Bug#580335: Please compile multithreaded ATLAS libraries

[Lucas Nussbaum]

On 09/07/10 at 16:10 +0200, Sylvestre Ledru wrote:
> Le vendredi 09 juillet 2010 à 14:08 +0300, Lucas Nussbaum a écrit :
> > On 09/07/10 at 11:16 +0200, Sylvestre Ledru wrote:
> > > Hello Ryo,
> > > 
> > > 
> > > > I have been able to link my program with threaded ATLAS successfully as:
> > > > 
> > > > $ gcc test.c -L/usr/lib/atlas-corei7sse3 -lptf77blas -latlas -lgfortran
> > > > 
> > > > However, I cannot get any speed gain on our 8-core with hyperthreading machine.
> > > Yes, and I guess my answer will disappoint you.
> > > To allow your 8 cores to be used, you must rebuild Atlas (fakeroot
> > > debian/rules custom).
> > > 
> > > Unlike the MKL, the number of thread is static:
> > > http://math-atlas.sourceforge.net/faq.html#tnum
> > Are there performance implications with building it for too many cores?
> Yes, if you run a multithread (let's say planned for 4 cores) algo on a
> single core, it will run many threads (as a 4-core CPU would manage
> them). 
> Many performances improvements of Atlas are done at build time.

I'm wondering whether it really makes sense to spend so much time
working on optimizing atlas for different CPU architectures then. Maybe
it would be better to just ship a default version, and advertise more
how to rebuild the package with optimizations.
-- 
| Lucas Nussbaum
| lucas at lucas-nussbaum.net   http://www.lucas-nussbaum.net/ |
| jabber: lucas at nussbaum.fr             GPG: 1024D/023B3F4F |