topic Re: Selective row/column to use for mat-mat and mat-vec multipl in Intel® oneAPI Math Kernel Library

Selective row/column to use for mat-mat and mat-vec multiplication

ssylee — Wed, 18 Feb 2009 19:17:18 GMT

Sorry to double-post in advance. I initially posted at http://icl.cs.utk.edu/lapack-forum/viewtopic.php?f=2&t=892 to ask about my question in the link. However, as I'm aware that there are probably more people using Intel or AMD math calculation tools when working with matrices, which includes Lapack, BLAS, and Sparse BLAS, I think I would have much better luck on asking for opinions here. It would be great if you have any tips on if there's a library function available for performing mat-mat, row-col, and mat-vec calculations while being able to select which rows/columns to multiply. The matrix is not densely sparse. It's more so randomly sparse along the matrix diagonal.

Re: Selective row/column to use for mat-mat and mat-vec multipl

Victor_Gladkikh — Thu, 19 Feb 2009 10:11:01 GMT

Quoting - ssylee

MKL BLAS level1 and MKL Sparse BLAS level 1 contain functions for vector-vector operations.

These routines can be used for performing multiplication of selected row of one matrix and a column of other matrix.

1.Dot product the i-th row of dense matrix A (m by n matrix stored in 2 dimensional FORTRAN array) and the j-th column of dense matrix B (n by n) can be computed with the help of "?dot" routine from BLAS level 1.

res = ddot(n, a(i, 1), m, b(1,j), 1)

2.Dot product the i-th row of sparse matrix A (m by n matrix stored in the CSR matrix format) and the j-th column of dense matrix B (n by n) can be computed with the help of "?doti" routine from Sparse BLAS level 1.

res = ddoti( pointere(i)-pointerb(i), values(pointerb(i)), columns(pointerb(i)), b(1,j))

See MKL Reference Manual for more details.

Re: Selective row/column to use for mat-mat and mat-vec multipl

ssylee — Thu, 19 Feb 2009 23:30:20 GMT

Quoting - Victor Gladkikh (Intel)

MKL BLAS level1 and MKL Sparse BLAS level 1 contain functions for vector-vector operations.

These routines can be used for performing multiplication of selected row of one matrix and a column of other matrix.

1.Dot product the i-th row of dense matrix A (m by n matrix stored in 2 dimensional FORTRAN array) and the j-th column of dense matrix B (n by n) can be computed with the help of "?dot" routine from BLAS level 1.

res = ddot(n, a(i, 1), m, b(1,j), 1)

2.Dot product the i-th row of sparse matrix A (m by n matrix stored in the CSR matrix format) and the j-th column of dense matrix B (n by n) can be computed with the help of "?doti" routine from Sparse BLAS level 1.

res = ddoti( pointere(i)-pointerb(i), values(pointerb(i)), columns(pointerb(i)), b(1,j))

See MKL Reference Manual for more details.

Would I need to extract a(i, 1) or b(1, j) in a vector manually first before calling the dot product function?

Re: Selective row/column to use for mat-mat and mat-vec multipl

Victor_Gladkikh — Fri, 20 Feb 2009 04:35:28 GMT

This is not actually needed.

Re: Selective row/column to use for mat-mat and mat-vec multipl

TimP — Fri, 20 Feb 2009 05:57:06 GMT

Quoting - ssylee

Would I need to extract a(i, 1) or b(1, j) in a vector manually first before calling the dot product function?

It seems that a Fortran dot_product would be appropriate.