Xref: utzoo comp.lang.fortran:4292 sci.math.num-analysis:1346
Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!samsung!sdd.hp.com!zaphod.mps.ohio-state.edu!ncar!gatech!uflorida!ufqtp!bernhold
From: bernhold@qtp.ufl.edu (David E. Bernholdt)
Newsgroups: comp.lang.fortran,sci.math.num-analysis
Subject: Sparse BLAS1 (SUMMARY)
Message-ID: <1202@orange19.qtp.ufl.edu>
Date: 4 Dec 90 20:35:03 GMT
Reply-To: bernhold@qtp.ufl.edu (David E. Bernholdt)
Organization: University of Florida Quantum Theory Project
Lines: 51
About 10 days ago, I sent out a request for information on sparse
BLAS-1 implementations. Here is a summary of the response...
There is a paper by Dodson, Grimes and Lewis (DGL) describing a set of
sparse BLAS-1 routines. The paper and model routines are available
from netlib. I understand it will be appearing in the ACM
Transactions on Mathematical Software as well. This is the most
"popular" such definition of which I am aware.
These routines have been implemented in the NAG Mark 14 Fortran
library.
The IBM ESSL library implements 10 of the routines -- the single and
double precision, but not the comple or double complex.
Cray's SCILIB has several sparse BLAS-1 type operations. They have
different names and arguments from the Dodson, Grimes and Lewis
definitions.
I have heard two different rumors about Cray implementing the DGL
definitions: From one source that they have been implemented, but
from another that they are not scheduled for release to the public.
In any event, it looks like the DGL proposal is begining to catch on
with the vendors.
As I suspected, there is little concensus on what "sparse blas" should
be. I think it is fairly clear for BLAS-1, but much less so for
higher levels. I get the feeling that most people working with sparse
problems are "rolling their own" basic routines. Some people
expressed doubt that sparse versions of the higher-level BLAS would ever
catch on.
That being said, I should note that the paper "Are there iterative
BLAS?" by Oppe and Kincaid includes a slightly different approach
sparse level-1 BLAS (as part of what they call "iterative BLAS") --
more general and somewhat more flexible than the DGL proposal, but
with basically the same operations available.
The Oppe & Kincaid paper is the only other proposal which received any
mention. They are aiming at a fundamental set of routines for the
development of sparse iterative solvers, so they include numerous
routines aimed at different common storage/problem structures.
Thanks to all who replied. I hope this summary is useful.
--
David Bernholdt bernhold@qtp.ufl.edu
Quantum Theory Project bernhold@ufpine.bitnet
University of Florida
Gainesville, FL 32611 904/392 6365