topic Sparse matrix and dense matrix multiplication in Intel® oneAPI Math Kernel Library

Sparse matrix and dense matrix multiplication

Wei_W_ — Thu, 17 Jul 2014 10:08:50 GMT

Hi,

Have a quick question to ask, I need to do A x B where A is a sparse matrix stored in CSR format, B is a symmetric dense matrix. Now I am using csrmm to compute the multiplication, but have to convert B into full dense matrix first, which doubles the memory for B. Is there any routine in MKL that can handles this?

Many thanks!

Any help would be appreciated

Wei_W_ — Fri, 18 Jul 2014 13:08:00 GMT

Any help would be appreciated, thanks!

Some questions that are

mecej4 — Fri, 18 Jul 2014 14:55:46 GMT

Some questions that are relevant to the request:

How is the symmetric matrix B currently stored (before being converted to full, dense)?
What do you intend to do with A x B , after computing the product matrix?
How do you wish to store the product matrix?

Thanks for reply.

Wei_W_ — Fri, 18 Jul 2014 15:06:14 GMT

Thanks for reply.

1. The B matrix is stored column wise, packed form (upper triangle) before converting to full dense matrix.

2. After computing AxB=C, I will do a for loop i=1:n, for each i, I need to do D_i x C, where D_i is also a sparse matrix. Here I use csrmm for sure since the dense C matrix is not symmetric.

3. The product matrix C is not symmetric, so just the regular way to store it, we use column major to store dense matrix in our code.

Hi Wei,

Ying_H_Intel — Mon, 21 Jul 2014 01:26:42 GMT

Hi Wei,

It seems there is no good function can operate a spase matrix A*B (parcked upper triangle) form. You may either convert B to dense matrix. Then use mkl_?csrmm. Get sparse C matrix. Then call mkl_?csrmultd. to compute D_i *C.

Or convert B to sparse matrix. use mkl_?csrmultd. to compute C= A*B. Then mkl_?csrmm do D_i*C. But if convert B to sparse matrix. it still need double memory.

or call mkl_?csrgemv to compute ci=A*bi (B's colum vector). , then call csrmm do D_i*C.

Best Regards,

Ying