topic Sparse Blas Matrix Multiplication in Intel® oneAPI Math Kernel Library

Sparse Blas Matrix Multiplication

Dan4 — Fri, 25 Sep 2009 11:17:27 GMT

Hi,

I have a problem using MKL Sparse Blas Level 3 routine "mkl_dcoomm" which calculates C=A*B where A is a sparse matrix in coordinate format and B is dense matrix. Clearly C is also a dense matrix. The problem is that I would like to perform C=A*B where B is sparse and A is dense (reverse of former case). Is there any other function for this (I didn't find!) or I should change my expression and use same matrix? My question might be very elementary, so sorry in advance!

Thanks for your help,

D.

Re: Sparse Blas Matrix Multiplication

jaewonj — Fri, 25 Sep 2009 14:41:00 GMT

Quoting - danltu.se

This might work.

C^T = (A*B)^T = B^T * A^T

B^T : set transa = 'T'
A^T : store A in row-major order, or transpose A if it is stored in column-major order
C^T : treat C as C-style dense array, or transpose C if you need column-major order

Re: Sparse Blas Matrix Multiplication

Dan4 — Fri, 25 Sep 2009 17:28:30 GMT

Quoting - jaewonj

Thank you janwonj for your answer. It can be a solution, but it sounds weired to me if MKL does not support such simple multiplication. Isn't really any suitable function to perform DENSE x SPARSE operation in MKL?

Re: Sparse Blas Matrix Multiplication

jaewonj — Sat, 26 Sep 2009 17:47:46 GMT

Quoting - danltu.se

As far as I know MKL does not directly support dense * sparse BLAS level 3 function. I know it's not pleasant, but I think we need to compute C' = (A*B)'.

Re: Sparse Blas Matrix Multiplication

Victor_Gladkikh — Mon, 28 Sep 2009 05:48:25 GMT

Quoting - danltu.se

Anyway there are different ways to used MKL, but it depend to problem that required computation of DENSE matrix C

First way: Size of SPARSE matrix is enough small, then you could convert from COO to DENSE and use MKL dgemm.

Second way: Your interest is compute product y = C*x (x is vector) or Y=C*X (X is DENSE matrix) then you could used MKL by next way:

compute z = B * x (use mkl_dcoomv) and compute y= A * z (use dgemv)

compute Z = B * X (use mkl_dcoomm) and compute Y= A * Z (use dgemm)