[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: ATLAS version 3.2 available

Hi Clint!  Congratulations!  A few quick questions to assist with

For the purposes of the following, lets denote the atlas general
architectures as (i386,ppc,sparc,alpha).  There may be others less
common, and some on which atlas won't compile due to gcc (e.g. m68k
and arm, I think), but lets discuss this list first.  The
sub-architectures are
and maybe some others as well.

1) My understanding is that all gcc compiled code in atlas when
   compiled on any machine of a given general arch, will run on any
   other machine of that same general arch.  True?
2) My understanding is that the only atlas code which will run on some
   but not all members of a general arch are user-supplied asm
   routines.  True?  Current list? (i.e. of asm, as opposed to compiled
   user contributions?)
3) I've been informed that libs relying on a certain capability of the
   cpu can be installed under /usr/lib/xmm, or /usr/lib/amd3d for
   example, and that the linker will look there first if the cpu has
   that capability.  These flags are reported in /proc/cpuinfo under
   Linux.  Since you've compiled this release on more machines than
   anyone else, could you please share your understanding of which cpu
   capabilities substantially affect the atlas build?  In other words,
   apart from cache sizes, bus speeds, clocks, etc., which cpu
   capabilities significantly change the contents of the produced
   atlas library?

Take care,

R Clint Whaley <rwhaley@cs.utk.edu> writes:

> Dear ATLAS User  **or** Enraged Spam Victim:
> This is an announcement of ATLAS v3.2.  If you are uninterested, please
> excuse us for the imposition, and just skip down until you see the line:
>                      GETTING RID OF THIS KIND OF MAIL
> and you can ensure we never do it again.
> ATLAS v3.2 has just been posted to netlib.  The big ticket items for this
> release are:
>    (1) SMP support via pthreads in the Level 3 BLAS
>    (2) All ATLAS kernels have timing and testing infrastructure so that users
>        can contribute kernels (see ATLAS/doc/atlas_contrib.ps)
> There has already been substantial user contribution using (2) above, see
> ATLAS/doc/AtlasCredits.txt for details.  User-supplied kernels are the main
> source of significant speedups for this release.  There are a bunch of bug
> fixes, minor speedups (eg., Level 1 speedups, etc) and other minutia as well,
> of course.  On the more significant side, the following is a list of
> architectures that get Level 2 and/or Level 3 BLAS speedups with v3.2 over 3.0
> (not including SMP; obviously, any SMP machine will get further Level 3
>  speedup via pthreads):
>    * PentiumIII (SSE support)
>    * AMD Athlon (3DNow! support)
>    * All Compaq ev5x (21164) and ev6x (21264) machines
>    * All UltraSparc-based systems
>    * All PowerPC-based architectures
>    * IA64 Itanium
>    * IBM POWER3
> The official announcment to sources like na-digest won't be sent out until
> January sometime (to give us time to take vacation, and then update the
> website, etc).
> Cheers,
> The ATLAS Team
> *******************************************************************************
>                           NOTE FOR WINDOWS USERS
> *******************************************************************************
> This release has not been compiled under Windows yet.  We expect to make any
> necessary changes with the first patch,  and should have details on what
> has to be done to enable windows compilation sometime in January.
> *******************************************************************************
>                      GETTING RID OF THIS KIND OF MAIL
> *******************************************************************************
> If you are receiving this mail, it means you have mailed to the ATLAS help
> list, atlas@cs.utk.edu.  We searched an archive of this list to generate
> a list of people to mail to about ATLAS releases.  We expect to send around
> 1-4 e-mails per year to this list.  If you do not want to be on it, just
> send mail to atlas@cs.utk.edu, and in the body of the message, say:
>    unsubscribe atlas-announce
> If you received multiple copies of this mail, it probably means you sent in
> user questions from two different e-mail addresses (i.e. rwhaley@cs.utk.edu
> and rwhaley@cupid.cs.utk.edu).  You can either get rid of multiple inclusions
> one at a time by sending mail from the appropriate address with
>    unsubscribe atlas-announce
> in the body of the message, or you can get rid of them via
>    unsubscribe atlas-announce <addresses to nuke>
> We apologize for this rather spamlike way of starting the list.
> *******************************************************************************
>                      GETTING MORE OF THIS KIND OF MAIL
> *******************************************************************************
> If you are actively using ATLAS, you might be interested in being informed
> whenever a new bug has been discovered, instead of having to periodically scope
> the errata file.  If you send mail to atlas@cs.utk.edu, and in the body of
> the message say:
>    subscribe atlas-error
> We will ping you to scope the errata whenever an error is found.

Camm Maguire			     			camm@enhanced.com
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah