error using cublasSgemmBatched in mex - matlab

Question

I am trying to multiply Multiple Matrices from within matlab using cublasSgemmBatched in a mex file.
My matlab code is very simple:
gpuDevice(1);
a = single(rand(400,10,1500,'gpuArray'));
b = single(rand(10,12,1500,'gpuArray'));
c = MatCuda(a,b)
I get the following error:
Error using gpuArray/subsref
An unexpected error occurred during CUDA execution. The CUDA error was:
unknown error
and here's the mexFunction code:
void mexFunction( int nlhs, mxArray *plhs[],
int nrhs, const mxArray *prhs[]){
char const * const errId = "parallel:gpu:mexGPUExample:InvalidInput";
char const * const errMsg = "Invalid input to MEX file.";
/* Declare all variables.*/
mxGPUArray const *A;
mxGPUArray const *B;
mxGPUArray *C;
/* Initialize the MathWorks GPU API. */
mxInitGPU();
/* Throw an error if the input is not a GPU array. */
if ((nrhs != 2) || !(mxIsGPUArray(prhs[0])) || !(mxIsGPUArray(prhs[1]))) {
mexErrMsgIdAndTxt(errId, errMsg);
}
A = mxGPUCreateFromMxArray(prhs[0]);
B = mxGPUCreateFromMxArray(prhs[1]);
if ((mxGPUGetClassID(A) != mxSINGLE_CLASS) || (mxGPUGetClassID(B) != mxSINGLE_CLASS)) {
mexErrMsgIdAndTxt(errId, errMsg);
}
float const *d_A;
float const *d_B;
d_A = (float const *)(mxGPUGetDataReadOnly(A));
d_B = (float const *)(mxGPUGetDataReadOnly(B));
const mwSize *dimsA = mxGPUGetDimensions(A);
size_t nrowsA = dimsA[0];
size_t ncolsA = dimsA[1];
size_t nMatricesA = dimsA[2];
mxFree((void*) dimsA);
const mwSize *dimsB = mxGPUGetDimensions(B);
size_t nrowsB = dimsB[0];
size_t ncolsB = dimsB[1];
size_t nMatricesB = dimsB[2];
mxFree((void*)dimsB);
size_t nrowsC = nrowsA;
size_t ncolsC = ncolsB;
mwSize dimsC[3] = { nrowsA, ncolsB, nMatricesB };
C = mxGPUCreateGPUArray(mxGPUGetNumberOfDimensions(A),
dimsC,
mxGPUGetClassID(A),
mxGPUGetComplexity(A),
MX_GPU_DO_NOT_INITIALIZE);
float *d_C;
d_C = (float *)(mxGPUGetData(C));
cublasHandle_t handle;
cublasStatus_t ret;
ret = cublasCreate(&handle);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCreate returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
const float alpha = 1.0f;
const float beta = 0.0f;
ret = cublasSgemmBatched(handle, CUBLAS_OP_N, CUBLAS_OP_N, nrowsA, ncolsB, ncolsA, &alpha, &d_A, nrowsA, &d_B, nrowsB, &beta, &d_C, nrowsC, nMatricesA);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasSgemm returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
ret = cublasDestroy(handle);
if (ret != CUBLAS_STATUS_SUCCESS)
{
printf("cublasCreate returned error code %d, line(%d)\n", ret, __LINE__);
exit(EXIT_FAILURE);
}
plhs[0] = mxGPUCreateMxArrayOnGPU(C);
mxGPUDestroyGPUArray(A);
mxGPUDestroyGPUArray(B);
mxGPUDestroyGPUArray(C);
}
I suspect it's related to the function cublasSgemmBatched because when I remove it from the code then I don't get the error.
Help will be very much appreciated !
Thanks !

Answer 1

You don't need a MEX file here, you can use pagefun to do this. Also, I'd recommend building a and b in single precision directly rather than casting. In other words,
a = rand(400,10,1500,'single','gpuArray');
b = rand(10,12,1500,'single','gpuArray');
c = pagefun(#mtimes, a, b);

Answer 2

cublasDgemm works for me. I just pass regular arrays to the mexfunction. Here is my example code.
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <time.h>
#include "mex.h"
#include "mxGPUArray.h"
typedef struct _matrixSize
{
unsigned int uiWA, uiHA, uiWB, uiHB, uiWC, uiHC;
} sMatrixSize;
void matrixMultiply(double const* d_A, double const* d_B, double* d_C, sMatrixSize &matrix_size);
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, mxArray const *prhs[])
{
mxGPUArray const *A;
mxGPUArray const *B;
mxGPUArray *C;
_matrixSize matrix_size;
mwSize const *A_sz;
mwSize const *B_sz;
double const *d_A;
double const *d_B;
double *d_C;
char const * const errId = "parallel:gpu:mexGPUExample:InvalidInput";
char const * const errMsg = "Invalid input to MEX file.";
if (nrhs != 2) {
mexErrMsgTxt("Need two inputs");
}
A = mxGPUCreateFromMxArray(prhs[0]);
B = mxGPUCreateFromMxArray(prhs[1]);
A_sz=mxGPUGetDimensions(A);
B_sz = mxGPUGetDimensions(B);
matrix_size.uiWA = (unsigned int)A_sz[0]; matrix_size.uiHA = (unsigned int)A_sz[1];
matrix_size.uiWB = (unsigned int)B_sz[0]; matrix_size.uiHB = (unsigned int)B_sz[1];
mwSize C_sz[3] = { matrix_size.uiWA, matrix_size.uiHB, 1 };
d_A = (double const *)(mxGPUGetDataReadOnly(A));
d_B = (double const *)(mxGPUGetDataReadOnly(B));
C = mxGPUCreateGPUArray(mxGPUGetNumberOfDimensions(A),
C_sz,
mxGPUGetClassID(A),
mxGPUGetComplexity(A),
MX_GPU_DO_NOT_INITIALIZE);
d_C = (double *)(mxGPUGetData(C));
matrixMultiply(d_A, d_B, d_C, matrix_size);
plhs[0] = mxGPUCreateMxArrayOnGPU(C);
mxFree((void*)A_sz);
mxFree((void*)B_sz);
mxGPUDestroyGPUArray(A);
mxGPUDestroyGPUArray(B);
mxGPUDestroyGPUArray(C);
}
void matrixMultiply(double const* d_A, double const* d_B, double* d_C, sMatrixSize &matrix_size)
{
cublasStatus_t status;
cublasHandle_t handle;
status=cublasCreate(&handle);
if (status != CUBLAS_STATUS_SUCCESS)
{
if (status == CUBLAS_STATUS_NOT_INITIALIZED) {
mexPrintf("CUBLAS initializing error");
}
getchar();
return;
}
const double alpha = 1.0f;
const double beta = 0.0f;
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, matrix_size.uiWB, matrix_size.uiHA, matrix_size.uiWA, &alpha, d_B, matrix_size.uiWB, d_A, matrix_size.uiWA, &beta, d_C, matrix_size.uiWB);
cudaThreadSynchronize();
cublasDestroy(handle);
}
There is also a more native cuda style. Both works fine for me.
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <time.h>
#include "mex.h"
#include "mxGPUArray.h"
typedef struct _matrixSize
{
unsigned int uiWA, uiHA, uiWB, uiHB, uiWC, uiHC;
} sMatrixSize;
void matrixMultiply(double const* d_A, double const* d_B, double* d_C, sMatrixSize &matrix_size);
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, mxArray const *prhs[])
{
mxArray const *mA;
mxArray const *mB;
_matrixSize matrix_size;
size_t A_w, A_h, B_w, B_h;
double *d_A;
double *d_B;
double *d_C;
double *h_A;
double *h_B;
double *h_C;
char const * const errId = "parallel:gpu:mexGPUExample:InvalidInput";
char const * const errMsg = "Invalid input to MEX file.";
if (nrhs != 2) {
mexErrMsgTxt("Need two inputs");
}
mA = prhs[0]; mB = prhs[1];
A_w = mxGetM(mA);A_h = mxGetN(mA);B_w = mxGetM(mB);B_h = mxGetN(mB);
matrix_size.uiWA = (unsigned int)A_w; matrix_size.uiHA = (unsigned int)A_h;
matrix_size.uiWB = (unsigned int)B_w; matrix_size.uiHB = (unsigned int)B_h;
matrix_size.uiWC = (unsigned int)A_w; matrix_size.uiHC = (unsigned int)B_h;
mwSize const C_sz[3] = { matrix_size.uiWA, matrix_size.uiHB, 1 };
unsigned int size_A = matrix_size.uiWA * matrix_size.uiHA;
unsigned int mem_size_A = sizeof(double) * size_A;
h_A = (double*)mxGetData(mA);
unsigned int size_B = matrix_size.uiWB * matrix_size.uiHB;
unsigned int mem_size_B = sizeof(double) * size_B;
h_B = (double*)mxGetData(mB);
unsigned int size_C = matrix_size.uiWC * matrix_size.uiHC;
unsigned int mem_size_C = sizeof(double) * size_C;
plhs[0] = mxCreateNumericArray(3, C_sz, mxDOUBLE_CLASS, mxREAL);
h_C = (double*)mxGetData(plhs[0]);
cudaMalloc((void **)&d_A, mem_size_A);
cudaMalloc((void **)&d_B, mem_size_B);
cudaMemcpy(d_A, h_A, mem_size_A, cudaMemcpyHostToDevice);
cudaMemcpy(d_B, h_B, mem_size_B, cudaMemcpyHostToDevice);
cudaMalloc((void **)&d_C, mem_size_C);
matrixMultiply(d_A, d_B, d_C, matrix_size);
cudaMemcpy(h_C, d_C, mem_size_C, cudaMemcpyDeviceToHost);
cudaThreadSynchronize();
cudaFree(d_A);
cudaFree(d_B);
cudaFree(d_C);
}
void matrixMultiply(double const* d_A, double const* d_B, double* d_C, sMatrixSize &matrix_size)
{
cublasHandle_t handle;
cublasCreate(&handle);
const double alpha = 1.0f;
const double beta = 0.0f;
cublasDgemm(handle, CUBLAS_OP_N, CUBLAS_OP_N, matrix_size.uiWB, matrix_size.uiHA, matrix_size.uiWA, &alpha, d_B, matrix_size.uiWB, d_A, matrix_size.uiWA, &beta, d_C, matrix_size.uiWB);
cublasDestroy(handle);
}