From: Eugen Leitl (eugen@leitl.org)
Date: Mon Apr 15 2002 - 00:24:32 MDT
Most users were reporting very poor performance, with a few notable
exceptions. SIMD on 16x 8 bit integers sounds extremely useful at least
for me.
---------- Forwarded message ----------
Date: Sun, 14 Apr 2002 21:57:56 -0400 (EDT)
From: William R. Pearson <wrp@alpha0.bioch.virginia.edu>
To: beowulf@beowulf.org
Subject: G4's for scientific computing
One of the advantages of the MacOSX gcc compiler is that in line
Altivec instructions are available at a high level. One can
define vector arrays, and do vector operations from 'C' code, e.g.
while(vec_any_gt(T2, NAUGHT)) {
T2 = vec_sub(LSHIFT(T2), RR);
FF = vec_max(FF, T2);
}
We are testing an Altivec FASTA version; a Altivec BLAST was announced
several months ago. We like Altivec because we can manipulate 8
16-bit integers or 16 8 bit integers at once - biological sequence
comparison code is essentially all integer. We see a 6-fold speedups
on when things are done 8-fold parallel. On our codes a dual 533 G4
and Altivec code is 6X-faster than a dual 1 GHz PIII (we don't have a
GHz G4 yet). Because of the high level Altivec primitives in the
Apple gcc compiler, vectorizing was very very easy; we would have to
be much more sophisticated to do the same thing on the PIII (and the
potential speed-up would be 1/2 as large, since the vector is 64, not
128 bits).
I might have agreed with the statement that one must have hand-tuned
Altivec code which pretty much excludes general purpose scientific
computing 4 months ago, but our experience has been very positive -
our programs are not specialized signal processing programs, but, in
retrospect, it was easy to get very dramatic speed up.
Bill Pearson
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:38 MDT