From help-octave-request at bevo dot che dot wisc dot edu  Sat Feb  5 05:46:02 2000
Subject: Re: ATLAS and octave
From: R Clint Whaley <rwhaley at cs dot utk dot edu>
To: jwe at bevo dot che dot wisc dot edu, keitt@nceas.ucsb.edu, rwhaley@cs.utk.edu
Cc: atlas at cs dot utk dot edu, help-octave@bevo.che.wisc.edu
Date: Sat, 5 Feb 2000 06:45:56 -0500

Guys,

As I was attempting to sleep, I realized I there might be a relatively
ease way to add ATLAS to octave.

All ATLAS internal routines begin with ATL_ (to avoid naming conflicts
with other libraries), and the rest of the defined routines are 
BLAS and LAPACK API routines.  This means, I think, that we can
put ATLAS into libcruft without too much problem.

The most direct method is, with a previously installed ATLAS, build 
octave's libcruft as normal, but you then ar into it the contents of:
   ATLAS/lib/<arch>/liblapack.a (overwriting some LAPACK routines)
   ATLAS/lib/<arch>/libatlas.a  (ATLAS internals)
   ATLAS/lib/<arch>/libf77blas.a (F77 BLAS interface -- overwrite BLAS)
   ATLAS/lib/<arch>/libcblas.a (C BLAS interface)

I have hacked up a quick version of this on my 600Mhz Athlon.  Everything
is not 100% I'm sure, but already I can see this:

Octave au natural:
===========================================================================
octave:1>  A = rand(1000,1000); 
octave:2> t1 = time(); lu(A) ; t2 = time() - t1
t2 = 30.460
octave:4> B = rand(1000,1000);
octave:5> t1 = time() ; A * B ; t2 = time() - t1
t2 = 41.757


My hacked up octave/atlas:
===========================================================================
octave:1> A = rand(1000,1000); 
octave:2> t1 = time(); lu(A) ; t2 = time() - t1
t2 = 5.1056
octave:7> B = rand(1000,1000);
octave:8> t1 = time() ; A * B ; t2 = time() - t1
t2 = 3.0831

So with trivial effort, we've got something running quite a bit faster.
All is not right with my hacked up octave, 'cause I'm not sure I got things
done in the right order.  I rather suspect I'm not using ATLAS's lapack
(I just hacked the libcruft/blas makefile and had it add all of atlas's
files).  I'm also not sure why I'm getting 324Mflop DGEMM on a platform
where `barebones' ATLAS DGEMM gets almost 700 . . .
However, the fact that the build succeeded tells me that ATLAS/octave are
playing together nicely . . .

Anyway, one easy approach is to formalize a process like this within the
octave framework.  Could be perhaps some kind of config option, or something.
If you are interested, I would be happy to supply the required
makefile / shell file / help needed  to shove the atlas stuff into
libcruft . . .

The cool think is that then ATLAS would not have to be incorperated into
octave, saving you from having to forward ATLAS queries on to us.  The
process for a normal octave remains the same.  For the high-octane version,
the user installs ATLAS first (sending mail to us if this doesn't work).
With the a good ATLAS install on the machine, you throw a configure option
in octave, and the faster stuff is built . . .

Anyway, the sun is about to come up, so it must be time for bed.  Let me
know what you think.

Cheers,
Clint



-----------------------------------------------------------------------
Octave is freely available under the terms of the GNU GPL.

Octave's home on the web:  http://www.che.wisc.edu/octave/octave.html
How to fund new projects:  http://www.che.wisc.edu/octave/funding.html
Subscription information:  http://www.che.wisc.edu/octave/archive.html
-----------------------------------------------------------------------