Could I use the `mxGPUArray` format data in the cuda kernel function?

I am new in compiling the mexcuda. I am trying to use the mxGPUArray as inputs of cuda kernel. My sample.cu code is as below. It is a simple plus function. The inputs are two vectors in CPU. The output is one vector in CPU.

But it gives the error at the line of

% z[row] = x[row] * y[row] line error
error: expression must be a pointer to a complete object type

Must I use the cudaMemcpy function, with transfering the mxGPUArray* data to a double* format?

#include "mex.h"
#include "gpu/mxGPUArray.h"
__global__ void plus(mxGPUArray*x, mxGPUArray* y, mxGPUArray* z, int N)
{
	int row = blockIdx.x*blockDim.x + threadIdx.x;
	if (row < N)
	{
		z[row] = x[row] * y[row];% error is here
	}
}

void mexFunction(int nlhs, mxArray *plhs[], int nrhs, mxArray const *prhs[])
{
    mxGPUArray* d_x = mxGPUCopyFromMxArray(prhs[0]);
    mxGPUArray* d_y = mxGPUCopyFromMxArray(prhs[1]);
    int N = (int)(mxGPUGetNumberOfElements(d_x));
    mxGPUArray* d_z = mxGPUCreateGPUArray(mxGPUGetNumberOfDimensions(d_x),
    mxGPUGetDimensions(d_x),
    mxGPUGetClassID(d_x),
    mxGPUGetComplexity(d_x),
    MX_GPU_DO_NOT_INITIALIZE);
    // =========================================================================
    // gpu computing
    // =========================================================================
    int const threadsPerBlock = 256;
    int blocksPerGrid;
    blocksPerGrid = (N + threadsPerBlock - 1) / threadsPerBlock;
    plus <<<blocksPerGrid, threadsPerBlock>>>(d_x, d_y, d_z,N);
    plhs[0] = mxGPUCreateMxArrayOnCPU(d_z);
    mxGPUDestroyGPUArray(d_x);
    mxGPUDestroyGPUArray(d_y);
    mxGPUDestroyGPUArray(d_z);
}

I don't want to do many copy steps. So I did not transfer the mxGPUArray into another double array. Any suggestion would be appreciated. Thank you.

Answer 1

[Joss Knight]

No. You have to get the GPU data pointer using the mxGPUGetData function.