Intel® oneAPI Math Kernel Library
Ask questions and share information with other developers who use Intel® Math Kernel Library.
The Intel sign-in experience has changed to support enhanced security controls. If you sign in, click here for more information.
6740 Discussions

from CSR to CSC for non-square (m x n) matrix


Hello everyone. I need to convert a non-square sparse matrix (m x n) from CSR format to CSC. But MKL has routines only for square matrix (m x m): mkl_?csrcsc. What can I try to make to complete this my own operation which is equivalent transpose operation for dense matrix? Thanks.   

0 Kudos
2 Replies

To obtain transpose of non-square matrix

I used "mkl_?csradd" function with trans = 'T' and zero-matrix.

This is not efficient. Any other way?


Black Belt

To convert from CSR to CSC or, which is equivalent, to find the CSR representation of AT given the CSR representation of the sparse rectangular matrix A of size m X n, one simple way involves (i) conceptually padding up the matrix with additional zero entries to a square matrix of size N X N where N = max (m,n), and (ii) calling the MKL converter. 

Note that, in the 3-array CSR representation, no changes need to be made to the values or columns  arrays (and, when m > n, to rowIndex) as a result of the padding-up, since none of the padding elements is non-zero. If m < n,  the rowIndex array should be padded up by replicating the prior last value, i.e., nnz+1.

If m and differ from each other so much that you are concerned about the wasted CPU cycles devoted to processing the additional zeros used to pad the matrix, you can write your own converter routine, noting that if log2 nnz is significantly less than max (m,n)  it may be worthwhile to sort the values and columns arrays by column indices before computing the 3-array representation of the transpose.