This is the mail archive of the
gsl-discuss@sources.redhat.com
mailing list for the GSL project.
Re: [Fwd: QR-Factorization in gsl (fwd)]
- From: Brian Gough <bjg at network-theory dot co dot uk>
- To: Joerg Wensch <wensch at informatik dot uni-halle dot de>
- Cc: gsl-discuss at sources dot redhat dot com
- Date: Wed, 24 Mar 2004 11:08:41 +0000
- Subject: Re: [Fwd: QR-Factorization in gsl (fwd)]
- References: <1079983057.5496.16.camel@bellerophon.lanl.gov>
> From: Joerg Wensch <wensch@informatik.uni-halle.de>
> To: jungman@lanl.gov
> Subject: QR-Factorization in gsl (fwd)
>
> we were using the QR-decomposition written by you
> for the gsl. We recogniced that the row-oriented
> format decreases performance considerably for larger
> matrices. We had a 10000 by 160 matrix were
> the composition took very long (about 45 sec.).
>
> It is clear to that it is a lot of extra work
> to allow both formats (row+col major)
> in the matrix class. I propose a simple workaroud:
> To add besides an QR-decomposition an LQ-decomposition too.
>
> I have used an LQ-decomposition for the large matrix
> (now in 160 times 10000 format) which gave a speedup of
> 13 on our architecture.
Thanks for the suggestion! I will look at that.
One thing I know is that the 'hm' routines in householder.c could also
benefit from using BLAS (although that is not a substitute for using
the right algorithm, of course).
--
Brian Gough
Network Theory Ltd
Tel: +44 (0)117 3179309
Fax: +44 (0)117 9048108
Web: http://www.network-theory.co.uk/