Markus Dittrich wrote:
> On Tue, 10 Jul 2007, Adam Pityszek wrote:
>> Hi All,
> 
>> I am forwarding my email to Sebastien regarding the merging of blas- and
>> cblas-reference into one package. Sebastien accepted this idea, but we
>> would like to hear your opinions as well, since you are the real users
>> of scientific libraries in Gentoo.
> 
>> BR,
>> /Adam
> 
> 
> This sounds like a fine plan to me, particularly since most
> blas implementations in portage do provide both blas and cblas
> at the same time. In addition, this would render the eselect framework
> less complicated which I find desirable.
> 
> cheers,
> Markus
> 
> 
> -- Markus Dittrich (markusle)
> Gentoo Linux Developer
> Scientific applications

I think the general scenario for users of BLAS and LAPACK is that they
will be happy with this. However, some people, especially in the
high-performance area, will want to be able to benchmark various
versions (ATLAS, GOTO or chip-specific Intel/AMD) and either static-link
to the winner or dynamic-link to the winner on an individual calling
application basis.

I haven't gotten around to doing that yet, but it's on my list for my
AMD64 dual-core system. I'm not sure the "eselect" facility provides for
this. And a number of packages carry their own BLAS or LAPACK. R is one
of them; in fact

1. R allows you to build a shared library of the BLAS and LAPACK that it
carries, and
2. The R team strongly suggests using the LAPACK that R carries rather
than an external one, and indeed, the "make check-all" tests fail on
some architectures with some external versions of LAPACK. On my Athlon
T-Bird, this happened with the Gentoo lapack-atlas and the R team
slapped me silly when I complained about it on their mailing list. ;)

So as long as I can do all of these things:

1. Choose between static and dynamic linking to BLAS/LAPACK with a USE
flag on any application that allows such a choice,
2. Choose between internal BLAS/LAPACK with a USE flag on any
application that allows such a choice, and
3. Choose between the various versions for the highest performance on an
individual application basis

I'm happy. I think most of the ebuilds are already set up this way, but
again, I haven't had a chance to test with really big problems on my
really big machine yet.
-- 
[EMAIL PROTECTED] mailing list

Reply via email to