-
Notifications
You must be signed in to change notification settings - Fork 108
Add static_multiset::for_each and its OA impl #506
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Changes from 6 commits
f14f521
4108a41
099503c
ca41a48
e7a8e03
7053703
18c5f60
22dab49
ee3dae7
bd309c1
6f6e5ff
7cd072c
d2ade79
6343ca1
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -962,6 +962,106 @@ class open_addressing_ref_impl { | |
| } | ||
| } | ||
|
|
||
| /** | ||
| * @brief Executes a callback on every element in the container with key equivalent to the probe | ||
| * key. | ||
| * | ||
| * @note Passes an un-incrementable input iterator to the element whose key is equivalent to | ||
| * `key` to the callback. | ||
| * | ||
| * @tparam ProbeKey Input type which is convertible to 'key_type' | ||
| + @tparam Callback Callback functor or lambda | ||
| * | ||
| * @param key The key to search for | ||
| * @param callback Function to call on every element found | ||
| */ | ||
| template <class ProbeKey, class Callback> | ||
| __device__ void for_each(ProbeKey const& key, Callback callback) const noexcept | ||
| { | ||
| static_assert(cg_size == 1, "Non-CG operation is incompatible with the current probing scheme"); | ||
| auto probing_iter = this->probing_scheme_(key, this->storage_ref_.window_extent()); | ||
|
|
||
| while (true) { | ||
| // TODO atomic_ref::load if insert operator is present | ||
| auto const window_slots = this->storage_ref_[*probing_iter]; | ||
|
|
||
| for (auto i = 0; i < window_size; ++i) { | ||
| switch ( | ||
| this->predicate_.operator()<is_insert::NO>(key, this->extract_key(window_slots[i]))) { | ||
| case detail::equal_result::EMPTY: { | ||
| return; | ||
| } | ||
| case detail::equal_result::EQUAL: { | ||
| callback(const_iterator{&(*(this->storage_ref_.data() + *probing_iter))[i]}); | ||
| if constexpr (allows_duplicates) { | ||
| continue; | ||
| } else { | ||
| return; | ||
| } | ||
| } | ||
| default: continue; | ||
| } | ||
| } | ||
| ++probing_iter; | ||
| } | ||
| } | ||
|
|
||
| /** | ||
| * @brief Executes a callback on every element in the container with key equivalent to the probe | ||
| * key. | ||
| * | ||
| * @note Passes an un-incrementable input iterator to the element whose key is equivalent to | ||
| * `key` to the callback. | ||
| * | ||
| * @note This function uses cooperative group semantics, meaning that any thread may call the | ||
| * callback if it finds a matching element. If multiple elements are found within the same group, | ||
| * each thread with a match will call the callback with its associated element. | ||
| * | ||
| * @tparam ProbeKey Input type which is convertible to 'key_type' | ||
| + @tparam Callback Callback functor or lambda | ||
| * | ||
| * @param group The Cooperative Group used to perform this operation | ||
| * @param key The key to search for | ||
| * @param callback Function to call on every element found | ||
| */ | ||
| template <class ProbeKey, class Callback> | ||
| __device__ void for_each(cooperative_groups::thread_block_tile<cg_size> const& group, | ||
|
Collaborator
Author
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Not sure why the unit test is failing. Seems like the logic in this function is flawed.
Collaborator
Author
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I think I found the problem. #509 should fix the issue. |
||
| ProbeKey const& key, | ||
| Callback callback) const noexcept | ||
| { | ||
| auto probing_iter = probing_scheme_(group, key, storage_ref_.window_extent()); | ||
|
|
||
| while (true) { | ||
| auto const window_slots = storage_ref_[*probing_iter]; | ||
|
|
||
| auto const [state, intra_window_index] = [&]() { | ||
| auto res = detail::equal_result::UNEQUAL; | ||
| for (auto i = 0; i < window_size; ++i) { | ||
| res = this->predicate_.operator()<is_insert::NO>(key, this->extract_key(window_slots[i])); | ||
| if (res != detail::equal_result::UNEQUAL) { return window_probing_results{res, i}; } | ||
| } | ||
| // returns dummy index `-1` for UNEQUAL | ||
| return window_probing_results{res, -1}; | ||
| }(); | ||
|
|
||
| // Find a match for the probe key, thus call the callback with an iterator to the entry | ||
| auto const equal = state == detail::equal_result::EQUAL; | ||
| if (equal) { | ||
| callback(const_iterator{&(*(storage_ref_.data() + *probing_iter))[intra_window_index]}); | ||
| } | ||
|
|
||
| if constexpr (not allows_duplicates) { | ||
| if (group.any(equal)) { return; } | ||
| } | ||
|
|
||
| // Find an empty slot, meaning that the probe key isn't present in the container | ||
| auto const empty = state == detail::equal_result::EMPTY; | ||
| if (group.any(empty)) { return; } | ||
|
|
||
| ++probing_iter; | ||
| } | ||
| } | ||
|
|
||
| /** | ||
| * @brief Compares the content of the address `address` (old value) with the `expected` value and, | ||
| * only if they are the same, sets the content of `address` to `desired`. | ||
|
|
||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -62,6 +62,12 @@ struct count_tag { | |
| struct find_tag { | ||
| } inline constexpr find; ///< `cuco::find` operator | ||
|
|
||
| /** | ||
| * @brief `for_each` operator tag | ||
| */ | ||
| struct for_each_tag { | ||
| } inline constexpr for_each; ///< `cuco::for_each` operator | ||
|
Comment on lines
+68
to
+69
Member
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I see
Collaborator
Author
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. From my standpoint I would treat it as an extension to the STL API that is more suitable for the GPU. Having a "cooperative iterator" instead, which would be closer to the spirit of modern C++ has its drawbacks. For example, how do we ensure users only increment the iterator with the same CG?
Collaborator
Author
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. On a side note I personally find this funtional approach, i.e., "for each found key do X" very appealing. Historic evidence that it is indeed useful comes from warpcore, where many downstream applications (mostly genomics stuff) implemented their custom lookup operations through |
||
|
|
||
| } // namespace op | ||
| } // namespace cuco | ||
|
|
||
|
|
||
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,139 @@ | ||
| /* | ||
| * Copyright (c) 2024, NVIDIA CORPORATION. | ||
| * | ||
| * Licensed under the Apache License, Version 2.0 (the "License"); | ||
| * you may not use this file except in compliance with the License. | ||
| * You may obtain a copy of the License at | ||
| * | ||
| * http://www.apache.org/licenses/LICENSE-2.0 | ||
| * | ||
| * Unless required by applicable law or agreed to in writing, software | ||
| * distributed under the License is distributed on an "AS IS" BASIS, | ||
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| * See the License for the specific language governing permissions and | ||
| * limitations under the License. | ||
| */ | ||
|
|
||
| #include <test_utils.hpp> | ||
|
|
||
| #include <cuco/detail/utility/cuda.hpp> | ||
| #include <cuco/static_multiset.cuh> | ||
|
|
||
| #include <cuda/atomic> | ||
| #include <cuda/functional> | ||
| #include <thrust/iterator/counting_iterator.h> | ||
| #include <thrust/iterator/transform_iterator.h> | ||
|
|
||
| #include <cooperative_groups.h> | ||
| #include <cooperative_groups/reduce.h> | ||
|
|
||
| #include <catch2/catch_template_test_macros.hpp> | ||
|
|
||
| #include <cstddef> | ||
|
|
||
| template <class Ref, class InputIt, class AtomicErrorCounter> | ||
| CUCO_KERNEL void for_each_check_scalar(Ref ref, | ||
| InputIt first, | ||
| std::size_t n, | ||
| std::size_t multiplicity, | ||
| AtomicErrorCounter* error_counter) | ||
| { | ||
| static_assert(Ref::cg_size == 1, "Scalar test must have cg_size==1"); | ||
| auto const loop_stride = cuco::detail::grid_stride(); | ||
| auto idx = cuco::detail::global_thread_id(); | ||
|
|
||
| while (idx < n) { | ||
| auto const& key = *(first + idx); | ||
| std::size_t matches = 0; | ||
| ref.for_each(key, [&] __device__(auto const it) { | ||
| if (ref.key_eq()(key, *it)) { matches++; } | ||
| }); | ||
| if (matches != multiplicity) { error_counter->fetch_add(1, cuda::memory_order_relaxed); } | ||
| idx += loop_stride; | ||
| } | ||
| } | ||
|
|
||
| template <class Ref, class InputIt, class AtomicErrorCounter> | ||
| CUCO_KERNEL void for_each_check_cooperative(Ref ref, | ||
| InputIt first, | ||
| std::size_t n, | ||
| std::size_t multiplicity, | ||
| AtomicErrorCounter* error_counter) | ||
| { | ||
| auto const loop_stride = cuco::detail::grid_stride() / Ref::cg_size; | ||
| auto idx = cuco::detail::global_thread_id() / Ref::cg_size; | ||
| ; | ||
|
|
||
| while (idx < n) { | ||
| auto const tile = | ||
| cooperative_groups::tiled_partition<Ref::cg_size>(cooperative_groups::this_thread_block()); | ||
| auto const& key = *(first + idx); | ||
| std::size_t thread_matches = 0; | ||
| ref.for_each(tile, key, [&] __device__(auto const it) { | ||
| if (ref.key_eq()(key, *it)) { thread_matches++; } | ||
| }); | ||
| auto const tile_matches = | ||
| cooperative_groups::reduce(tile, thread_matches, cooperative_groups::plus<std::size_t>()); | ||
| if (tile_matches != multiplicity and tile.thread_rank() == 0) { | ||
| error_counter->fetch_add(1, cuda::memory_order_relaxed); | ||
| } | ||
| idx += loop_stride; | ||
| } | ||
| } | ||
|
|
||
| TEMPLATE_TEST_CASE_SIG( | ||
| "static_multiset for_each tests", | ||
| "", | ||
| ((typename Key, cuco::test::probe_sequence Probe, int CGSize), Key, Probe, CGSize), | ||
| (int32_t, cuco::test::probe_sequence::double_hashing, 1), | ||
| (int32_t, cuco::test::probe_sequence::double_hashing, 2), | ||
| (int64_t, cuco::test::probe_sequence::double_hashing, 1), | ||
| (int64_t, cuco::test::probe_sequence::double_hashing, 2), | ||
| (int32_t, cuco::test::probe_sequence::linear_probing, 1), | ||
| (int32_t, cuco::test::probe_sequence::linear_probing, 2), | ||
| (int64_t, cuco::test::probe_sequence::linear_probing, 1), | ||
| (int64_t, cuco::test::probe_sequence::linear_probing, 2)) | ||
| { | ||
| constexpr size_t num_unique_keys{400}; | ||
| constexpr size_t key_multiplicity{5}; | ||
| constexpr size_t num_keys{num_unique_keys * key_multiplicity}; | ||
|
|
||
| using probe = std::conditional_t<Probe == cuco::test::probe_sequence::linear_probing, | ||
| cuco::linear_probing<CGSize, cuco::default_hash_function<Key>>, | ||
| cuco::double_hashing<CGSize, cuco::default_hash_function<Key>>>; | ||
|
|
||
| auto set = | ||
| cuco::static_multiset{num_keys, cuco::empty_key<Key>{-1}, {}, probe{}, {}, cuco::storage<2>{}}; | ||
|
|
||
| auto unique_keys_begin = thrust::counting_iterator<Key>(0); | ||
| auto gen_duplicate_keys = cuda::proclaim_return_type<Key>( | ||
| [] __device__(auto const& k) { return static_cast<Key>(k % num_unique_keys); }); | ||
| auto keys_begin = thrust::make_transform_iterator(unique_keys_begin, gen_duplicate_keys); | ||
|
|
||
| set.insert(keys_begin, keys_begin + num_keys); | ||
|
|
||
| using error_counter_type = cuda::atomic<std::size_t, cuda::thread_scope_system>; | ||
| error_counter_type* error_counter; | ||
| CUCO_CUDA_TRY(cudaMallocHost(&error_counter, sizeof(error_counter_type))); | ||
| new (error_counter) error_counter_type{0}; | ||
|
|
||
| auto const grid_size = cuco::detail::grid_size(num_unique_keys, CGSize); | ||
| auto const block_size = cuco::detail::default_block_size(); | ||
|
|
||
| // test scalar for_each | ||
| if constexpr (CGSize == 1) { | ||
| for_each_check_scalar<<<grid_size, block_size>>>( | ||
| set.ref(cuco::for_each), unique_keys_begin, num_unique_keys, key_multiplicity, error_counter); | ||
| CUCO_CUDA_TRY(cudaDeviceSynchronize()); | ||
| REQUIRE(error_counter->load() == 0); | ||
| error_counter->store(0); | ||
| } | ||
|
|
||
| // test CG for_each | ||
| for_each_check_cooperative<<<grid_size, block_size>>>( | ||
| set.ref(cuco::for_each), unique_keys_begin, num_unique_keys, key_multiplicity, error_counter); | ||
| CUCO_CUDA_TRY(cudaDeviceSynchronize()); | ||
| REQUIRE(error_counter->load() == 0); | ||
|
|
||
| CUCO_CUDA_TRY(cudaFreeHost(error_counter)); | ||
| } |
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Unsure if this needs to be a mutable or even universal reference instead. Let's say we define a
countfunctor as such:And then call
Then we want the functor to be taken as a mutable reference, right?
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
pass by value is preferred.
The above example is a good example of a bad callback, especially in a parallel context
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is it best practice to pass a callback by-value? I'd have to skim some stackoverflow/cppreference pages to get familiar with the topic. With pass-by-value we lose the ability of giving the callback an internal state that can hold the result of the operation. How would we solve the above example with a callback passed by-value? Pass a pointer to
thread_countto the callback?There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes
I see a callable defining the operations to be performed on the output instead of being the output itself.