Re: Linear Algebra Libraries
On Sat, Mar 20, 2010 at 06:00:04PM +0100, Sylvestre Ledru wrote:
> Le samedi 20 mars 2010 à 13:49 +0100, Marc 'HE' Brockschmidt a écrit :
> > Sylvestre Ledru <sylvestre@debian.org> writes:
> > > I am planning to upload some modifications in the way Linear Algebra
> > > Libraries BLAS / LAPACK and ATLAS are handle in Debian. These libraries
> > > are very used by many scientific software.
> > > They didn't have many attentions during the last few years and ATLAS in
> > > unstable is in a pretty bad shape (old version, plenty of bugs, hard to
> > > use ...)
> > > I described what I am planning to do in this wiki page:
> > > http://wiki.debian.org/DebianScience/LinearAlgebraLibraries
> > > and my changes have been tested.
> >
> > I have to admit that I'm a bit uncomfortable with using
> > update-alternatives to switch between libary implementations. I am not
> > aware of any other package doing this.
> This is done with the different MPI implementation (openmpi, mpich2,
> etc)
>
> > Is there a reason for not doing
> > it on the package level, allowing users to just install the package
> > fitting their needs and machine?
> Yes, it does improve the performance of software based this libraries.
>
> > Is it really common to switch
> > implementations on the same machine very often?
> Yes, it is. This kind of library are used by Scilab, R or Octave. They
> relies strongly on these libraries and for precision, confirmation of
> the result and performances, it is common to switch between the
> different implementation.
As Octave maintainer, I can confirm that switching the libraries is a
routine[1]. Having a reference implementation and some highly tuned (but
occasionally buggy) implementation is extremely valuable.
Example of such a case on upstream's list:
http://www-old.cae.wisc.edu/pipermail/bug-octave/2010-January/010231.html
Example in Debian:
http://bugs.debian.org/570231
[1] Actually, I'm pretty sad that ATLAS isn't allowed any more to ship a
configuration file in /etc/ld.so.conf.d/ .
Thomas
Reply to: