Path: utzoo!censor!geac!torsqnt!news-server.csri.toronto.edu!cs.utexas.edu!wuarchive!emory!hubcap!rcarter
From: rcarter@nas.nasa.gov (Russell L. Carter)
Newsgroups: comp.parallel
Subject: Re: CM Fortran (actually, CRI array syntax performance)
Message-ID: <12444@hubcap.clemson.edu>
Date: 2 Jan 91 18:06:35 GMT
References: <12301@hubcap.clemson.edu> <12324@hubcap.clemson.edu> <12381@hubcap.clemson.edu> <12398@hubcap.clemson.edu> <12410@hubcap.clemson.edu>
Sender: fpst@hubcap.clemson.edu
Reply-To: rcarter@wilbur.nas.nasa.gov (Russell L. Carter)
Organization: NAS Program, NASA Ames Research Center, Moffett Field, CA
Lines: 67
Approved: parallel@hubcap.clemson.edu

In article <12410@hubcap.clemson.edu> john%ghostwheel.unm.edu@ariel.unm.edu (John Prentice) writes:
>In article <12398@hubcap.clemson.edu> serafini@amelia.nas.nasa.gov (David B. Serafini) writes:
>>
>>CFT77 has had a subset of the Fortran 90 array syntax for quite a while now,
>>at least a couple of years.  It hasn't caught on much, (to my knowledge)
>>probably because of the portability problems it creates.  The Cray Fortran
>>preprocessor, fpp, now can put out do-loop based code when given array code,
>>so portability is less of a concern now. 
>>
>This is rather interesting.  Does fpp generate a new do-loop for every
>array construct or is it smart enough to combine operations the way
>you would if you were writing in Fortran 77?  What is the effect on
>the optimizer of all this?
>
>John K. Prentice
>john@unmfys.unm.edu

Well, let's look at some data. I converted the NAS Kernels, a popular
CFD benchmark (at least here at NAS) to array syntax, and ran
it on our YMP.  Here is what I get:

Using array syntax:

                THE NAS KERNEL BENCHMARK PROGRAM


 PROGRAM        ERROR          FP OPS       SECONDS      MFLOPS

 MXM          1.8085E-13     4.1943E+08      1.5879      264.15
 CFFT2D       3.2001E-12     4.9807E+08     11.1267       44.76
 CHOLSKY      1.8256E-10     2.2103E+08      4.9911       44.29
 BTRIX        6.0622E-12     3.2197E+08      4.5159       71.30
 GMTRY        1.0082E+00     2.2650E+08      3.5258       64.24
 EMIT         1.5609E-13     2.2604E+08      1.3055      173.15
 VPENTA       2.3541E-13     2.5943E+08      7.1281       36.40

 TOTAL        1.0082E+00     2.1725E+09     34.1810       63.56


And using plain vanilla fortran 77:


                THE NAS KERNEL BENCHMARK PROGRAM


 PROGRAM        ERROR          FP OPS       SECONDS      MFLOPS

 MXM          1.8085E-13     4.1943E+08      1.5705      267.06
 CFFT2D       3.2001E-12     4.9807E+08      7.0951       70.20
 CHOLSKY      1.8256E-10     2.2103E+08      2.6393       83.75
 BTRIX        6.0622E-12     3.2197E+08      2.3717      135.76
 GMTRY        6.5609E-13     2.2650E+08      2.0910      108.32
 EMIT         1.5609E-13     2.2604E+08      1.2987      174.05
 VPENTA       2.3541E-13     2.5943E+08      4.7900       54.16

 TOTAL        1.9305E-10     2.1725E+09     21.8563       99.40



The MFLOPS definitely decrease for this code when the DO loops
are expressed in the array section syntax.  We have smart users;
They HATE to go slower.  So few apparently use the array
syntax here at NAS, unless they want a code that can be run
with minimal changes on the CM.

russell
rcarter@wilbur.nas.nasa.gov