Page 1 of 1

Matrix equilibration

PostPosted: Tue Apr 30, 2013 1:24 am
by coruun

I wanted to try to convert some matrix calculations from MKL to CUDA. My problem involves some matrix multiplications and additions and finally the solution of a linear equation system.

Unfortunately, the condition of the matrix seems to be very high and the solution of the equation system with MKL's zgesv was unsuccessful. As a result of this, I switched to the expert driver zgesvx. This routine equilibrates the matrix and solves the resulting equation system.

So my question is, if there is a possible workaround or any plans to include routines for matrix equilibration (especially the routine ?gesvx or the general routine ?geequ would be of interest).

Best Regards,

Re: Matrix equilibration

PostPosted: Fri May 10, 2013 11:29 am
by coruun
Update: I wrote my own kernels to perform matrix equilibration (?geequ) and iterative refinement (?getrs) to recreate LAPACK's ?gesvx.

Re: Matrix equilibration

PostPosted: Thu Dec 12, 2013 10:29 pm
by siamon

I would like to try CULA Sparse in an internally developed CFD code.
I tried the Academic version but our institution is not in you list of recognized academic institutions.

Re: Matrix equilibration

PostPosted: Thu Oct 16, 2014 5:02 am
by rdokoye
I guess there’s no legitimate alternative to converting matrix calculations from MKL to CUDA. But I’m going to continue looking, myself or just take the high leap.