Denormalized values and accuracy

Post here if you have a question about LAPACK performance

Denormalized values and accuracy

Postby ldm001 » Sun May 04, 2014 10:38 am

I am looking for advice about how to improve scalapack accuracy when denormalized values matter.

I have a large code doing eigenvalue calculations using scalapack which I have determined is sensitive to how denormalized values are handled. If "standard" flush-to-zero (ftz) is used it is slightly faster (a few percent) but the accuracy is lower. Since these eigenvalues are later used (after some transformations) within a quasi-Newton (QN) code, the loss of accuracy slows down the convergence of QN iterations so the nett effect of ftz is to slow down the overall code, and the QN may not converge well.

I am using Intel's mkl version, and have turned down the optimization somewhat and I am also setting the FPE hardware flags in the software. I am wondering if there is anything within scalapack itself that I might be able to exploit.

N.B., of course one option is to do some rescaling within the code itself. Unfortunately it is a rather big code written by someone else, and finding the "right" method that does not break something else is a different project by itself.

N.N.B., if relevant I can post the relevant parts of the code (Fortran)
ldm001
 
Posts: 7
Joined: Tue Sep 16, 2008 10:21 pm

Return to Performance

Who is online

Users browsing this forum: No registered users and 2 guests