texture(depthImage, uv + vec2(0.0, 1.0) * sceneData.texelSize).r; -// float rDepth3 = texture(depthImage, uv + vec2(1.0, 1.0) * sceneData.texelSize).r; + vec4 depths = textureGatherOffset(depthImage, vec2(screenPos) * sceneData.texelSize, ivec2(1,1), 0); + float rDepth0 = depths.w; + float rDepth1 = depths.z; + float rDepth2 = depths.x; + float rDepth3 = depths.y; float depth0 = clampDepth(screenToViewSpaceDepth(rDepth0, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); float depth1 = clampDepth(screenToViewSpaceDepth(rDepth1, pushConstants.depthLinearizeMult, pushConstants.depthLinearizeAdd)); diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index 8b75a447..d721fb5c 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -134,8 +134,9 @@ void main() { vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - vec4 valuesUL = textureGather(prefilteredDepth, uv, 0); - vec4 valuesBR = textureGatherOffset(prefilteredDepth, uv, ivec2(1,1), 0); + vec2 gatherCenter = vec2(screenPos) * sceneData.texelSize; + vec4 valuesUL = textureGather(prefilteredDepth, gatherCenter, 0); + vec4 valuesBR = textureGatherOffset(prefilteredDepth, gatherCenter, ivec2(1, 1), 0); float viewSpaceZM = valuesUL.y; const float viewSpaceZL = valuesUL.x; const float viewSpaceZR = valuesUL.z; @@ -146,12 +147,6 @@ void main() { float packedEdges = XeGTAO_PackEdges(edges); imageStore(edgeDataOutput, screenPos, vec4(packedEdges)); - float minEdge = min(min(edges.x, edges.y), min(edges.z, edges.w)); - - if (pushConstants.debug == 2){ - imageStore(debugImage, screenPos, vec4(vec3(minEdge), 1.0f)); - return; - } // Get view space normal by sampling normal buffer and converting from world to view (code not relevant) vec3 worldNormal = texture(normalBuffer, uv).rgb; @@ -159,7 +154,6 @@ void main() { if (pushConstants.debug == 2){ imageStore(debugImage, screenPos, vec4(viewNormal, 1.0f)); - return; } // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used @@ -174,12 +168,11 @@ void main() { if (pushConstants.debug == 3){ vec3 worldPos = (sceneData.invView * vPosAlt).xyz; imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); - return; } // Per Intel // prevents normals that are facing away from the view vector - xeGTAO struggles with extreme cases, but in Vanilla it seems rare so it's disabled by default - viewNormal = normalize(viewNormal + max(0, -dot(viewNormal, viewVec)) * viewVec); + viewNormal = normalize(viewNormal + max(0, -dot(viewNormal, viewVec)) * viewVec); const float effectRadius = pushConstants.effectRadius * pushConstants.radiusMultiplier; @@ -215,7 +208,6 @@ void main() { visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); if (pushConstants.debug == 4){ imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); - return; } imageStore(aoOutput, screenPos, vec4(visibility)); @@ -374,7 +366,6 @@ void main() { visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); if (pushConstants.debug == 4){ imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); - return; } imageStore(aoOutput, screenPos, vec4(visibility)); diff --git a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp index 7ffe6fb6..63fc4cef 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp @@ -31,8 +31,6 @@ void main() { return; } - vec2 uv = (vec2(screenPos) + 0.5) * sceneData.texelSize; - // if final apply use below (see, using multiple denoise passes) float blurAmount = pushConstants.isFinalDenoisePass == 1 ? pushConstants.denoiseBlurBeta : pushConstants.denoiseBlurBeta / 5.0f; @@ -46,15 +44,16 @@ void main() { float weightBR[2]; + vec2 gatherCenter = vec2(screenPos) * sceneData.texelSize; - vec4 edgesQ0 = textureGatherOffset(edgeData, uv, ivec2(0, 0), 0); - vec4 edgesQ1 = textureGatherOffset(edgeData, uv, ivec2(2, 0), 0); - vec4 edgesQ2 = textureGatherOffset(edgeData, uv, ivec2(1, 2), 0); + vec4 edgesQ0 = textureGatherOffset(edgeData, gatherCenter, ivec2(0, 0), 0); + vec4 edgesQ1 = textureGatherOffset(edgeData, gatherCenter, ivec2(2, 0), 0); + vec4 edgesQ2 = textureGatherOffset(edgeData, gatherCenter, ivec2(1, 2), 0); - vec4 visQ0 = textureGatherOffset(rawAO, uv, ivec2(0, 0), 0); - vec4 visQ1 = textureGatherOffset(rawAO, uv, ivec2(2, 0), 0); - vec4 visQ2 = textureGatherOffset(rawAO, uv, ivec2(0, 2), 0); - vec4 visQ3 = textureGatherOffset(rawAO, uv, ivec2(2, 2), 0); + vec4 visQ0 = textureGatherOffset(rawAO, gatherCenter, ivec2(0, 0), 0); + vec4 visQ1 = textureGatherOffset(rawAO, gatherCenter, ivec2(2, 0), 0); + vec4 visQ2 = textureGatherOffset(rawAO, gatherCenter, ivec2(0, 2), 0); + vec4 visQ3 = textureGatherOffset(rawAO, gatherCenter, ivec2(2, 2), 0); for (int side = 0; side < 2; side++) { diff --git a/shaders/deferredResolve.comp b/shaders/deferredResolve.comp index e393bb23..3a26434f 100644 --- a/shaders/deferredResolve.comp +++ b/shaders/deferredResolve.comp @@ -138,10 +138,11 @@ void main() { vec3 reflectionSpecular = SpecularReflection(environmentDiffuseAndSpecular, lut, V, N, roughness, F) * indirectAttenuation; - vec3 ambient = (kD * reflectionDiffuse + reflectionSpecular); + //vec3 ambient = (kD * reflectionDiffuse + reflectionSpecular); + float ao = texture(aoBuffer, uv).r; + vec3 ambient = (kD * reflectionDiffuse + reflectionSpecular) * ao; - - vec3 finalColor = (diffuse + specular) * vec3(1.0f) * nDotL * shadowFactor; + vec3 finalColor = (diffuse + specular) * nDotL * shadowFactor; finalColor += ambient; imageStore(outputImage, screenPos, vec4(finalColor, albedo.w)); From 2b9d1c7b373c09a12b10c3222effbce03c47a4b1 Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 31 Mar 2025 18:50:15 +0700 Subject: [PATCH 26/27] GTAo debug improvements. --- .../ground_truth/gtao_depth_prefilter.comp | 5 +- .../ground_truth/gtao_main_pass.comp | 45 ++-- .../ground_truth/gtao_spatial_filter.comp | 6 + .../gtao_temporal_accumulation.comp | 15 -- shaders/include/gtao.glsl | 3 + src/renderer/imgui_wrapper.cpp | 214 ++++++++++++++---- src/renderer/imgui_wrapper.h | 1 + .../ambient_occlusion_types.h | 24 +- .../ground_truth_ambient_occlusion.cpp | 31 ++- .../ground_truth_ambient_occlusion.h | 4 +- .../visibility_pass/visibility_pass.h | 2 - 11 files changed, 232 insertions(+), 118 deletions(-) delete mode 100644 shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp diff --git a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp index ab682687..0283c028 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_depth_prefilter.comp @@ -54,6 +54,10 @@ float depthMipFilter(float depth0, float depth1, float depth2, float depth3, flo } void main() { + if (pushConstants.debug == -1){ + return; + } + ivec2 groupThreadID = ivec2(gl_LocalInvocationID.xy); // MIP 0 @@ -88,7 +92,6 @@ void main() { imageStore(debugImage, screenPos + ivec2(1, 0), vec4(vec3(depth1 / 1000.0f), 1.0f)); imageStore(debugImage, screenPos + ivec2(0, 1), vec4(vec3(depth2 / 1000.0f), 1.0f)); imageStore(debugImage, screenPos + ivec2(1, 1), vec4(vec3(depth3 / 1000.0f), 1.0f)); - return; } // MIP 1 diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index d721fb5c..fa3d99e7 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -21,16 +21,6 @@ layout (rgba8, set = 1, binding = 4) uniform image2D debugImage; #define XE_GTAO_PI (3.1415926535897932384626433832795) #define XE_GTAO_PI_HALF (1.5707963267948966192313216916398) -#define XE_GTAO_SLICE_COUNT_LOW 1.0f -#define XE_GTAO_SLICE_COUNT_MEDIUM 2.0f -#define XE_GTAO_SLICE_COUNT_HIGH 3.0f -#define XE_GTAO_SLICE_COUNT_ULTRA 9.0f - -#define XE_GTAO_STEPS_PER_SLICE_COUNT_LOW 2.0f -#define XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM 2.0f -#define XE_GTAO_STEPS_PER_SLICE_COUNT_HIGH 3.0f -#define XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA 3.0f - #define XE_GTAO_OCCLUSION_TERM_SCALE 1.5f // http://h14s.p5r.org/2012/09/0x5f3759df.html, [Drobot2014a] Low Level Optimizations for GCN, https://blog.selfshadow.com/publications/s2016-shading-course/activision/s2016_pbs_activision_occlusion.pdf slide 63 @@ -50,7 +40,7 @@ float XeGTAO_FastACos(float inX) return (inX >= 0) ? res : PI - res; } -uint hilbertIndex(uint posX, uint posY) +uint XE_GTAO_HilbertIndex(uint posX, uint posY) { uint index = 0u; for (uint curLevel = XE_HILBERT_WIDTH/2u; curLevel > 0u; curLevel /= 2u) @@ -74,14 +64,14 @@ uint hilbertIndex(uint posX, uint posY) return index; } -vec2 spatioTemporalNoise(ivec2 pixCoord, uint temporalIndex)// without TAA, temporalIndex is always 0 +vec2 XE_GTAO_SpatioTemporalNoise(ivec2 pixCoord, uint temporalIndex)// without TAA, temporalIndex is always 0 { - uint index = hilbertIndex(uint(pixCoord.x), uint(pixCoord.y)); + uint index = XE_GTAO_HilbertIndex(uint(pixCoord.x), uint(pixCoord.y)); index += 288u * (temporalIndex % 64u); return vec2(fract(0.5 + index * vec2(0.75487766624669276005, 0.5698402909980532659114))); } -vec4 calculateDepthEdges(const float centerZ, const float leftZ, const float rightZ, const float topZ, const float bottomZ) +vec4 XE_GTAO_CalculateDepthEdges(const float centerZ, const float leftZ, const float rightZ, const float topZ, const float bottomZ) { vec4 edgesLRTB = vec4(leftZ, rightZ, topZ, bottomZ) - vec4(centerZ); @@ -97,6 +87,7 @@ vec3 cheapReconstructViewSpacePosition(vec2 uv, float viewspaceDepth) { vec3 ret; ret.xy = (pushConstants.ndcToViewMul * uv.xy + pushConstants.ndcToViewAdd) * viewspaceDepth; + //ret.y = -ret.y; ret.z = -viewspaceDepth; return ret; } @@ -111,11 +102,6 @@ vec4 reconstructViewSpacePosition(vec2 uv, float viewDepth) { return positionVS; } -void outputWorkingTerm(ivec2 screenPos, float visibility, vec3 bentNormal, image2D outputImage){ - visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); - imageStore(outputImage, screenPos, vec4(visibility + 0.5f / 255.0f)); -} - mat3 adjugate(mat4 m) { return mat3( cross(m[1].xyz, m[2].xyz), @@ -126,6 +112,10 @@ mat3 adjugate(mat4 m) { } void main() { + if (pushConstants.debug == -1){ + return; + } + const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); if (screenPos.x > sceneData.renderTargetSize.x || screenPos.y > sceneData.renderTargetSize.y) { @@ -143,7 +133,7 @@ void main() { const float viewSpaceZT = valuesBR.z; const float viewSpaceZB = valuesBR.x; - vec4 edges = calculateDepthEdges(viewSpaceZM, viewSpaceZL, viewSpaceZR, viewSpaceZT, viewSpaceZB); + vec4 edges = XE_GTAO_CalculateDepthEdges(viewSpaceZM, viewSpaceZL, viewSpaceZR, viewSpaceZT, viewSpaceZB); float packedEdges = XeGTAO_PackEdges(edges); imageStore(edgeDataOutput, screenPos, vec4(packedEdges)); @@ -191,7 +181,7 @@ void main() { { // NOISE - vec2 noise = spatioTemporalNoise(screenPos, pushConstants.noiseIndex); + vec2 noise = XE_GTAO_SpatioTemporalNoise(screenPos, pushConstants.noiseIndex); float noiseSlice = noise.x; float noiseSample = noise.y; @@ -211,15 +201,14 @@ void main() { } imageStore(aoOutput, screenPos, vec4(visibility)); - // todo: look at how this will return 1/1.5 instead of 1 for visibility always? - // todo: (bent normal) need to write `viewNormal` value to the buffer (i.e. no change to trajectory of normal) + // (Bent Normals) need to write `viewNormal` value to the buffer (i.e. no change to trajectory of normal) return; } const float minS = pixelTooCloseThreshold / screenspaceRadius; - float sliceCount = XE_GTAO_SLICE_COUNT_ULTRA; - float stepsPerSlice = XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA; + float sliceCount = pushConstants.sliceCount; + float stepsPerSlice = pushConstants.stepsPerSliceCount; for (float slice = 0; slice < sliceCount; slice++){ float sliceK = (slice+noiseSlice) / sliceCount; @@ -345,7 +334,7 @@ void main() { float localVisibility = projectedNormalVecLength * (iarc0+iarc1); visibility += localVisibility; - // todo: uncomment and fix if outputting bent normals + // (Bent Normals) // see "Algorithm 2 Extension that computes bent normals b." // lpfloat t0 = (6*sin(h0-n)-sin(3*h0-n)+6*sin(h1-n)-sin(3*h1-n)+16*sin(n)-3*(sin(h0+n)+sin(h1+n)))/12; // lpfloat t1 = (-cos(3 * h0-n)-cos(3 * h1-n) +8 * cos(n)-3 * (cos(h0+n) +cos(h1+n)))/12; @@ -358,11 +347,11 @@ void main() { visibility = pow(visibility, pushConstants.finalValuePower); visibility = max(0.03, visibility);// disallow total occlusion (which wouldn't make any sense anyhow since pixel is visible but also helps with packing bent normals) - // todo (bent normal) + // (Bent Normals) // bentNormal = normalize(bentNormal) ; } - // todo (bent normal) + // (Bent Normals) visibility = clamp(visibility / XE_GTAO_OCCLUSION_TERM_SCALE, 0, 1); if (pushConstants.debug == 4){ imageStore(debugImage, screenPos, vec4(vec3(visibility), 1.0f)); diff --git a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp index 63fc4cef..f077594f 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_spatial_filter.comp @@ -27,6 +27,12 @@ void main() { //const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy); const ivec2 screenPos = ivec2(gl_GlobalInvocationID.xy) * ivec2(2, 1); + if (pushConstants.debug == -1){ + imageStore(filteredAO, screenPos, vec4(1.0f)); + imageStore(filteredAO, screenPos + ivec2(1, 0), vec4(1.0f)); + return; + } + if (screenPos.x > sceneData.renderTargetSize.x || screenPos.y > sceneData.renderTargetSize.y) { return; } diff --git a/shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp b/shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp deleted file mode 100644 index bf32d9ad..00000000 --- a/shaders/ambient_occlusion/ground_truth/gtao_temporal_accumulation.comp +++ /dev/null @@ -1,15 +0,0 @@ -#version 460 - -#include "scene.glsl" - -// layout (std140, set = 0, binding = 0) uniform SceneData - scene.glsl - -layout (set = 1, binding = 0) uniform sampler2D filteredAO; -layout (set = 1, binding = 1) uniform sampler2D historyOutputAO; -layout (set = 1, binding = 2) uniform sampler2D depthBuffer; -layout (set = 1, binding = 2) uniform sampler2D velocityBuffer; -layout (r8, set = 1, binding = 2) uniform image2D outputAO; - -void main() { - -} diff --git a/shaders/include/gtao.glsl b/shaders/include/gtao.glsl index da3a9f5a..5fd64f10 100644 --- a/shaders/include/gtao.glsl +++ b/shaders/include/gtao.glsl @@ -21,6 +21,9 @@ layout (push_constant) uniform PushConstants { int noiseIndex; int isFinalDenoisePass; + float sliceCount; + float stepsPerSliceCount; + int debug; } pushConstants; diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index 603fd10d..aea583b3 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -192,30 +192,187 @@ void ImguiWrapper::imguiInterface(Engine* engine) if (ImGui::Begin("Renderer")) { if (ImGui::BeginTabBar("RendererTabs")) { - if (ImGui::BeginTabItem("Shaders")) { - ImGui::Text("Shaders"); + if (ImGui::BeginTabItem("Debugging")) { ImGui::SetNextItemWidth(75.0f); if (ImGui::Button("Hot-Reload Shaders")) { engine->hotReloadShaders(); } ImGui::Separator(); + ImGui::Text("Temporal Anti-Aliasing"); ImGui::Checkbox("Enable TAA", &engine->bEnableTaa); ImGui::DragFloat("Taa Blend Value", &engine->taaBlendValue, 0.01, 0.1f, 0.5f); - ImGui::EndTabItem(); - } + ImGui::Separator(); - if (ImGui::BeginTabItem("Pipelines")) { ImGui::Text("Deferred Debug"); const char* deferredDebugOptions[]{"None", "Depth", "Velocity", "Albedo", "Normal", "PBR", "Shadows", "Cascade Level", "nDotL", "AO"}; ImGui::Combo("Deferred Debug", &engine->deferredDebug, deferredDebugOptions, IM_ARRAYSIZE(deferredDebugOptions)); + ImGui::Separator(); + ImGui::Text("Frustum Cull Debug Draw"); + ImGui::Checkbox("Enable Frustum Cull Debug Draw", &engine->bEnableDebugFrustumCullDraw); ImGui::EndTabItem(); } - if (ImGui::BeginTabItem("Debug View")) { - ImGui::Text("Frustum Cull Debug Draw"); - ImGui::Checkbox("Enable Frustum Cull Debug Draw", &engine->bEnableDebugFrustumCullDraw); + if (ImGui::BeginTabItem("Ambient Occlusion")) { + ambient_occlusion::GTAOPushConstants& gtao = engine->ambientOcclusionPipeline->gtaoPush; + if (ImGui::CollapsingHeader("GTAO Settings")) { + ImGui::Text("Effect Parameters"); + ImGui::Separator(); + ImGui::SliderFloat("Effect Radius", >ao.effectRadius, 0.1f, 2.0f); + ImGui::SliderFloat("Effect Falloff Range", >ao.effectFalloffRange, 0.0f, 1.0f); + + ImGui::Spacing(); + ImGui::Text("Denoise Parameters"); + ImGui::Separator(); + float blurBeta = gtao.denoiseBlurBeta; + if (ImGui::SliderFloat("Denoise Blur Beta", &blurBeta, 0.0f, 5.0f)) { + if (ambient_occlusion::GTAO_DENOISE_PASSES != 0) { + gtao.denoiseBlurBeta = blurBeta; + } + } + ImGui::Checkbox("Final Denoise Pass", (bool*) >ao.isFinalDenoisePass); + + ImGui::Spacing(); + ImGui::Text("Sampling Parameters"); + ImGui::Separator(); + ImGui::SliderFloat("Radius Multiplier", >ao.radiusMultiplier, 0.1f, 3.0f); + ImGui::SliderFloat("Sample Distribution Power", >ao.sampleDistributionPower, 1.0f, 4.0f); + ImGui::SliderFloat("Thin Occluder Compensation", >ao.thinOccluderCompensation, 0.0f, 1.0f); + ImGui::SliderFloat("Final Value Power", >ao.finalValuePower, 1.0f, 4.0f); + ImGui::SliderFloat("Depth Mip Sampling Offset", >ao.depthMipSamplingOffset, 0.0f, 5.0f); + + ImGui::Spacing(); + ImGui::Text("Sample Count"); + ImGui::Separator(); + const char* qualityPresets[] = {"Low", "Medium", "High", "Ultra"}; + int slicePreset = 0; + + if (gtao.sliceCount == ambient_occlusion::XE_GTAO_SLICE_COUNT_LOW) slicePreset = 0; + else if (gtao.sliceCount == ambient_occlusion::XE_GTAO_SLICE_COUNT_MEDIUM) slicePreset = 1; + else if (gtao.sliceCount == ambient_occlusion::XE_GTAO_SLICE_COUNT_HIGH) slicePreset = 2; + else slicePreset = 3; + + if (ImGui::Combo("Slice Count Preset", &slicePreset, qualityPresets, IM_ARRAYSIZE(qualityPresets))) { + switch (slicePreset) { + case 0: gtao.sliceCount = ambient_occlusion::XE_GTAO_SLICE_COUNT_LOW; + break; + case 1: gtao.sliceCount = ambient_occlusion::XE_GTAO_SLICE_COUNT_MEDIUM; + break; + case 2: gtao.sliceCount = ambient_occlusion::XE_GTAO_SLICE_COUNT_HIGH; + break; + case 3: gtao.sliceCount = ambient_occlusion::XE_GTAO_SLICE_COUNT_ULTRA; + break; + default: + break; + } + } + + int stepsPreset = 0; + if (gtao.stepsPerSliceCount == ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM) stepsPreset = 1; + else if (gtao.stepsPerSliceCount == ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_LOW) stepsPreset = 0; + else if (gtao.stepsPerSliceCount == ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA) stepsPreset = 3; + else stepsPreset = 2; + + if (ImGui::Combo("Steps Per Slice Preset", &stepsPreset, qualityPresets, IM_ARRAYSIZE(qualityPresets))) { + switch (stepsPreset) { + case 0: gtao.stepsPerSliceCount = ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_LOW; + break; + case 1: gtao.stepsPerSliceCount = ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM; + break; + case 2: gtao.stepsPerSliceCount = ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_HIGH; + break; + case 3: gtao.stepsPerSliceCount = ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA; + break; + default: + break; + } + } + + ImGui::Spacing(); + ImGui::Text("Other Parameters"); + ImGui::Separator(); + ImGui::InputInt("Debug Mode", >ao.debug); + + ImGui::Spacing(); + if (ImGui::Button("Reset to Defaults")) { + gtao.effectRadius = 0.5f; + gtao.effectFalloffRange = 0.615f; + gtao.denoiseBlurBeta = (ambient_occlusion::GTAO_DENOISE_PASSES == 0) ? (1e4f) : (1.2f); + gtao.radiusMultiplier = 1.457f; + gtao.sampleDistributionPower = 2.0f; + gtao.thinOccluderCompensation = 0.0f; + gtao.finalValuePower = 2.2f; + gtao.depthMipSamplingOffset = 3.30f; + gtao.noiseIndex = 0; + gtao.isFinalDenoisePass = 1; + gtao.sliceCount = ambient_occlusion::XE_GTAO_SLICE_COUNT_ULTRA; + gtao.stepsPerSliceCount = ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA; + gtao.debug = 0; + } + } + + if (ImGui::CollapsingHeader("GTAO Debug Preview")) { + ImGui::Checkbox("Show GTAO Debug Preview", &showGtaoDebugPreview); + + if (showGtaoDebugPreview) { + if (aoDebugTextureImguiId == VK_NULL_HANDLE) { + if (engine->ambientOcclusionPipeline->debugImage.image != VK_NULL_HANDLE) { + aoDebugTextureImguiId = ImGui_ImplVulkan_AddTexture( + engine->resourceManager->getDefaultSamplerNearest(), + engine->ambientOcclusionPipeline->debugImage.imageView, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); + } + } + + ImGui::BeginChild("GTAODebugPreview", ImVec2(0, 0), false, ImGuiWindowFlags_None); + + if (aoDebugTextureImguiId == VK_NULL_HANDLE) { + ImGui::TextColored(ImVec4(1.0f, 0.3f, 0.3f, 1.0f), "Debug texture not available."); + } + else { + // Calculate best fit size + float maxSize = ImGui::GetContentRegionAvail().x; + maxSize = glm::min(maxSize, 1024.0f); + + VkExtent3D imageExtent = engine->ambientOcclusionPipeline->debugImage.imageExtent; + float width = std::min(maxSize, static_cast(imageExtent.width)); + float aspectRatio = static_cast(imageExtent.width) / static_cast(imageExtent.height); + float height = width / aspectRatio; + + ImGui::Image(reinterpret_cast(aoDebugTextureImguiId), ImVec2(width, height)); + + ImGui::SameLine(); + if (ImGui::Button("Save GTAO Debug Image")) { + if (file::getOrCreateDirectory(file::imagesSavePath)) { + const std::filesystem::path path = file::imagesSavePath / "gtao_debug.png"; + + vk_helpers::saveImageR8G8B8A8UNORM( + *engine->resourceManager, + *engine->immediate, + engine->ambientOcclusionPipeline->debugImage, + VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, + path.string().c_str(), + 0 + ); + + ImGui::OpenPopup("SaveConfirmation"); + } + } + + if (ImGui::BeginPopupModal("SaveConfirmation", NULL, ImGuiWindowFlags_AlwaysAutoResize)) { + ImGui::Text("Image saved to %s/gtao_debug.png", file::imagesSavePath.string().c_str()); + if (ImGui::Button("OK", ImVec2(120, 0))) { + ImGui::CloseCurrentPopup(); + } + ImGui::EndPopup(); + } + } + + ImGui::EndChild(); + } + } + ImGui::EndTabItem(); } @@ -877,46 +1034,7 @@ void ImguiWrapper::imguiInterface(Engine* engine) ImGui::End(); if (ImGui::Begin("Discardable Debug")) { - ImGui::InputInt("GTAO Debug", &engine->gtaoDebug); - - if (aoDebugTextureImguiId == VK_NULL_HANDLE) { - if (engine->ambientOcclusionPipeline->debugImage.image != VK_NULL_HANDLE) { - aoDebugTextureImguiId = ImGui_ImplVulkan_AddTexture(engine->resourceManager->getDefaultSamplerNearest(), - engine->ambientOcclusionPipeline->debugImage.imageView, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL); - } - } - - - if (aoDebugTextureImguiId == VK_NULL_HANDLE) { - ImGui::Text("Issue."); - } - else { - float maxSize = ImGui::GetContentRegionAvail().x; - maxSize = glm::min(maxSize, 1024.0f); - - VkExtent3D imageExtent = engine->ambientOcclusionPipeline->debugImage.imageExtent; - float width = std::min(maxSize, static_cast(imageExtent.width)); - float aspectRatio = static_cast(imageExtent.width) / static_cast(imageExtent.height); - float height = width / aspectRatio; - - ImGui::Image(reinterpret_cast(aoDebugTextureImguiId), ImVec2(width, height)); - } - - if (ImGui::Button("Save GTAO Debug Image")) { - if (file::getOrCreateDirectory(file::imagesSavePath)) { - const std::filesystem::path path = file::imagesSavePath / "gtao_debug.png"; - - vk_helpers::saveImageR8G8B8A8UNORM( - *engine->resourceManager, - *engine->immediate, - engine->ambientOcclusionPipeline->debugImage, - VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, - path.string().c_str(), - 0 - ); - } - } + ImGui::Text("Empty"); } ImGui::End(); diff --git a/src/renderer/imgui_wrapper.h b/src/renderer/imgui_wrapper.h index c77f836d..701dc596 100644 --- a/src/renderer/imgui_wrapper.h +++ b/src/renderer/imgui_wrapper.h @@ -82,6 +82,7 @@ class ImguiWrapper VkDescriptorSet currentlySelectedTextureImguiId{VK_NULL_HANDLE}; + bool showGtaoDebugPreview = false; VkDescriptorSet aoDebugTextureImguiId{VK_NULL_HANDLE}; }; } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index d9afb53e..5139ad51 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -13,17 +13,26 @@ namespace will_engine::ambient_occlusion static constexpr int32_t DEPTH_PREFILTER_MIP_COUNT = 5; static constexpr int32_t GTAO_DENOISE_PASSES = 1; +static constexpr float XE_GTAO_SLICE_COUNT_LOW = 1.0f; +static constexpr float XE_GTAO_SLICE_COUNT_MEDIUM = 2.0f; +static constexpr float XE_GTAO_SLICE_COUNT_HIGH = 3.0f; +static constexpr float XE_GTAO_SLICE_COUNT_ULTRA = 9.0f; +static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_LOW = 2.0f; +static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM = 2.0f; +static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_HIGH = 3.0f; +static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA = 3.0f; + struct GTAOPushConstants { - glm::vec2 cameraTanHalfFOV; + glm::vec2 cameraTanHalfFOV{}; - glm::vec2 ndcToViewMul; - glm::vec2 ndcToViewAdd; + glm::vec2 ndcToViewMul{}; + glm::vec2 ndcToViewAdd{}; - glm::vec2 ndcToViewMul_x_PixelSize; + glm::vec2 ndcToViewMul_x_PixelSize{}; - float depthLinearizeMult; - float depthLinearizeAdd; + float depthLinearizeMult{0.0f}; + float depthLinearizeAdd{0.0f}; float effectRadius = 0.5f; float effectFalloffRange = 0.615f; @@ -37,6 +46,9 @@ struct GTAOPushConstants uint32_t noiseIndex{0}; int32_t isFinalDenoisePass{1}; + float sliceCount{XE_GTAO_SLICE_COUNT_ULTRA}; + float stepsPerSliceCount{XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA}; + int32_t debug{0}; }; diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp index 26ad79fd..1f8ddba1 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.cpp @@ -376,31 +376,30 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::setupS resourceManager.setupDescriptorBufferSampler(spatialFilteringDescriptorBuffer, imageDescriptors, 0); } -void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const +void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) { VkDebugUtilsLabelEXT label{}; label.sType = VK_STRUCTURE_TYPE_DEBUG_UTILS_LABEL_EXT; label.pLabelName = "GT Ambient Occlusion"; vkCmdBeginDebugUtilsLabelEXT(cmd, &label); - GTAOPushConstants push = drawInfo.pushConstants; glm::mat4 projMatrix = drawInfo.camera->getProjMatrix(); - push.depthLinearizeMult = -projMatrix[3][2]; - push.depthLinearizeAdd = projMatrix[2][2]; - if (push.depthLinearizeMult * push.depthLinearizeAdd < 0) { - push.depthLinearizeAdd = -push.depthLinearizeAdd; + gtaoPush.depthLinearizeMult = -projMatrix[3][2]; + gtaoPush.depthLinearizeAdd = projMatrix[2][2]; + if (gtaoPush.depthLinearizeMult * gtaoPush.depthLinearizeAdd < 0) { + gtaoPush.depthLinearizeAdd = -gtaoPush.depthLinearizeAdd; } float tanHalfFOVY = 1.0f / projMatrix[1][1]; float tanHalfFOVX = 1.0F / projMatrix[0][0]; - push.cameraTanHalfFOV = {tanHalfFOVX, tanHalfFOVY}; - push.ndcToViewMul = {push.cameraTanHalfFOV.x * 2.0f, push.cameraTanHalfFOV.y * -2.0f}; - push.ndcToViewAdd = {push.cameraTanHalfFOV.x * -1.0f, push.cameraTanHalfFOV.y * 1.0f}; + gtaoPush.cameraTanHalfFOV = {tanHalfFOVX, tanHalfFOVY}; + gtaoPush.ndcToViewMul = {gtaoPush.cameraTanHalfFOV.x * 2.0f, gtaoPush.cameraTanHalfFOV.y * -2.0f}; + gtaoPush.ndcToViewAdd = {gtaoPush.cameraTanHalfFOV.x * -1.0f, gtaoPush.cameraTanHalfFOV.y * 1.0f}; constexpr glm::vec2 texelSize = {1.0f / RENDER_EXTENT_WIDTH, 1.0f / RENDER_EXTENT_HEIGHT}; - push.ndcToViewMul_x_PixelSize = {push.ndcToViewMul.x * texelSize.x, push.ndcToViewMul.y * texelSize.y}; + gtaoPush.ndcToViewMul_x_PixelSize = {gtaoPush.ndcToViewMul.x * texelSize.x, gtaoPush.ndcToViewMul.y * texelSize.y}; - push.noiseIndex = GTAO_DENOISE_PASSES > 0 ? drawInfo.currentFrame % 64 : 0; + gtaoPush.noiseIndex = GTAO_DENOISE_PASSES > 0 ? drawInfo.currentFrame % 64 : 0; vk_helpers::transitionImage(cmd, debugImage.image, VK_IMAGE_LAYOUT_UNDEFINED, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_ASPECT_COLOR_BIT); @@ -408,7 +407,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V // Depth Prefilter { vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, depthPrefilterPipeline); - vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); + vkCmdPushConstants(cmd, depthPrefilterPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), >aoPush); VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; bindingInfos[0] = drawInfo.sceneDataBinding; @@ -434,7 +433,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V // Ambient Occlusion { vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, ambientOcclusionPipeline); - vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); + vkCmdPushConstants(cmd, ambientOcclusionPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), >aoPush); VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; bindingInfos[0] = drawInfo.sceneDataBinding; @@ -459,7 +458,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V // Spatial Filtering { vkCmdBindPipeline(cmd, VK_PIPELINE_BIND_POINT_COMPUTE, spatialFilteringPipeline); - vkCmdPushConstants(cmd, spatialFilteringPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), &push); + vkCmdPushConstants(cmd, spatialFilteringPipelineLayout, VK_SHADER_STAGE_COMPUTE_BIT, 0, sizeof(GTAOPushConstants), >aoPush); VkDescriptorBufferBindingInfoEXT bindingInfos[2] = {}; bindingInfos[0] = drawInfo.sceneDataBinding; @@ -478,9 +477,7 @@ void will_engine::ambient_occlusion::GroundTruthAmbientOcclusionPipeline::draw(V } vk_helpers::transitionImage(cmd, denoisedFinalAO.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, - VK_IMAGE_ASPECT_COLOR_BIT); - - + VK_IMAGE_ASPECT_COLOR_BIT); vk_helpers::transitionImage(cmd, debugImage.image, VK_IMAGE_LAYOUT_GENERAL, VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL, diff --git a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h index 5964b1f3..f5a78a74 100644 --- a/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h +++ b/src/renderer/lighting/ambient_occlusion/ground_truth/ground_truth_ambient_occlusion.h @@ -27,7 +27,7 @@ class GroundTruthAmbientOcclusionPipeline void setupSpatialFilteringDescriptorBuffer(const VkImageView& depthImageView, const VkImageView& normalsImageView); - void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo) const; + void draw(VkCommandBuffer cmd, const GTAODrawInfo& drawInfo); void reloadShaders(); @@ -87,6 +87,8 @@ class GroundTruthAmbientOcclusionPipeline private: ResourceManager& resourceManager; + GTAOPushConstants gtaoPush{}; + friend void ImguiWrapper::imguiInterface(Engine* engine); }; } diff --git a/src/renderer/pipelines/visibility_pass/visibility_pass.h b/src/renderer/pipelines/visibility_pass/visibility_pass.h index cb7695d6..1c7bf721 100644 --- a/src/renderer/pipelines/visibility_pass/visibility_pass.h +++ b/src/renderer/pipelines/visibility_pass/visibility_pass.h @@ -39,8 +39,6 @@ class VisibilityPassPipeline void draw(VkCommandBuffer cmd, const VisibilityPassDrawInfo& drawInfo) const; - void indirectBufferSynchronize(); - void reloadShaders() { createPipeline(); } private: From 339415ac06aee640b65486cb64517574ce2745ec Mon Sep 17 00:00:00 2001 From: Williscool13 Date: Mon, 31 Mar 2025 19:11:44 +0700 Subject: [PATCH 27/27] Final GTAO version. --- .../ground_truth/gtao_main_pass.comp | 14 +++++--------- src/renderer/imgui_wrapper.cpp | 14 +------------- .../ambient_occlusion/ambient_occlusion_types.h | 15 ++++++++++++--- src/renderer/renderer_constants.h | 4 ++-- 4 files changed, 20 insertions(+), 27 deletions(-) diff --git a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp index fa3d99e7..6aff7c77 100644 --- a/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp +++ b/shaders/ambient_occlusion/ground_truth/gtao_main_pass.comp @@ -87,7 +87,7 @@ vec3 cheapReconstructViewSpacePosition(vec2 uv, float viewspaceDepth) { vec3 ret; ret.xy = (pushConstants.ndcToViewMul * uv.xy + pushConstants.ndcToViewAdd) * viewspaceDepth; - //ret.y = -ret.y; + ret.y = -ret.y; ret.z = -viewspaceDepth; return ret; } @@ -149,14 +149,12 @@ void main() { // Per Intel: Move center pixel slightly towards camera to avoid imprecision artifacts due to depth buffer imprecision; offset depends on depth texture format used viewSpaceZM = viewSpaceZM * 0.998f; - //vec3 vPos = cheapReconstructViewSpacePosition(uv, viewSpaceZM); - vec4 vPosAlt = reconstructViewSpacePosition(uv, viewSpaceZM); - vec3 vPos = vPosAlt.xyz; + vec3 vPos = cheapReconstructViewSpacePosition(uv, viewSpaceZM); vec3 viewVec = normalize(-vPos); // debug world pos if (pushConstants.debug == 3){ - vec3 worldPos = (sceneData.invView * vPosAlt).xyz; + vec3 worldPos = (sceneData.invView * vec4(vPos, 1.0f)).xyz; imageStore(debugImage, screenPos, vec4(worldPos / 1000.0f, 1.0f)); } @@ -284,13 +282,11 @@ void main() { vec2 sampleScreenPos0 = uv + sampleOffset; float SZ0 = textureLod(prefilteredDepth, sampleScreenPos0, mipLevel).r; - //vec3 samplePos0 = cheapReconstructViewSpacePosition(sampleScreenPos0, SZ0); - vec3 samplePos0 = reconstructViewSpacePosition(sampleScreenPos0, SZ0).xyz; + vec3 samplePos0 = cheapReconstructViewSpacePosition(sampleScreenPos0, SZ0); vec2 sampleScreenPos1 = uv - sampleOffset; float SZ1 = textureLod(prefilteredDepth, sampleScreenPos1, mipLevel).r; - //vec3 samplePos1 = cheapReconstructViewSpacePosition(sampleScreenPos1, SZ1); - vec3 samplePos1 = reconstructViewSpacePosition(sampleScreenPos1, SZ1).xyz; + vec3 samplePos1 = cheapReconstructViewSpacePosition(sampleScreenPos1, SZ1); vec3 sampleDelta0 = (samplePos0 - vec3(vPos));// using lpfloat for sampleDelta causes precision issues vec3 sampleDelta1 = (samplePos1 - vec3(vPos));// using lpfloat for sampleDelta causes precision issues diff --git a/src/renderer/imgui_wrapper.cpp b/src/renderer/imgui_wrapper.cpp index aea583b3..b8605098 100644 --- a/src/renderer/imgui_wrapper.cpp +++ b/src/renderer/imgui_wrapper.cpp @@ -296,19 +296,7 @@ void ImguiWrapper::imguiInterface(Engine* engine) ImGui::Spacing(); if (ImGui::Button("Reset to Defaults")) { - gtao.effectRadius = 0.5f; - gtao.effectFalloffRange = 0.615f; - gtao.denoiseBlurBeta = (ambient_occlusion::GTAO_DENOISE_PASSES == 0) ? (1e4f) : (1.2f); - gtao.radiusMultiplier = 1.457f; - gtao.sampleDistributionPower = 2.0f; - gtao.thinOccluderCompensation = 0.0f; - gtao.finalValuePower = 2.2f; - gtao.depthMipSamplingOffset = 3.30f; - gtao.noiseIndex = 0; - gtao.isFinalDenoisePass = 1; - gtao.sliceCount = ambient_occlusion::XE_GTAO_SLICE_COUNT_ULTRA; - gtao.stepsPerSliceCount = ambient_occlusion::XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA; - gtao.debug = 0; + gtao = {}; } } diff --git a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h index 5139ad51..18a2904d 100644 --- a/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h +++ b/src/renderer/lighting/ambient_occlusion/ambient_occlusion_types.h @@ -22,6 +22,15 @@ static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM = 2.0f; static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_HIGH = 3.0f; static constexpr float XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA = 3.0f; +#ifdef NDEBUG +#define DEFAULT_GTAO_SLICE_COUNT XE_GTAO_SLICE_COUNT_ULTRA +#define DEFAULT_GTAO_STEPS_PER_SLICE_COUNT XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA +#else +#define DEFAULT_GTAO_SLICE_COUNT XE_GTAO_SLICE_COUNT_MEDIUM +#define DEFAULT_GTAO_STEPS_PER_SLICE_COUNT XE_GTAO_STEPS_PER_SLICE_COUNT_MEDIUM +#endif + + struct GTAOPushConstants { glm::vec2 cameraTanHalfFOV{}; @@ -46,10 +55,10 @@ struct GTAOPushConstants uint32_t noiseIndex{0}; int32_t isFinalDenoisePass{1}; - float sliceCount{XE_GTAO_SLICE_COUNT_ULTRA}; - float stepsPerSliceCount{XE_GTAO_STEPS_PER_SLICE_COUNT_ULTRA}; + float sliceCount{DEFAULT_GTAO_SLICE_COUNT}; + float stepsPerSliceCount{DEFAULT_GTAO_STEPS_PER_SLICE_COUNT}; - int32_t debug{0}; + int32_t debug{4}; }; struct GTAODrawInfo diff --git a/src/renderer/renderer_constants.h b/src/renderer/renderer_constants.h index 54db7310..b383e60b 100644 --- a/src/renderer/renderer_constants.h +++ b/src/renderer/renderer_constants.h @@ -10,8 +10,8 @@ constexpr int32_t FRAME_OVERLAP = 2; constexpr char ENGINE_NAME[] = "Will Engine"; constexpr bool USING_REVERSED_DEPTH_BUFFER = true; constexpr VkDeviceSize ZERO_DEVICE_SIZE = 0; -constexpr VkExtent2D RENDER_EXTENTS{1920, 1080}; -//constexpr VkExtent2D RENDER_EXTENTS{3840, 2160}; +//constexpr VkExtent2D RENDER_EXTENTS{1920, 1080}; +constexpr VkExtent2D RENDER_EXTENTS{3840, 2160}; constexpr float RENDER_EXTENT_WIDTH{RENDER_EXTENTS.width}; constexpr float RENDER_EXTENT_HEIGHT{RENDER_EXTENTS.height};