*BSD News Article 39908


Return to BSD News archive

Xref: sserve comp.sys.powerpc:30653 comp.sys.intel:26968 comp.os.misc:3580 comp.unix.bsd:15728 comp.unix.pc-clone.32bit:7884 comp.unix.sys5.r4:8928 comp.unix.misc:15260 comp.os.linux.development:21741 comp.os.linux.misc:32356 comp.os.linux.misc:32357 comp.os.386bsd.development:2903 comp.os.386bsd.misc:4545
Path: sserve!newshost.anu.edu.au!harbinger.cc.monash.edu.au!news.mira.net.au!otis.apana.org.au!serval.net.wsu.edu!netnews.nwnet.net!oracle.pnl.gov!osi-east2.es.net!lll-winken.llnl.gov!uwm.edu!spool.mu.edu!bloom-beacon.mit.edu!panix!berke
From: berke@panix.com (Wayne Berke)
Newsgroups: comp.sys.powerpc,comp.sys.intel,comp.os.misc,comp.unix.bsd,comp.unix.pc-clone.32bit,comp.unix.sys5.r4,comp.unix.misc,comp.os.linux.development,comp.os.linux.misc,comp.os.linux.misc,comp.os.386bsd.development,comp.os.386bsd.misc
Subject: Re: Interested in PowerPC for Linux / FreeBSD / NetBSD?
Date: 22 Dec 1994 16:12:25 GMT
Organization: yes
Lines: 26
Message-ID: <3dc8h9$iug@news.panix.com>
References: <3cphs0$l6e@ddi2.digital.net> <stephenkD0vyyn.1GG@netcom.com> <D12MGt.1L9@tyrell.net> <D16ony.K5z@cogsci.ed.ac.uk>
NNTP-Posting-Host: 166.84.251.57
X-Newsreader: NN version 6.5.0 #1

In <D16ony.K5z@cogsci.ed.ac.uk> richard@cogsci.ed.ac.uk (Richard Tobin) writes:

>In article <D12MGt.1L9@tyrell.net> john@tyrell.net (John A. Matzen) writes:
>>If a bank is using the FPU to do numerical calculations, they could be loosing
>>hundreds of pennies every day due to rounding errors.  It doesn't matter
>>how many digits of precision the FPU is using.

>Sorry to be rude, but this is just drivel.  IEEE floating point is
>guaranteed (when correctly implemented!) to be exact for operations
>such as addition on integers up to 53 (?) bits.  Floating point isn't
>just any old approximation; it's a well-defined approximation that
>can be safely used by people who understand it (except on Pentiums).

There's no need to be rude.  Who said that calculations were restricted to
additions?  If they were, obviously, they would perform identically on
Pentiums.

The point is that the type of calculations that result in reduced precision
on Pentiums are exactly those type of calculations ((x * y)/ x) that should
be analyzed for roundoff error anyway.  Of course having the Pentium's
reduced precision in certain cases makes this analysis more difficult, but
you seem to be making the assumption that a correctly implemented IEEE fp
obviates the need for any such analysis.
--
Wayne Berke
berke@panix.com