[Scilab-Dev] Switching linear algebra library on Scilab 5.2.2 binaries

[Vergnes Nicolas (STERIA)]

Ok
thank you very much for these informations !
Best regards;

Nicolas Vergnes Produits & Logiciels Cnes
3IRT/AP /POL

-----Message d'origine-----
De : Sylvestre Ledru [mailto:sylvestre.ledru at scilab.org] 
Envoyé : vendredi 11 juin 2010 12:02
À : dev at lists.scilab.org
Objet : Re: [Scilab-Dev] Switching linear algebra library on Scilab 5.2.2 binaries

Hello,

Le jeudi 10 juin 2010 à 17:26 +0200, Vergnes Nicolas (STERIA) a écrit :
> Hi,
> 
> actually I still can't use scilab 5.x on our linux 64b cluster (I think
> because of graphic problem) and the scilab binaries dont offer good
> performances with libatlas.so supplied with scilab 5.x binaries.
Actually, thanks for pointing this out. libatlas in the Scilab binaries
are not used. We are still using refblas which explains bad
performances.
Because we don't have an installer under Linux, it is hard to do like we
do with Windows (ie, detect the arch and use the best arch available)

> I have replaced libatlas.so.3gf.0, libblas.so.3gf.0, libf77blas.so.3gf.0
> and liblapack.so.3gf.0 files by an sym-link of the libacml_mp.so .
> Multithreading is used and we have quasi same performances than Matlab
> in matrix computation.
> 
> Say me if i'm wrong, but it is correct to say that atlas, gotoblas,
> acml, and mkl have the same routines names or so for the main routines ?
Yes, you should be able to replace one by the other. 
For your information, under Debian and Ubuntu, you can switch from any
blas & lapack implementations with the commands:
update-alternatives --config libblas.so.3gf 
update-alternatives --config liblapack.so.3gf

> With acml remplacing atlas, matrix multiplications are ok. Do you have
> any advice or examples files for testing it and see if all routines used
> by scilab are ok ?
> 
> Do you think it's really safe to use scilab configured like that ?
Yes, I think it is.

Sylvestre