将cudaMalloc同步主机和设备？

Question

我知道cudaMemcpy会同步主机和设备,但cudaMalloc或cudaFree怎么样？

基本上我想在多个GPU设备上异步内存分配/复制和内核执行,我的代码的简化版本是这样的:

void wrapper_kernel(const int &ngpu, const float * const &data)
{
 cudaSetDevice(ngpu);
 cudaMalloc(...);
 cudaMemcpyAsync(...);
 kernels<<<...>>>(...);
 cudaMemcpyAsync(...);
 some host codes;
}

int main()
{
 const int NGPU=3;
 static float *data[NGPU];
 for (int i=0; i<NGPU; i++) wrapper_kernel(i,data[i]);
 cudaDeviceSynchronize();
 some host codes;
}

但是,GPU按顺序运行,无法找到原因.

Answer 1

尝试cudaStream_t对每个 GPU 使用。下面是来自 CUDA 示例的 simpleMultiGPU.cu。

 //Solver config                                                          
TGPUplan      plan[MAX_GPU_COUNT];
//GPU reduction results                                                                                   
float     h_SumGPU[MAX_GPU_COUNT];

....memory init....

//Create streams for issuing GPU command asynchronously and allocate memory (GPU and System page-locked)                             for (i = 0; i < GPU_N; i++)
{
    checkCudaErrors(cudaSetDevice(i));
    checkCudaErrors(cudaStreamCreate(&plan[i].stream));
    //Allocate memory                                                                                                                    checkCudaErrors(cudaMalloc((void **)&plan[i].d_Data, plan[i].dataN * sizeof(float)));
    checkCudaErrors(cudaMalloc((void **)&plan[i].d_Sum, ACCUM_N * sizeof(float)));
    checkCudaErrors(cudaMallocHost((void **)&plan[i].h_Sum_from_device, ACCUM_N * sizeof(float)));
    checkCudaErrors(cudaMallocHost((void **)&plan[i].h_Data, plan[i].dataN * sizeof(float)));

    for (j = 0; j < plan[i].dataN; j++)
    {
        plan[i].h_Data[j] = (float)rand() / (float)RAND_MAX;
    }
}

....kernel, memory copyback....

这是一些使用多 GPU 的指南。