Re: blas, lapack, slatec,scalapack
On Sat, 1 Aug 1998, John Lapeyre wrote:
> These are some of the more popular libraries from netlib
> (according to access statistics) . Lapack is packaged for debian already.
> I have been looking at the others. slatec did not have any makefiles or
> nice structure. I added enough to build static and shared libs.
Yes, BLAS and LAPACK are pretty important in scientific computation (and
many packages rely on BLAS), so it is important to get these ones right. I
am just beguinning to use Debian for a Linux cluster (Beowulf-like), and I
hope to help testing/packaging these in the future, if needed.
It should be noted that there are two branches of assembler-optimized BLAS
for Pentiums & P-II under Linux:
1) http://cip.physik.uni-wuerzburg.de/~mlkessle/blas1.html
2) http://www.cs.utk.edu/~ghenry/distrib/
(1) Is the project of M Kessler, and seems to be all GPL. But only level 1
BLAS are finished. (2) will probably be much more powerful and complete,
since these come directly from Intel people working on the ASCI Option Red
Supercomputer at Sandia. But these cannot be distributed freely, you have
to register (although they're "gratis").
So maybe the packager/mantainer of the BLAS would have to take these
versions of the blas into account too. (to plan for different names or
whatever)
> But there are a few routines in slatec which contain machine
> specific data on word size, etc. I am not sure what values to enter for
> these. One is supposed to set these in the blas source as well. However,
> in the debian lapack source, which includes blas, I can not find any trace
> of the machine dependent files. I don't use the library, so I don't know
> if it's useful without these. Does anyone have information on this ?
> scalapack, which appears to be actively developed relies on
> an installed blas library. Perhaps we could build a separate blas library
> from the lapack source (the stand alone blas source is a mess) , or maybe
> make a sym link to liblapack for programs that want libblas.
It is almost two years since I compiled and installed the whole lapack
package on a Sun (where we didn't have Sun's assembler tuned version,
libperf), and I vaguely remember that prior to the compiling stage the
instalation probed for the machine parameters (routine "machar" from
netlib, which comes with the BLAS). I'm not sure, but anyway I don't
remember having to input values for the machine floating-point parameters.
What I know is that I came up with *two* different static libs, libblas.a
and liblapack2.a.
Do you mean that the current lapack package in Debian bundles the blas
inside lapack2.a? In that case I'd agree that it's better to have them
separate. Otherwise, how could one use the lapack and one of the above
assembler blas?
Jose
=============================================================================
Jose L. Marín jose@ma.hw.ac.uk
Dept of Maths marin@wanda.unizar.es
Heriot-Watt University
Edinburgh EH14 4AS, U.K.
Phone: +44 131 4513230
Fax: +44 131 451 3249
Former address: Dept. de Física de la Materia Condensada
Facultad de Ciencias, Universidad de Zaragoza
50009 Zaragoza, SPAIN
=============================================================================
--
To UNSUBSCRIBE, email to debian-devel-request@lists.debian.org
with a subject of "unsubscribe". Trouble? Contact listmaster@lists.debian.org
Reply to: