Re: Atlas proposal
On Wed, Aug 18, 2010 at 01:59:59PM +0200, Sylvestre Ledru wrote:
> I am more concern about advanced users of scientific computing software
> (Scilab, R, Octave...) which are familiar with such tools but not
> familiar enough with the internals.
> They just see these software as a whole and would not guess that
> changing RefBLAS => Atlas could improve the performances to a 40
> factor...
Oh, that's not a problem in practice. They will be told so when asking
on the mailing lists of those projects about the speed. Apart from that,
putting ATLAS into Recommends: has served fine, at least for the Octave
packages in Debian.
> Here, at Scilab (and other people at the DebConf reported me the same
> experiences), it is not rare that people are complaining about the speed
> of the software because RefBLAS is used as the linear algebra library.
That's because you are a bad, bad boy ;) and don't recommend at least the
libatlas3gf-base package.
Personally, I would prefer to have just one compiled ATLAS package with
support for different instruction sets. Yes, this doesn't cater for
different cache sizes, but there's a limit on what we as distribution
can provide.
I don't have any preference about the -auto package for local
compilation.
Thomas
Reply to: