0

I am developing my MEX file for sparse matrix computing with CUDA. I am using CUSP library. I don't know how to return cusp::csr_matrix back to Matlab. For example, I have 

cusp::csr_matrix<mwIndex,double,cusp::host_memory> At(m,n,N);

So, it is At matrix in CSR format, which, lets say, I have computed. Now, I have to return it to Matlab. Something like this:

plhs[0] = At;

But, of course, it doesn't work like that, firstly because At is on GPU. I guess I should use At.values and methods for indexes. But, also, how to assign them to host data?

Could somebody suggest how to do all that? :)

The Hiary
  • 109
  • 1
  • 13
Aurimas Šimkus
  • 159
  • 10

1 Answers1

0

Matlab stores sparse matrices in CSR format too, so it's not complicated. All you have to do is to allocate the sparse matrix using mxArray *mxCreateSparse(mwSize m, mwSize n, mwSize nzmax, mxComplexity ComplexFlag); and then setting the pr, ir, jc arrays (using mxGetPr, mxGetIr, mxGetJc). Pr corresponds to the values array in cusp, ir to column_indices and jc to row_offsets. If the matrix is in device memory, copy it using cudaMemcpy with cudaMemcpyDeviceToHost. Here is some examples using sparse matrices (its for Octave, but should work for Matlab as well).

Josef Borkovec
  • 1,069
  • 8
  • 13
  • Thank you. In addition, what I figured out, Thrust can be used for that in case of using CUSP. For example, device vector Xg can be returned to Matlab array T like that: mxArray *T = mxCreateDoubleMatrix(n, 1, mxREAL); double *x = mxGetPr(T); thrust::copy(xg.begin(), xg.end(), x); – Aurimas Šimkus Apr 09 '13 at 17:39