dedf08166a
Added 863 files from NVIDIA/cccl sparse checkout: - c2h/ (27 files): Catch2 test helpers — generators, validators, runner - nvbench_helper/ (10 files): Benchmark harness utilities - cmake/ (29 files): CMake presets and build helpers - cudax/ (794 files): Experimental CUDA extensions - AGENTS.md: NVIDIA's official AI agent instructions for CCCL - CMakePresets.json: Standardized build configurations - cccl-version.json: Version tracking Also added CCCL_ASSET_MAP.md mapping all 4295 CCCL files to competition value and PRD items. cccl_upstream now covers 100% of competition-critical assets: - 27 tuning headers (SM80/90/100 benchmark data) - 32 dispatch headers (algorithm implementations) - 60 Thrust examples (correctness verification) - 217 CUB Catch2 tests (regression matrix) - 153 CUB benchmarks (parameter space search) - 18 CUB examples (API verification) - 27 test helpers + benchmark harness - 794 cudax experimental extensions
132 lines
3.7 KiB
Plaintext
132 lines
3.7 KiB
Plaintext
// SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
|
|
// SPDX-License-Identifier: SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
|
|
#include <thrust/device_vector.h>
|
|
|
|
#include <cuda/std/cstddef>
|
|
#include <cuda/std/cstdint>
|
|
|
|
#include <cuda/experimental/__cuco/hash_functions.cuh>
|
|
|
|
#include <nvbench/nvbench.cuh>
|
|
#include <nvbench/range.cuh>
|
|
|
|
namespace cudax = cuda::experimental;
|
|
|
|
// repeat hash computation n times
|
|
static constexpr auto n_repeats = 100;
|
|
|
|
template <cuda::std::int32_t Words>
|
|
struct large_key
|
|
{
|
|
constexpr __host__ __device__ large_key(cuda::std::int32_t seed) noexcept
|
|
{
|
|
for (cuda::std::int32_t i = 0; i < Words; ++i)
|
|
{
|
|
data_[i] = seed;
|
|
}
|
|
}
|
|
|
|
private:
|
|
cuda::std::int32_t data_[Words];
|
|
};
|
|
|
|
template <cuda::std::int32_t BlockSize, typename Key, typename Hasher, typename OutputIt>
|
|
__global__ void hash_bench_kernel(Hasher hash, size_t n, OutputIt out, bool materialize_result)
|
|
{
|
|
size_t const gid = static_cast<size_t>(BlockSize) * blockIdx.x + threadIdx.x;
|
|
size_t const loop_stride = static_cast<size_t>(gridDim.x) * BlockSize;
|
|
size_t idx = gid;
|
|
using result_t = decltype(hash(0));
|
|
|
|
result_t agg{};
|
|
|
|
while (idx < n)
|
|
{
|
|
Key key(idx);
|
|
for (cuda::std::int32_t i = 0; i < n_repeats; ++i)
|
|
{ // execute hash func n times
|
|
agg += hash(key);
|
|
}
|
|
idx += loop_stride;
|
|
}
|
|
|
|
if (materialize_result)
|
|
{
|
|
out[gid] = agg;
|
|
}
|
|
}
|
|
|
|
// benchmark evaluating performance of various hash functions
|
|
template <typename HasherTag, typename Key>
|
|
void hash_eval(nvbench::state& state, nvbench::type_list<HasherTag, Key>)
|
|
{
|
|
using Hash = typename HasherTag::template fn<Key>;
|
|
|
|
bool const materialize_result = false;
|
|
constexpr auto block_size = 128;
|
|
auto const num_keys = state.get_int64("NumInputs");
|
|
auto const grid_size = (num_keys + block_size * 16 - 1) / block_size * 16;
|
|
using result_t = decltype(std::declval<Hash>()(std::declval<cuda::std::int32_t>()));
|
|
|
|
thrust::device_vector<result_t> hash_values((materialize_result) ? num_keys : 1);
|
|
|
|
state.add_element_count(num_keys);
|
|
|
|
state.exec([&](nvbench::launch& launch) {
|
|
hash_bench_kernel<block_size, Key>
|
|
<<<grid_size, block_size, 0, launch.get_stream()>>>(Hash{}, num_keys, hash_values.begin(), materialize_result);
|
|
});
|
|
}
|
|
|
|
struct xxhash_32_tag
|
|
{
|
|
template <typename Key>
|
|
using fn = cudax::cuco::hash<Key, cudax::cuco::hash_algorithm::xxhash_32>;
|
|
};
|
|
|
|
struct xxhash_64_tag
|
|
{
|
|
template <typename Key>
|
|
using fn = cudax::cuco::hash<Key, cudax::cuco::hash_algorithm::xxhash_64>;
|
|
};
|
|
|
|
struct murmurhash3_32_tag
|
|
{
|
|
template <typename Key>
|
|
using fn = cudax::cuco::hash<Key, cudax::cuco::hash_algorithm::murmurhash3_32>;
|
|
};
|
|
|
|
#if _CCCL_HAS_INT128()
|
|
|
|
struct murmurhash3_x86_128_tag
|
|
{
|
|
template <typename Key>
|
|
using fn = cudax::cuco::hash<Key, cudax::cuco::hash_algorithm::murmurhash3_x86_128>;
|
|
};
|
|
|
|
struct murmurhash3_x64_128_tag
|
|
{
|
|
template <typename Key>
|
|
using fn = cudax::cuco::hash<Key, cudax::cuco::hash_algorithm::murmurhash3_x64_128>;
|
|
};
|
|
|
|
#endif // _CCCL_HAS_INT128()
|
|
|
|
NVBENCH_BENCH_TYPES(
|
|
hash_eval,
|
|
NVBENCH_TYPE_AXES(
|
|
nvbench::type_list<xxhash_32_tag,
|
|
xxhash_64_tag,
|
|
murmurhash3_32_tag
|
|
#if _CCCL_HAS_INT128()
|
|
,
|
|
murmurhash3_x86_128_tag,
|
|
murmurhash3_x64_128_tag
|
|
#endif // _CCCL_HAS_INT128()
|
|
>,
|
|
nvbench::type_list<cuda::std::int32_t, large_key<4>, large_key<8>, large_key<16>, large_key<32>>))
|
|
.set_name("hash_function_eval")
|
|
.set_type_axes_names({"Hash", "Key"})
|
|
.add_int64_power_of_two_axis("NumInputs", nvbench::range(18, 26, 4));
|