Parallelizer and vectorizers available

This is a followup to my message requesting references to vectorizers,
parallelizers, and dependence analyzers. I am indebted to the many
resondents to my query who provided me with a wealth of useful
information.


I know of three commercial products: FORGE, KAP, and VAST. I don't know
anything more about FORGE. KAP and VAST are source-to-source FORTRAN
translators which do various sorts of program transformations such as loop
unrolling, loop fusion, and other recognition of particular paradigms and
replacement by more efficient equivalents. Both KAP and VAST are
available on IBM RS/6000s running AIX 3.2 as the -Pk and -Pv preprocessors
for XL FORTRAN version 2.2. KAP can be acquired from
Kuck & Associates
1906 Fox Drive
Champaign, IL 61820
VAST can be acquired from Pacific-Sierra Research.


There are a number of research products available.


The FORTRAN-D compiler can supposedly be obtained for a minimal fee from
the Rice University Center for Research on Parallel Computation.


The Omega test, an array data dependence analyzer, is implemented on top
of tiny (see below) as well as in a stand-alone test driver. It is
available via anonymous ftp from ftp.cs.umd.edu (128.8.128.8) in the
directory pub/omega, along with a slightly modified tiny system.


Parafrase-2 is supposedly available from the University of Illinois CSRD.


The ParaScope programming environment is available from Rice; contact
Theresa Chatman <tlc@cs.rice.edu> for details. It includes tools for
editing, analyzing, parallelizing and debugging Fortran.


Pat is an interactive source-to-source Parallelizing Assistant Tool. It
includes two passes, the first taken from a Fortran 77 compiler and used
to generate intermediate files, and the second providing interactive code
transformations to generate a parallel program. You can get pat by
anonymous ftp from 130.207.6.52:/pub/pat.


The PIPS parallelizing compiler is available from Ecole des Mines in
France; is is only a prototype, but people outside the research group are
working with it. Contact Pierre Jouvelot
<jouvelot@alexandria.lcs.mit.edu> for more information, or see the paper
in the proceedings of the International Conference on Supercomputing'91.


The PSERVE dependence server at Rice is based on the PFC batch
vectorization/parallelization system. (PFC is a Rice project related to
ParaScope; the distinction between the two isn't completely clear to me.)
Distributions of ParaScope include support for obtaining PSERVE dependence
graphs and using them for parallelization or debugging. If you want
information about accessing PSERVE outside of ParaScope, contact Paul
Havlak <paco@cs.rice.edu>.


There is something called "Ptool" that I don't know anything more about,
except that one correspondent says it's buggy.


The Sigma system has a component that does parallelization/vectorization
transformations. It is not automatic, but it works by user directives.
The transformations that it now supports include, vectorization, loop
distribution, loop blocking & multiloop tiling, loop interchange, loop
unrolling, loop merging and scalar expansion and localization. These
constitute the standard suite of transformations used by parallelizers.
Contact Dennis Gannon <gannon@cs.indiana.edu> for more details.


The "tiny" system (by Michael Wolfe <mwolfe@cse.ogi.edu> at the Oregon
Graduate Institute) is a research implementation of an interactive program
restructuring tool. Tiny computes data dependence relations and
interactively performs many restructuring transformations, such as loop
interchanging, distribution, skewing, ... . The source is available by
anonymous ftp from cse.ogi.edu.


-Michael Ernst
mernst@theory.lcs.mit.edu
--

Return to the comp.compilers page.
Search the comp.compilers archives again.