Path: utzoo!attcan!uunet!husc6!mailrus!cornell!batcomputer!braner
From: braner@batcomputer.tn.cornell.edu (Moshe Braner)
Newsgroups: comp.sys.transputer
Subject: Re: Transputer Benchmarks
Summary: Yes, the compiler can make all the difference.
Message-ID: <6879@batcomputer.tn.cornell.edu>
Date: 25 Nov 88 02:46:51 GMT
References: <1389@thumper.bellcore.com> <22@microsoft.UUCP> <228@cs-spool.calgary.UUCP> <1000@etive.ed.ac.uk>
Reply-To: braner@tcgould.tn.cornell.edu (Moshe Braner)
Organization: Cornell Theory Center, Cornell University, Ithaca NY
Lines: 21

[response to doubts expressed about posted benchmark]

I've seen compiler differences of 100-fold before, using the same
program on the same hardware.  In this case (a ray-tracing program)
I would guess that the key compiler feature is the trigonometry library,
since the T800 FPU only does +-*/ plus sqrt (sort of).  I have seen
rather naive trig libs using simple algorithms written in C, and, in
contrast, good algorithms hand-optimized in assembler.  The Logical
Systems math library is of the latter sort, and is the fastest there is
on the T800 as far as I know.  So the quoted 25-fold advantage (over a
beta version of an unnamed compiler) seems believable to me.

BTW for graphical rendition purposes single precision is frequently
adequate and a compiler that allows breaking the standard C rule of
doing everything in double wins if it can do single precision faster.
A library that does it even faster in half-precision can also be useful.

- Moshe Braner

"compiled BASIC can be faster than interpreted C"
"Spaghetti code is possible even in Pascal"