如何在 CUDA 中执行多个矩阵乘法？

我有一个方阵数组int *M[10];以便M[i]定位第一个元素i-th 矩阵。我想将所有矩阵相乘M[i]通过另一个矩阵N，这样我就收到了方阵数组int *P[10]作为输出。

我看到有不同的可能性：

1. 分配不同元素的计算M[i]到不同的线程；例如，我有10矩阵，4x4大小，以便涉及的线程数为160;如何使用CUDA来实现这种方法？
2. 在上面例子的框架中，创建一个复合矩阵大小40x40（即收集10, 4x4大小矩阵在一起）并使用40x40线程；但这种方法似乎需要更多时间；我正在尝试使用矩阵数组，但我认为我做错了；我怎样才能使用这种方法10矩阵？如何在内核函数中编写它？

这就是我正在尝试的；

void GPU_Multi(int *M[2], int *N, int *P[2], size_t width)
{

    int *devM[2];
    int *devN[2];
    int *devP[2];
    size_t allocasize =sizeof(int) *width*width;

    for(int i = 0 ; i < 10 ; i ++ ) 
    {
        cudaMalloc((void**)&devM[ i ], allocasize );
        cudaMalloc((void**)&devP[ i ], allocasize ); 
    }

    cudaMalloc((void**)&devN, allocasize );

    for(int i = 0 ; i < 10 ; i ++ ) {

        cudaMemcpy(devM[ i ],M[ i ], allocasize , cudaMemcpyHostToDevice);
        cudaMemcpy(devN, N, allocasize , cudaMemcpyHostToDevice);
        dim3 block(width*2, width*2);
        dim3 grid(1,1,1);
        Kernel_Function<<<grid, block>>>  (devM[2], devN, devP[2],width);

        for(int i = 0 ; i < 10 ; i ++ ) 
        {
            cudaMemcpy(P[ i ], P[ i ], allocatesize, cudaMemcpyDeviceToHost);
            cudaFree(devM[ i ]);   
            cudaFree(devP[ i ]);
        }

    }

我认为使用以下方法可能会实现最快的性能CUBLAS批量gemm函数 http://docs.nvidia.com/cuda/cublas/index.html#cublas-lt-t-gt-gemmbatched它是专门为此目的而设计的（执行大量“相对较小”的矩阵乘法运算）。

Even though you want to multiply your array of matrices (M[]) by a single matrix (N), the batch gemm function will require you to pass also an array of matrices for N (i.e. N[]), which will all be the same in your case.

EDIT:现在我已经完成了一个示例，对我来说很清楚，通过对下面的示例进行修改，我们可以传递一个N矩阵并有GPU_Multi函数只需发送单个N矩阵到设备，同时传递一个指针数组N, i.e. d_Narray在下面的示例中，所有指针都指向同一个N设备上的矩阵。

这是一个完整的批量 GEMM 示例：

#include <stdio.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <assert.h>

#define ROWM 4
#define COLM 3
#define COLN 5

#define cudaCheckErrors(msg) \
    do { \
        cudaError_t __err = cudaGetLastError(); \
        if (__err != cudaSuccess) { \
            fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
                msg, cudaGetErrorString(__err), \
                __FILE__, __LINE__); \
            fprintf(stderr, "*** FAILED - ABORTING\n"); \
            exit(1); \
        } \
    } while (0)


typedef float mytype;
// Pi = Mi x Ni
// pr = P rows = M rows
// pc = P cols = N cols
// mc = M cols = N rows
void GPU_Multi(mytype **M, mytype **N, mytype **P
  , size_t pr, size_t pc, size_t mc
  , size_t num_mat, mytype alpha, mytype beta)
{

    mytype *devM[num_mat];
    mytype *devN[num_mat];
    mytype *devP[num_mat];
    size_t p_size =sizeof(mytype) *pr*pc;
    size_t m_size =sizeof(mytype) *pr*mc;
    size_t n_size =sizeof(mytype) *mc*pc;
    const mytype **d_Marray, **d_Narray;
    mytype **d_Parray;
    cublasHandle_t myhandle;
    cublasStatus_t cublas_result;

    for(int i = 0 ; i < num_mat; i ++ )
    {
        cudaMalloc((void**)&devM[ i ], m_size );
        cudaMalloc((void**)&devN[ i ], n_size );
        cudaMalloc((void**)&devP[ i ], p_size );
    }
    cudaMalloc((void**)&d_Marray, num_mat*sizeof(mytype *));
    cudaMalloc((void**)&d_Narray, num_mat*sizeof(mytype *));
    cudaMalloc((void**)&d_Parray, num_mat*sizeof(mytype *));
    cudaCheckErrors("cudaMalloc fail");
    for(int i = 0 ; i < num_mat; i ++ ) {

        cudaMemcpy(devM[i], M[i], m_size , cudaMemcpyHostToDevice);
        cudaMemcpy(devN[i], N[i], n_size , cudaMemcpyHostToDevice);
        cudaMemcpy(devP[i], P[i], p_size , cudaMemcpyHostToDevice);
    }
    cudaMemcpy(d_Marray, devM, num_mat*sizeof(mytype *), cudaMemcpyHostToDevice);
    cudaMemcpy(d_Narray, devN, num_mat*sizeof(mytype *), cudaMemcpyHostToDevice);
    cudaMemcpy(d_Parray, devP, num_mat*sizeof(mytype *), cudaMemcpyHostToDevice);
    cudaCheckErrors("cudaMemcpy H2D fail");
    cublas_result = cublasCreate(&myhandle);
    assert(cublas_result == CUBLAS_STATUS_SUCCESS);
    // change to    cublasDgemmBatched for double
    cublas_result = cublasSgemmBatched(myhandle, CUBLAS_OP_N, CUBLAS_OP_N
      , pr, pc, mc
      , &alpha, d_Marray, pr, d_Narray, mc
      , &beta, d_Parray, pr
      , num_mat);
    assert(cublas_result == CUBLAS_STATUS_SUCCESS);

    for(int i = 0 ; i < num_mat ; i ++ )
    {
        cudaMemcpy(P[i], devP[i], p_size, cudaMemcpyDeviceToHost);
        cudaFree(devM[i]);
        cudaFree(devN[i]);
        cudaFree(devP[i]);
    }
    cudaFree(d_Marray);
    cudaFree(d_Narray);
    cudaFree(d_Parray);
    cudaCheckErrors("cudaMemcpy D2H fail");

}

int main(){

  mytype h_M1[ROWM][COLM], h_M2[ROWM][COLM];
  mytype h_N1[COLM][COLN], h_N2[COLM][COLN];
  mytype h_P1[ROWM][COLN], h_P2[ROWM][COLN];
  mytype *h_Marray[2], *h_Narray[2], *h_Parray[2];
  for (int i = 0; i < ROWM; i++)
    for (int j = 0; j < COLM; j++){
      h_M1[i][j] = 1.0f; h_M2[i][j] = 2.0f;}
  for (int i = 0; i < COLM; i++)
    for (int j = 0; j < COLN; j++){
      h_N1[i][j] = 1.0f; h_N2[i][j] = 1.0f;}
  for (int i = 0; i < ROWM; i++)
    for (int j = 0; j < COLN; j++){
      h_P1[i][j] = 0.0f; h_P2[i][j] = 0.0f;}

  h_Marray[0] = &(h_M1[0][0]);
  h_Marray[1] = &(h_M2[0][0]);
  h_Narray[0] = &(h_N1[0][0]);
  h_Narray[1] = &(h_N2[0][0]);
  h_Parray[0] = &(h_P1[0][0]);
  h_Parray[1] = &(h_P2[0][0]);

  GPU_Multi(h_Marray, h_Narray, h_Parray, ROWM, COLN, COLM, 2, 1.0f, 0.0f);
  for (int i = 0; i < ROWM; i++)
    for (int j = 0; j < COLN; j++){
      if (h_P1[i][j] != COLM*1.0f)
      {
        printf("h_P1 mismatch at %d,%d was: %f should be: %f\n"
          , i, j, h_P1[i][j], COLM*1.0f); return 1;
      }
      if (h_P2[i][j] != COLM*2.0f)
      {
        printf("h_P2 mismatch at %d,%d was: %f should be: %f\n"
          , i, j, h_P2[i][j], COLM*2.0f); return 1;
      }
    }
  printf("Success!\n");
  return 0;
}