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
61 lines
1.9 KiB
Plaintext
61 lines
1.9 KiB
Plaintext
//===----------------------------------------------------------------------===//
|
|
//
|
|
// Part of CUDASTF 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) 2022-2024 NVIDIA CORPORATION & AFFILIATES.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include <cuda/experimental/__stf/allocators/adapters.cuh>
|
|
#include <cuda/experimental/stf.cuh>
|
|
|
|
using namespace cuda::experimental::stf;
|
|
|
|
int main()
|
|
{
|
|
double* d_ptrA;
|
|
const size_t N = 128 * 1024;
|
|
const size_t NITER = 10;
|
|
|
|
// User allocated memory
|
|
cuda_safe_call(cudaMalloc(&d_ptrA, N * sizeof(double)));
|
|
|
|
async_resources_handle handle;
|
|
|
|
cudaStream_t stream;
|
|
cuda_safe_call(cudaStreamCreate(&stream));
|
|
|
|
for (size_t i = 0; i < NITER; i++)
|
|
{
|
|
graph_ctx ctx(stream, handle);
|
|
|
|
// The uncached allocator of the context will be using cudaMallocAsync(...,
|
|
// stream) to avoid creating memory nodes in the graph (because they are
|
|
// costly and caching the graph also keeps memory allocated)
|
|
auto wrapper = stream_adapter(ctx, stream);
|
|
|
|
ctx.set_allocator(block_allocator<buddy_allocator>(ctx, wrapper.allocator()));
|
|
|
|
auto A = ctx.logical_data(make_slice(d_ptrA, N), data_place::current_device());
|
|
|
|
for (size_t k = 0; k < 4; k++)
|
|
{
|
|
auto tmp = ctx.logical_data(A.shape());
|
|
auto tmp2 = ctx.logical_data(A.shape());
|
|
// Test device and managed memory
|
|
ctx.parallel_for(A.shape(), A.read(), tmp.write(), tmp2.write(data_place::managed()))
|
|
->*[] __device__(size_t i, auto a, auto tmp, auto tmp2) {
|
|
tmp(i) = a(i);
|
|
tmp2(i) = a(i);
|
|
};
|
|
}
|
|
|
|
ctx.finalize();
|
|
|
|
wrapper.clear();
|
|
}
|
|
cuda_safe_call(cudaStreamSynchronize(stream));
|
|
}
|