topic Thank you Fiona, that's in Intel® oneAPI Math Kernel Library

gemv for sparse, rectangular matrices?

damienhocking — Tue, 09 May 2017 04:13:05 GMT

Is there a recommended routine in the MKL that can do a GEMV with the transpose of a rectangular sparse matrix? I know I can pad it to square but I'd prefer not to have to do that. These sparse matrices could be large, 10 million+ nonzeroes and a million rows or columns.

Damien

Hi,

Zhen_Z_Intel — Tue, 09 May 2017 06:21:00 GMT

Hi,

You provided BLAS level 2 function for sparse matrix. For sparse matrix, we normally use compressed data structure that Intel MKL supports CSR, CSC, COO, diagonal and skyline. You could learn more from developer reference about usage with different matrix structure& data compression format:

https://software.intel.com/en-us/node/520802

Yes, but the docs are all for

damienhocking — Tue, 09 May 2017 06:35:23 GMT

Yes, but the docs are all for a square m x m matrix. I have rectangular matrices.

Hi,

Zhen_Z_Intel — Tue, 09 May 2017 07:14:11 GMT

Hi,

mkl_?<format>gemv is designed for square, but mkl_?<format>mv can be used for rectangular matrix (m * k). Or, you could use Inspector-execute sparse function mkl_sparse_?_mv, normally would provide a better performance.

Thank you Fiona, that's

damienhocking — Tue, 09 May 2017 15:20:31 GMT

Thank you Fiona, that's exactly what I wanted.