Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!zaphod.mps.ohio-state.edu!uakari.primate.wisc.edu!aplcen!haven!adm!smoke!gwyn
From: gwyn@smoke.BRL.MIL (Doug Gwyn)
Newsgroups: comp.unix.wizards
Subject: Re: Floating Point Expectations
Message-ID: <13008@smoke.BRL.MIL>
Date: 28 May 90 05:55:34 GMT
References: <995@s8.Morgan.COM> <13000@smoke.BRL.MIL> <1603:May2720:32:5190@stealth.acf.nyu.edu>
Organization: U.S. Army Ballistic Research Laboratory, APG, MD.
Lines: 16

In article <1603:May2720:32:5190@stealth.acf.nyu.edu> brnstnd@stealth.acf.nyu.edu (Dan Bernstein) writes:
-In article <13000@smoke.BRL.MIL> gwyn@smoke.BRL.MIL (Doug Gwyn) writes:
-> Rather than attempting to model exact behavior of floating-point systems,
-> I recommend devising algorithms that are robust in the face of
-> considerable cruft from the floating-point unit.
-This doesn't remove from FPU makers the responsibility to make chips
-satisfying reasonably simple rules. You really can get a lot better
-performance out of numerical algorithms when, e.g., you're working under
-Knuth's model of floating-point computation. ANSI floating-point is an
-excellent start, and it's worth the effort for strict compliance.

Unfortunately, in the real world the programmer cannot specify the
properties of the floating-point hardware.  There are numerous kinds
of FP architectures and implementations, each with its own set of
inaccuracies.  I stand by what I said -- instead of relying on the
FP model, make the algorithms robust.