Re: Sisal?

Robert Bernecky <bernecky@eecg.toronto.edu>
Wed, 21 Sep 1994 22:43:07 GMT

          From comp.compilers
Related articles
Sisal? wchang@genome.lbl.gov (1994-09-14)
Re: Sisal? preston@tera.com (1994-09-18)
Re: Sisal? buyukisik_o_f@ae.ge.com (U-E59264-Osman Buyukisik) (1994-09-19)
Re: Sisal? miller@diego.llnl.gov (1994-09-19)
Re: Sisal? bernecky@eecg.toronto.edu (Robert Bernecky) (1994-09-19)
Re: Sisal? bernecky@eecg.toronto.edu (Robert Bernecky) (1994-09-21)
Re: Sisal? bernecky@eecg.toronto.edu (Robert Bernecky) (1994-09-23)
| List of all articles for this month |
Newsgroups: comp.compilers,comp.lang.misc,comp.lang.functional
From: Robert Bernecky <bernecky@eecg.toronto.edu>
Keywords: functional, performance, Fortran
Organization: University of Toronto, Computer Engineering
References: 94-09-038 94-09-102
Date: Wed, 21 Sep 1994 22:43:07 GMT



>> Where might I find discussions/critiques/rebuttals regarding
>> Sisal? Is this dataflow/single-assignment language
>> "functional", and how does it manage to beat Fortran?


>It only beats FORTRAN on multi-processor parallel machines. It is first


I hope you'll pardon me for taking some of the wind out of your sails,
but about 5 minutes ago, I completed a benchmark of a simple convolution
code on the SUN-4, comparing SISAL, Fortran, and my APL compiler.




For a fairly hefty convolution (250x4000), I observed times as follows:


F77 -O3 0.730u 0.180s
SISAL -O -nobounds -cc="-O4" 0.610u 0.090s


I'm not ready to announce APL results yet.


Note that this is a vanilla, single processor machine, so your claims
about "only... multiprocessor.,." ain't strictly true.


Basis for comparison: I tried to get the most optimized version of both
fortran and SISAL. I am not that familiar with the SUN-4, and it may be
that there are better-performing compiler options. If so, please let me
know.


One note of interest: I had a fancy, hand-unrolled (10) code, originally
destined for a CRAY X-MP, and it ran SLOWER on the SUN-4 than did the
naive loop. May have been unrolled too far? I don't know.


Bob
--

Post a followup to this message

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