Path: utzoo!attcan!uunet!lll-winken!lll-lcc!well!rchrd
From: rchrd@well.UUCP (Richard Friedman)
Newsgroups: comp.lang.fortran
Subject: Re: Fastest Fortran Engine?
Message-ID: <6386@well.UUCP>
Date: 27 Jun 88 18:41:19 GMT
References: <5377@cup.portal.com> <2852@mmintl.UUCP> <1005@cresswell.quintus.UUCP> <701@garth.UUCP> <2157@sugar.UUCP> <777@garth.UUCP> <10757@ames.arc.nasa.gov> <7568@boring.cwi.nl> <10889@ames.arc.nasa.gov>
Reply-To: rchrd@well.UUCP (Richard Friedman)
Organization: RCHRD 2855 Telegraph #415 Berkeley CA 94705
Lines: 21

One reason why the NEC SX-2 does so well on matrix operations is that their
compiler is one of the best!.  See "Vectorizing Compilers: A Test Suite and
Results" from Argonne (Tech Memo #109, March 31, 1988) from Dongarra,
Callahan and Levine.  It gives the results of running a large (100+) suite
of routines thru a number of vectorizing compilers.  Altho execution times
are not given, (that WOULD be interesting!), they were only rating these
compilers on their ability to vectorize/optimize source code.  

On a composite scale, the NEC compiler vectorize 59 out of 100 loops while
CFT 1.15 did 51% and CFT77 did 52%.  I should mention that the Alliant
compiler did 71%  and so did the Amdahl FORTRAN 77/VP.  The FTN77 compiler on
the ETA-10 did 69%.

The point is that the NEC SX-2 runs faster than the X-MP in vector mode
(clock is 6ns) so the more code you get into vector mode the better you
are.  I suppose that more loops in the LINPACK test were vectorized by
the NEC compiler than the Cray.  But thats just a guess.  
The SX-2 is a significant machine.  
-- 
...Richard Friedman [rchrd]                         
uucp:  {ucbvax,lll-lcc,ptsfa,hplabs}!well!rchrd
- or -   rchrd@well.uucp