From octave-sources-request at bevo dot che dot wisc dot edu Fri Feb 4 17:50:36 2000 Subject: Re: ATLAS and octave From: "Timothy H. Keitt" To: R Clint Whaley CC: octave-sources at bevo dot che dot wisc dot edu, atlas@cs.utk.edu Date: Fri, 04 Feb 2000 15:50:21 -0800 Sounds good to me. If atlas is a drop in replacement for blas and lapack, then this could all be handled in a configure script. (I'm assuming here that the octave's blas and lapack routines are the same as the standard ones, so no recoding would be necessary.) I've actually never quite understood why octave includes blas and lapack code. It seems to defeat the purpose of having standard (and tunable) external libraries. (At best it duplicates some code; at worst its an unofficial code fork.) Cheers, Tim R Clint Whaley wrote: > Hi, > > First, let me say octave is a beautiful package, and it is a real lifesaver > for a linear algebra guy like myself . . . Also, I hope this is the correct > mailing list to bring up the following idea: > > We have a free software project known as ATLAS (www.netlib.org/atlas) which > provides optimized BLAS for pretty much arbitrary architectures. These > BLAS may be orders of magnitude faster than the reference BLAS. The idea is > that it would be nice if a user could use ATLAS, or some other optimized BLAS, > with octave, thus speeding up all of the lapack-handled routines significantly. > I realize speed is not always important (eg, small problems), but some engineers > use matlab, for instance, to solve quite large problems, where this speedup > would be invaluable. > > ATLAS takes a while to install, so I doubt you are interested in including > it in octave, but I thought you might be interesting in facilitating a > user's use of optimized BLAS in order to see these kinds of speedups. > I hacked a previous version of octave to use ATLAS quite a while ago, but > I doubt your average user would be able to do so unless it were made a > little easier. > > ATLAS presently includes all the BLAS, and LAPACK's LU and Cholesky > factorization routines. To give you an idea of the kinds of speedups I'm > talking about, I include a few timings below. These timings are for a > previous release of ATLAS (which did not have Level 2 or 1 BLAS), but they > can give you an idea of the speedups to be had. I show the MFLOP rate > for a 500x500 LU factorization, using various configurations an a > 266Mhz PII and a 533Mhz DEC ev56: > > PII266 533 ev56 > ======= ======== > ATLAS LU & BLAS : 157 441 > LAPACK LU, ATLAS BLAS : 143 330 > LAPACK LU, reference BLAS : 68 60 > > Anyway, if you are interested in pursuing this, please let us know at > atlas at cs dot utk dot edu dot > > Thanks, > Clint -- Timothy H. Keitt National Center for Ecological Analysis and Synthesis 735 State Street, Suite 300, Santa Barbara, CA 93101 Phone: 805-892-2519, FAX: 805-892-2510 http://www.nceas.ucsb.edu/~keitt/