多gpu CUDA推力 我有一个CUDA C++代码，它使用当前在单个GPU上正确工作的推力。我现在想修改它的多gpu。我有一个主机函数，它包括许多在设备阵列上排序、复制、计算差异等的推力调用。我想使用每个GPU同时在它自己的（独立的）阵列集上运行这一系列推力调用。我已经读到返回值的推力函数是同步的，但是我可以使用OpenMP让每个主机线程调用一个在单独的GPU上运行的函数（带有推力调用）吗

例如（在浏览器中编码）：

#pragma omp parallel for
对于（int-dev=0；devYes），您可以将推力和OpenMP结合使用

下面是一个完整的工作示例和结果：

$ cat t340.cu
#include <omp.h>
#include <stdio.h> 
#include <stdlib.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
#include <thrust/copy.h>
#include <time.h>
#include <sys/time.h>

#define DSIZE 200000000

using namespace std;

int main(int argc, char *argv[])
{
    timeval t1, t2;
    int num_gpus = 0;   // number of CUDA GPUs

    printf("%s Starting...\n\n", argv[0]);

    // determine the number of CUDA capable GPUs
    cudaGetDeviceCount(&num_gpus);

    if (num_gpus < 1)
    {
        printf("no CUDA capable devices were detected\n");
        return 1;
    }

    // display CPU and GPU configuration
    printf("number of host CPUs:\t%d\n", omp_get_num_procs());
    printf("number of CUDA devices:\t%d\n", num_gpus);

    for (int i = 0; i < num_gpus; i++)
    {
        cudaDeviceProp dprop;
        cudaGetDeviceProperties(&dprop, i);
        printf("   %d: %s\n", i, dprop.name);
    }

    printf("initialize data\n");


    // initialize data
    typedef thrust::device_vector<int> dvec;
    typedef dvec *p_dvec;
    std::vector<p_dvec> dvecs;

    for(unsigned int i = 0; i < num_gpus; i++) {
      cudaSetDevice(i);
      p_dvec temp = new dvec(DSIZE);
      dvecs.push_back(temp);
      }

    thrust::host_vector<int> data(DSIZE);
    thrust::generate(data.begin(), data.end(), rand);

    // copy data
    for (unsigned int i = 0; i < num_gpus; i++) {
      cudaSetDevice(i);
      thrust::copy(data.begin(), data.end(), (*(dvecs[i])).begin());
      }

    printf("start sort\n");
    gettimeofday(&t1,NULL);

    // run as many CPU threads as there are CUDA devices
    omp_set_num_threads(num_gpus);  // create as many CPU threads as there are CUDA devices
    #pragma omp parallel
    {
        unsigned int cpu_thread_id = omp_get_thread_num();
        cudaSetDevice(cpu_thread_id);
        thrust::sort((*(dvecs[cpu_thread_id])).begin(), (*(dvecs[cpu_thread_id])).end());
        cudaDeviceSynchronize();
    }
    gettimeofday(&t2,NULL);
    printf("finished\n");
    unsigned long et = ((t2.tv_sec * 1000000)+t2.tv_usec) - ((t1.tv_sec * 1000000) + t1.tv_usec);
    if (cudaSuccess != cudaGetLastError())
        printf("%s\n", cudaGetErrorString(cudaGetLastError()));
    printf("sort time = %fs\n", (float)et/(float)(1000000));
    // check results
    thrust::host_vector<int> result(DSIZE);
    thrust::sort(data.begin(), data.end());
    for (int i = 0; i < num_gpus; i++)
    {
        cudaSetDevice(i);
        thrust::copy((*(dvecs[i])).begin(), (*(dvecs[i])).end(), result.begin());
        for (int j = 0; j < DSIZE; j++)
          if (data[j] != result[j]) { printf("mismatch on device %d at index %d, host: %d, device: %d\n", i, j, data[j], result[j]); return 1;}
    }
    printf("Success\n");
    return 0;

}
$ nvcc -Xcompiler -fopenmp -O3 -arch=sm_20 -o t340 t340.cu -lgomp
$ CUDA_VISIBLE_DEVICES="0" ./t340
./t340 Starting...

number of host CPUs:    12
number of CUDA devices: 1
   0: Tesla M2050
initialize data
start sort
finished
sort time = 0.398922s
Success
$ ./t340
./t340 Starting...

number of host CPUs:    12
number of CUDA devices: 4
   0: Tesla M2050
   1: Tesla M2070
   2: Tesla M2050
   3: Tesla M2070
initialize data
start sort
finished
sort time = 0.460058s
Success
$

$cat t340.cu
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#定义DSIZE 200000000
使用名称空间std；
int main（int argc，char*argv[]）
{
时间值t1，t2；
int num_gpu=0；//CUDA GPU的数量
printf（“%s开始…\n\n”，argv[0]）；
//确定支持CUDA的GPU的数量
cudaGetDeviceCount（&numgpu）；
如果（numgpus<1）
{
printf（“未检测到支持CUDA的设备\n”）；
返回1；
}
//显示CPU和GPU配置
printf（“主机CPU数量：\t%d\n”，omp_get_num_procs（））；
printf（“CUDA设备的数量：\t%d\n”，num\u gpu）；
对于（int i=0；i

我们可以看到，当我将程序限制为使用单个设备时，排序操作大约需要0.4秒。然后，当我允许它使用所有4个设备（在所有4个设备上重复相同的排序）时，整个操作只需要0.46秒，即使我们做了4倍的工作

对于这个特殊情况，我碰巧使用了CUDA 5.0和推力v1.7，以及gcc 4.4.6（RHEL 6.2）
为什么您需要从不同的主机线程设置设备？您能用一个像中那样简单的例子吗？大概是一种解决某些推力操作阻塞主机线程这一事实的方法。这样的加速比不是理想的4x，而是0.4*4/0.46=3.48x。干得好！
$ cat t340.cu
#include <omp.h>
#include <stdio.h> 
#include <stdlib.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
#include <thrust/copy.h>
#include <time.h>
#include <sys/time.h>

#define DSIZE 200000000

using namespace std;

int main(int argc, char *argv[])
{
    timeval t1, t2;
    int num_gpus = 0;   // number of CUDA GPUs

    printf("%s Starting...\n\n", argv[0]);

    // determine the number of CUDA capable GPUs
    cudaGetDeviceCount(&num_gpus);

    if (num_gpus < 1)
    {
        printf("no CUDA capable devices were detected\n");
        return 1;
    }

    // display CPU and GPU configuration
    printf("number of host CPUs:\t%d\n", omp_get_num_procs());
    printf("number of CUDA devices:\t%d\n", num_gpus);

    for (int i = 0; i < num_gpus; i++)
    {
        cudaDeviceProp dprop;
        cudaGetDeviceProperties(&dprop, i);
        printf("   %d: %s\n", i, dprop.name);
    }

    printf("initialize data\n");


    // initialize data
    typedef thrust::device_vector<int> dvec;
    typedef dvec *p_dvec;
    std::vector<p_dvec> dvecs;

    for(unsigned int i = 0; i < num_gpus; i++) {
      cudaSetDevice(i);
      p_dvec temp = new dvec(DSIZE);
      dvecs.push_back(temp);
      }

    thrust::host_vector<int> data(DSIZE);
    thrust::generate(data.begin(), data.end(), rand);

    // copy data
    for (unsigned int i = 0; i < num_gpus; i++) {
      cudaSetDevice(i);
      thrust::copy(data.begin(), data.end(), (*(dvecs[i])).begin());
      }

    printf("start sort\n");
    gettimeofday(&t1,NULL);

    // run as many CPU threads as there are CUDA devices
    omp_set_num_threads(num_gpus);  // create as many CPU threads as there are CUDA devices
    #pragma omp parallel
    {
        unsigned int cpu_thread_id = omp_get_thread_num();
        cudaSetDevice(cpu_thread_id);
        thrust::sort((*(dvecs[cpu_thread_id])).begin(), (*(dvecs[cpu_thread_id])).end());
        cudaDeviceSynchronize();
    }
    gettimeofday(&t2,NULL);
    printf("finished\n");
    unsigned long et = ((t2.tv_sec * 1000000)+t2.tv_usec) - ((t1.tv_sec * 1000000) + t1.tv_usec);
    if (cudaSuccess != cudaGetLastError())
        printf("%s\n", cudaGetErrorString(cudaGetLastError()));
    printf("sort time = %fs\n", (float)et/(float)(1000000));
    // check results
    thrust::host_vector<int> result(DSIZE);
    thrust::sort(data.begin(), data.end());
    for (int i = 0; i < num_gpus; i++)
    {
        cudaSetDevice(i);
        thrust::copy((*(dvecs[i])).begin(), (*(dvecs[i])).end(), result.begin());
        for (int j = 0; j < DSIZE; j++)
          if (data[j] != result[j]) { printf("mismatch on device %d at index %d, host: %d, device: %d\n", i, j, data[j], result[j]); return 1;}
    }
    printf("Success\n");
    return 0;

}
$ nvcc -Xcompiler -fopenmp -O3 -arch=sm_20 -o t340 t340.cu -lgomp
$ CUDA_VISIBLE_DEVICES="0" ./t340
./t340 Starting...

number of host CPUs:    12
number of CUDA devices: 1
   0: Tesla M2050
initialize data
start sort
finished
sort time = 0.398922s
Success
$ ./t340
./t340 Starting...

number of host CPUs:    12
number of CUDA devices: 4
   0: Tesla M2050
   1: Tesla M2070
   2: Tesla M2050
   3: Tesla M2070
initialize data
start sort
finished
sort time = 0.460058s
Success
$