获取 CUDA Thrust 以使用您选择的 CUDA 流

Question

查看 CUDA Thrust 代码中的内核启动，似乎它们总是使用默认流。我可以让 Thrust 使用我选择的流吗？我是否遗漏了 API 中的某些内容？

Answer 1

我想在 Thrust 1.8 发布后更新 talonmies 提供的答案，该版本引入了将 CUDA 执行流指示为的可能性

thrust::cuda::par.on(stream)

另请参阅

推力发布1.8.0.

在下面，我将重新编写

中的示例

Fermi 架构的错误依赖问题

就 CUDA Thrust API 而言。

#include <iostream>

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

#include <stdio.h>

#include <thrust\device_vector.h>
#include <thrust\execution_policy.h>

#include "Utilities.cuh"

using namespace std;

#define NUM_THREADS 32
#define NUM_BLOCKS 16
#define NUM_STREAMS 3

struct BinaryOp{ __host__ __device__ int operator()(const int& o1,const int& o2) { return o1 * o2; } };

int main()
{
    const int N = 6000000;

    // --- Host side input data allocation and initialization. Registering host memory as page-locked (required for asynch cudaMemcpyAsync).
    int *h_in = new int[N]; for(int i = 0; i < N; i++) h_in[i] = 5;
    gpuErrchk(cudaHostRegister(h_in, N * sizeof(int), cudaHostRegisterPortable));

    // --- Host side input data allocation and initialization. Registering host memory as page-locked (required for asynch cudaMemcpyAsync).
    int *h_out = new int[N]; for(int i = 0; i < N; i++) h_out[i] = 0;
    gpuErrchk(cudaHostRegister(h_out, N * sizeof(int), cudaHostRegisterPortable));

    // --- Host side check results vector allocation and initialization
    int *h_checkResults = new int[N]; for(int i = 0; i < N; i++) h_checkResults[i] = h_in[i] * h_in[i];

    // --- Device side input data allocation.
    int *d_in = 0;              gpuErrchk(cudaMalloc((void **)&d_in, N * sizeof(int)));

    // --- Device side output data allocation. 
    int *d_out = 0;             gpuErrchk( cudaMalloc((void **)&d_out, N * sizeof(int)));

    int streamSize = N / NUM_STREAMS;
    size_t streamMemSize = N * sizeof(int) / NUM_STREAMS;

    // --- Set kernel launch configuration
    dim3 nThreads       = dim3(NUM_THREADS,1,1);
    dim3 nBlocks        = dim3(NUM_BLOCKS, 1,1);
    dim3 subKernelBlock = dim3((int)ceil((float)nBlocks.x / 2));

    // --- Create CUDA streams
    cudaStream_t streams[NUM_STREAMS];
    for(int i = 0; i < NUM_STREAMS; i++)
        gpuErrchk(cudaStreamCreate(&streams[i]));

    /**************************/
    /* BREADTH-FIRST APPROACH */
    /**************************/

    for(int i = 0; i < NUM_STREAMS; i++) {
        int offset = i * streamSize;
        cudaMemcpyAsync(&d_in[offset], &h_in[offset], streamMemSize, cudaMemcpyHostToDevice,     streams[i]);
    }

    for(int i = 0; i < NUM_STREAMS; i++)
    {
        int offset = i * streamSize;

        thrust::transform(thrust::cuda::par.on(streams[i]), thrust::device_pointer_cast(&d_in[offset]), thrust::device_pointer_cast(&d_in[offset]) + streamSize/2, 
                                                            thrust::device_pointer_cast(&d_in[offset]), thrust::device_pointer_cast(&d_out[offset]), BinaryOp());
        thrust::transform(thrust::cuda::par.on(streams[i]), thrust::device_pointer_cast(&d_in[offset + streamSize/2]), thrust::device_pointer_cast(&d_in[offset + streamSize/2]) + streamSize/2, 
                                                            thrust::device_pointer_cast(&d_in[offset + streamSize/2]), thrust::device_pointer_cast(&d_out[offset + streamSize/2]), BinaryOp());

    }

    for(int i = 0; i < NUM_STREAMS; i++) {
        int offset = i * streamSize;
        cudaMemcpyAsync(&h_out[offset], &d_out[offset], streamMemSize, cudaMemcpyDeviceToHost,   streams[i]);
    }

    for(int i = 0; i < NUM_STREAMS; i++)
        gpuErrchk(cudaStreamSynchronize(streams[i]));

    gpuErrchk(cudaDeviceSynchronize());

    // --- Release resources
    gpuErrchk(cudaHostUnregister(h_in));
    gpuErrchk(cudaHostUnregister(h_out));
    gpuErrchk(cudaFree(d_in));
    gpuErrchk(cudaFree(d_out));

    for(int i = 0; i < NUM_STREAMS; i++)
        gpuErrchk(cudaStreamDestroy(streams[i]));

    cudaDeviceReset();  

    // --- GPU output check
    int sum = 0;
    for(int i = 0; i < N; i++) {     
        //printf("%i %i\n", h_out[i], h_checkResults[i]);
        sum += h_checkResults[i] - h_out[i];
    }

    cout << "Error between CPU and GPU: " << sum << endl;

    delete[] h_in;
    delete[] h_out;
    delete[] h_checkResults;

    return 0;
}

运行此类示例所需的 Utilities.cu 和 Utilities.cuh 文件均在此 github 页面中维护。

Visual Profiler 时间线显示 CUDA Thrust 操作和内存传输的并发性

Answer 2

不，您没有遗漏任何内容（至少直到 CUDA 6.0 附带的发布快照为止）。

最初的基于 Thrust 标签的调度系统故意抽象了所有底层 CUDA API 调用，为了易用性和一致性而牺牲了一些性能（请记住，thrust 有 CUDA 以外的后端）。如果您想要这种程度的灵活性，您将需要尝试另一个库（例如 CUB）。

在 CUDA 7.0 快照以来的版本中，可以通过执行策略和调度功能设置推力操作的选择流。

Answer 3

截至最近（？），语法为：

thrust::device.on(stream)

：https://github.com/NVIDIA/thrust/issues/1626