让CUDA推力使用您选择的CUDA流

看看CUDA推力代码中的内核启动，它们似乎总是使用默认流。我可以使用我选择的流吗？我在API中遗漏了什么吗？

不，您没有遗漏任何东西（至少在CUDA 6.0附带的发布快照之前）

原始的基于推力标签的调度系统有意地抽象了所有底层CUDA API调用，牺牲了一些性能以便于使用和一致性（请记住，推力具有CUDA以外的后端）。如果您想要这种级别的灵活性，您需要尝试另一个库（例如CUB）

---

在CUDA 7.0快照之后的版本中，可以通过设置推力操作的选择流。

我想在推力1.8发布后更新Talonmes提供的答案，该版本引入了将CUDA执行流指示为

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

另见

在下面，我将重铸中的示例

就CUDA推力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;
}

#包括
#包括“cuda_runtime.h”
#包括“设备启动参数.h”
#包括
#包括
#包括
#包括“Utilities.cuh”
使用名称空间std；
#定义NUM_线程32
#定义NUM_块16
#定义NUM_流3
结构二进制操作{{uuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuu；
int main（）
{
常数整数N=6000000；
//---主机端输入数据分配和初始化。将主机内存注册为页面锁定（异步cudaMemcpyAsync需要）。
int*h_in=new int[N]；for（int i=0；i是否应该可以通过推力的master/development分支开始使用推力流进行实验。实验性公告是。示例语法：
推力：：排序（推力：：cuda：：par（stream），key.begin（），key.end（））
par

表示什么？@einpoklum这是一个推力执行策略，就像

推力：：seq

或

推力：：设备

一样。另一个工作示例是