[CCCL] 瘦身 + 补全: 移除 cudax/python/libcudacxx-tests 冗余文件, 新增 c2h 测试助手 + cmake 构建系统 + 8 个 CUDA thrust examples
变更摘要:
- 删除: cudax/ (783 files, 7.2M) — 实验性组件，竞赛不需要
- 删除: python/ (226 files, 2.0M) — Python 绑定，竞赛不需要
- 删除: libcudacxx/{test,benchmarks,codegen,cmake,share} (4432 files, 31M)
保留: libcudacxx/include/ (1463 headers, cuda::std 编译依赖)
- 新增: c2h/ (27 files) — CUB Catch2 测试辅助头文件，编译 243 个测试必需
- 新增: cmake/ (29 files) — CCCL 原生 CMake 构建系统
- 新增: thrust/examples/cuda/ (7 files) + cpp_integration/ (1 file)
async_reduce, custom_temporary_allocation, explicit_cuda_stream,
global_device_vector, range_view, unwrap_pointer, wrap_pointer, device
结果: cccl_upstream 从 74M→35M (瘦身 53%), 核心内容 100% 保留:
27/27 tuning headers, 78 benchmarks, 243 tests,
60 thrust examples, 18 CUB examples, 全部编译头文件
This commit is contained in:
muh-bot
@@ -1,252 +0,0 @@
|
||||
//===----------------------------------------------------------------------===//
|
||||
//
|
||||
// Part of CUDA Experimental in CUDA C++ Core Libraries,
|
||||
// under the Apache License v2.0 with LLVM Exceptions.
|
||||
// See https://llvm.org/LICENSE.txt for license information.
|
||||
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
||||
// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
|
||||
//
|
||||
//===----------------------------------------------------------------------===//
|
||||
|
||||
#include <cuda/devices>
|
||||
#include <cuda/functional>
|
||||
#include <cuda/hierarchy>
|
||||
#include <cuda/launch>
|
||||
#include <cuda/std/algorithm>
|
||||
#include <cuda/std/type_traits>
|
||||
#include <cuda/stream>
|
||||
|
||||
#include <cuda/experimental/coop.cuh>
|
||||
#include <cuda/experimental/group.cuh>
|
||||
|
||||
#include <testing.cuh>
|
||||
|
||||
#include <c2h/catch2_test_helper.h>
|
||||
#include <c2h/extended_types.h>
|
||||
#include <c2h/generators.h>
|
||||
#include <catch2/matchers/catch_matchers_floating_point.hpp>
|
||||
|
||||
constexpr int cluster_size = 2;
|
||||
constexpr int block_size = 128;
|
||||
|
||||
/***********************************************************************************************************************
|
||||
* Thread Reduce Wrapper Kernels
|
||||
**********************************************************************************************************************/
|
||||
|
||||
template <bool Broadcasted>
|
||||
struct ReduceKernel
|
||||
{
|
||||
template <class Config, int NumItems, class T, class RedOp>
|
||||
__device__ void operator()(
|
||||
Config config,
|
||||
cuda::std::integral_constant<int, NumItems>,
|
||||
const T* __restrict__ d_in,
|
||||
T* __restrict__ d_out,
|
||||
RedOp red_op)
|
||||
{
|
||||
cudax::this_grid grid{config};
|
||||
|
||||
T thread_data[NumItems];
|
||||
for (int i = 0; i < NumItems; ++i)
|
||||
{
|
||||
thread_data[i] = d_in[cuda::gpu_thread.rank_as<int>(grid) + i * cuda::gpu_thread.count_as<int>(grid)];
|
||||
}
|
||||
|
||||
if constexpr (Broadcasted)
|
||||
{
|
||||
const auto result = cudax::coop::reduce(cudax::broadcasted, grid, thread_data, red_op);
|
||||
|
||||
d_out[cuda::gpu_thread.rank(grid)] = result;
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto result = cudax::coop::reduce(grid, thread_data, red_op);
|
||||
|
||||
REQUIRE(result.has_value() == cuda::gpu_thread.is_root_rank(grid));
|
||||
if (cuda::gpu_thread.is_root_rank(grid))
|
||||
{
|
||||
*d_out = result.value();
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
/***********************************************************************************************************************
|
||||
* Type list definition
|
||||
**********************************************************************************************************************/
|
||||
|
||||
using integral_type_list =
|
||||
c2h::type_list<cuda::std::int8_t, cuda::std::int16_t, cuda::std::uint16_t, cuda::std::int32_t, cuda::std::int64_t>;
|
||||
|
||||
using fp_type_list = c2h::type_list<float, double>;
|
||||
|
||||
using operator_integral_list =
|
||||
c2h::type_list<cuda::std::plus<>,
|
||||
cuda::std::multiplies<>,
|
||||
cuda::std::bit_and<>,
|
||||
cuda::std::bit_or<>,
|
||||
cuda::std::bit_xor<>,
|
||||
cuda::minimum<>,
|
||||
cuda::maximum<>>;
|
||||
|
||||
using operator_fp_list = c2h::type_list<cuda::std::plus<>, cuda::std::multiplies<>, cuda::minimum<>, cuda::maximum<>>;
|
||||
|
||||
using grid_size_list = c2h::enum_type_list<int, 1, 12, 32>;
|
||||
|
||||
/***********************************************************************************************************************
|
||||
* Verify results and kernel launch`
|
||||
**********************************************************************************************************************/
|
||||
|
||||
template <class T>
|
||||
void verify_results(const T& expected_data, const T& test_results)
|
||||
{
|
||||
if constexpr (cuda::std::is_floating_point_v<T>)
|
||||
{
|
||||
REQUIRE_THAT(expected_data, Catch::Matchers::WithinRel(test_results, T{0.05}));
|
||||
}
|
||||
else
|
||||
{
|
||||
REQUIRE(expected_data == test_results);
|
||||
}
|
||||
}
|
||||
|
||||
template <int GridSize, class T, class RedOp, bool Broadcasted = false>
|
||||
void run_reduce_kernel(
|
||||
cuda::stream_ref stream,
|
||||
cuda::std::integral_constant<int, GridSize>,
|
||||
int num_items,
|
||||
const c2h::device_vector<T>& in,
|
||||
c2h::device_vector<T>& out,
|
||||
RedOp red_op,
|
||||
cuda::std::bool_constant<Broadcasted> = {})
|
||||
{
|
||||
const auto config = cuda::make_config(
|
||||
cuda::grid_dims<GridSize>(),
|
||||
cuda::cluster_dims<cluster_size>(),
|
||||
cuda::block_dims<block_size>(),
|
||||
cuda::cooperative_launch{});
|
||||
const auto in_ptr = thrust::raw_pointer_cast(in.data());
|
||||
const auto out_ptr = thrust::raw_pointer_cast(out.data());
|
||||
const ReduceKernel<Broadcasted> kernel{};
|
||||
|
||||
switch (num_items)
|
||||
{
|
||||
case 1:
|
||||
cuda::launch(stream, config, kernel, cuda::std::integral_constant<int, 1>{}, in_ptr, out_ptr, red_op);
|
||||
break;
|
||||
case 4:
|
||||
cuda::launch(stream, config, kernel, cuda::std::integral_constant<int, 4>{}, in_ptr, out_ptr, red_op);
|
||||
break;
|
||||
default:
|
||||
FAIL("Unsupported number of items");
|
||||
}
|
||||
stream.sync();
|
||||
}
|
||||
|
||||
constexpr int max_size = 4;
|
||||
constexpr int num_seeds = 10;
|
||||
|
||||
/***********************************************************************************************************************
|
||||
* Test cases
|
||||
**********************************************************************************************************************/
|
||||
|
||||
_CCCL_DIAG_SUPPRESS_MSVC(4244) // warning C4244: '=': conversion from 'int' to '_Tp', possible loss of data
|
||||
|
||||
C2H_TEST("reduce/this_grid Integral Type Tests",
|
||||
"[reduce][this_grid]",
|
||||
integral_type_list,
|
||||
operator_integral_list,
|
||||
grid_size_list)
|
||||
{
|
||||
const auto device = cuda::devices[0];
|
||||
if (cuda::device_attributes::compute_capability_major(device) < 9)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
using value_t = c2h::get<0, TestType>;
|
||||
using op_t = c2h::get<1, TestType>;
|
||||
using grid_size_t = c2h::get<2, TestType>;
|
||||
constexpr auto reduce_op = op_t{};
|
||||
constexpr auto operator_identity = cuda::identity_element<op_t, value_t>();
|
||||
CAPTURE(c2h::type_name<value_t>(), max_size, c2h::type_name<decltype(reduce_op)>());
|
||||
c2h::device_vector<value_t> d_in(max_size * grid_size_t::value * cluster_size * block_size);
|
||||
c2h::device_vector<value_t> d_out(1);
|
||||
c2h::gen(C2H_SEED(num_seeds), d_in, cuda::std::numeric_limits<value_t>::min());
|
||||
c2h::host_vector<value_t> h_in = d_in;
|
||||
cuda::stream stream{device};
|
||||
for (int num_items : {1, 4})
|
||||
{
|
||||
auto reference_result = cuda::std::accumulate(
|
||||
h_in.begin(),
|
||||
h_in.begin() + num_items * grid_size_t::value * cluster_size * block_size,
|
||||
operator_identity,
|
||||
reduce_op);
|
||||
run_reduce_kernel(stream, grid_size_t{}, num_items, d_in, d_out, reduce_op);
|
||||
verify_results(reference_result, c2h::host_vector<value_t>(d_out)[0]);
|
||||
}
|
||||
}
|
||||
|
||||
C2H_TEST(
|
||||
"reduce/this_grid Floating-Point Type Tests", "[reduce][this_grid]", fp_type_list, operator_fp_list, grid_size_list)
|
||||
{
|
||||
const auto device = cuda::devices[0];
|
||||
if (cuda::device_attributes::compute_capability_major(device) < 9)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
using value_t = c2h::get<0, TestType>;
|
||||
using op_t = c2h::get<1, TestType>;
|
||||
using grid_size_t = c2h::get<2, TestType>;
|
||||
constexpr auto reduce_op = op_t{};
|
||||
const auto operator_identity = cuda::identity_element<op_t, value_t>();
|
||||
CAPTURE(c2h::type_name<value_t>(), max_size, c2h::type_name<decltype(reduce_op)>());
|
||||
c2h::device_vector<value_t> d_in(max_size * grid_size_t::value * cluster_size * block_size);
|
||||
c2h::device_vector<value_t> d_out(1);
|
||||
c2h::gen(C2H_SEED(num_seeds), d_in, cuda::std::numeric_limits<value_t>::min());
|
||||
c2h::host_vector<value_t> h_in = d_in;
|
||||
cuda::stream stream{device};
|
||||
for (int num_items : {1, 4})
|
||||
{
|
||||
auto reference_result = cuda::std::accumulate(
|
||||
h_in.begin(),
|
||||
h_in.begin() + num_items * grid_size_t::value * cluster_size * block_size,
|
||||
operator_identity,
|
||||
reduce_op);
|
||||
run_reduce_kernel(stream, grid_size_t{}, num_items, d_in, d_out, reduce_op);
|
||||
verify_results(reference_result, c2h::host_vector<value_t>(d_out)[0]);
|
||||
}
|
||||
}
|
||||
|
||||
C2H_TEST("reduce/this_grid Broadcasted", "[reduce][this_grid]", integral_type_list, grid_size_list)
|
||||
{
|
||||
const auto device = cuda::devices[0];
|
||||
if (cuda::device_attributes::compute_capability_major(device) < 9)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
using value_t = c2h::get<0, TestType>;
|
||||
using op_t = cuda::std::plus<>;
|
||||
using grid_size_t = c2h::get<1, TestType>;
|
||||
constexpr auto reduce_op = op_t{};
|
||||
constexpr auto operator_identity = cuda::identity_element<op_t, value_t>();
|
||||
CAPTURE(c2h::type_name<value_t>(), max_size, c2h::type_name<decltype(reduce_op)>());
|
||||
c2h::device_vector<value_t> d_in(max_size * grid_size_t::value * cluster_size * block_size);
|
||||
c2h::gen(C2H_SEED(num_seeds), d_in, cuda::std::numeric_limits<value_t>::min());
|
||||
c2h::host_vector<value_t> h_in = d_in;
|
||||
cuda::stream stream{device};
|
||||
for (int num_items : {1, 4})
|
||||
{
|
||||
c2h::device_vector<value_t> d_out(grid_size_t::value * cluster_size * block_size);
|
||||
auto reference_result = cuda::std::accumulate(
|
||||
h_in.begin(),
|
||||
h_in.begin() + num_items * grid_size_t::value * cluster_size * block_size,
|
||||
operator_identity,
|
||||
reduce_op);
|
||||
run_reduce_kernel(stream, grid_size_t{}, num_items, d_in, d_out, reduce_op, cuda::std::true_type{});
|
||||
verify_results(c2h::host_vector<value_t>(grid_size_t::value * cluster_size * block_size, reference_result),
|
||||
c2h::host_vector<value_t>(d_out));
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user