Keeping data in device memory via Matlab mex files.

Accelerated Computing CUDA CUDA Programming and Performance
#1

I’m trying to set up Matlab mex files that copy data to the device memory and return a pointer to the data (I’m doing estimations that operate on the same data many times, so want to avoid transferring the data with each call). My problem is I am unsure of how to return the pointer to Matlab (and then back again).

So far I have:

[codebox]#include <cuda.h>

#include “mex.h”

#include “matrix.h”

void mexFunction(int nargout, mxArray *argout, int nargin, const mxArray *argin)

{

float *input, *cuda_mem;

size_t size;

input = (float *)mxGetPr(argin[0]);

size = mxGetM(argin[0]) * mxGetN(argin[0]) * sizeof(float);

cudaMalloc((void **) &cuda_mem, size);

cudaMemcpy(cuda_mem, input, size, cudaMemcpyHostToDevice);

/* Now return the pointer to the data.*/

argout[0] = mxCreateNumericMatrix(1, 1, mxSINGLE_CLASS, mxREAL);

}[/codebox]

but I don’t know the type that the pointer to the single array on the device will be. Any “pointers” would be greatly appreciated.

#2

If you use the pointer in the same mex file afterwards you can just use a persistent variable. Otherwise my guess would be that you cast it to a int32.

Matlab 2010b has GPGPU support in the parallel computing toolbox where you can do all of this from matlab without the need to write mex files. I have not tried it yet as 2010b is one day old and this functionality was not yet in the beta version, but it sounds very promising.

#3

If you use the pointer in the same mex file afterwards you can just use a persistent variable. Otherwise my guess would be that you cast it to a int32.

Matlab 2010b has GPGPU support in the parallel computing toolbox where you can do all of this from matlab without the need to write mex files. I have not tried it yet as 2010b is one day old and this functionality was not yet in the beta version, but it sounds very promising.