[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Getfem-users] Gmm matrix multiply accuracy issue
From: |
Yves Renard |
Subject: |
Re: [Getfem-users] Gmm matrix multiply accuracy issue |
Date: |
Thu, 20 Oct 2011 16:07:45 +0200 |
User-agent: |
KMail/1.13.5 (Linux/2.6.32-34-server; KDE/4.4.5; x86_64; ; ) |
Dear Umut,
Is it possible that the error comes from the MUMPS interface ?
What kind of sparse matrix do you use ?
Yes, you can send me the program, I will try to check.
Yves.
On mercredi 19 octobre 2011, Umut Tabak wrote:
> On 10/19/2011 10:29 AM, Umut Tabak wrote:
> > Dear all,
> >
> > I have run into a problem today with the gmm++ library. I am not very
> > sure where the problem could be. I coded a symmetric lanczos solver
> > for my problem and used gmm++ and MUMPS (as the linear solver).
> > However, the accuracy of the basic linear algebra operations
> > matrix-vector product and gmm::add, scale and so seem to be
> > questionable to me.
> >
> > Well it could be a bit cryptic to understand for you, maybe, but let
> > me rephrase with an example:
> >
> > I am computing some coefficients to form a tridiagonal matrix and
> > these coefficients that I compute in gmm++ start to deviate from the
> > ones I computed in MATLAB. They seem to be the same at the starting
> > iterations then conducting these matrix-vector operations(for
> > orthogonalizations), they start to diverge, I am not sure what the
> > problem could be.
> >
> > I went over my code several times, however, the end results tell me
> > that there is indeed a problem in the orthogonalization phase, which
> > is basically a combination of matrix-vector operations.
> >
> > Any comments are appreciated on this issue.
> >
> > Best regards,
> > Umut
>
> As a reply to my own post, there is definitely a problem in the linear
> algebra operations in gmm++ from accuracy point of view.
>
> I tested my problem with MTL4 and the results seem to be far more close
> to what they should be.
>
> I am not sure what the problem is in gmm++ but this deserves some
> attention.
>
> I can provide a test case in MATLAB and gmm++ where I compare these
> tridiagonal coefficients of the lanczos proces. To test the MATLAB code,
> one should have the MUMPS MATLAB interface installed.
> Best regards,
> Umut
--
Yves Renard (address@hidden) tel : (33) 04.72.43.87.08
Pole de Mathematiques, INSA-Lyon fax : (33) 04.72.43.85.29
20, rue Albert Einstein
69621 Villeurbanne Cedex, FRANCE
http://math.univ-lyon1.fr/~renard
---------