Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!watmath!clyde!cbosgd!gatech!seismo!harvard!stew
From: stew@harvard.UUCP (Stew Rubenstein)
Newsgroups: net.micro.mac
Subject: Re: Floating Point bench-marks
Message-ID: <735@harvard.UUCP>
Date: Tue, 25-Feb-86 10:56:07 EST
Article-I.D.: harvard.735
Posted: Tue Feb 25 10:56:07 1986
Date-Received: Fri, 28-Feb-86 06:31:52 EST
References: <887@umn-cs.UUCP>
Reply-To: stew@harvard.UUCP (Stew Rubenstein)
Distribution: net.micro.mac
Organization: Aiken Comp Lab, Harvard
Lines: 25

In article <887@umn-cs.UUCP> waddingt@umn-cs.UUCP (Paul Fink ) writes:
>
>I, with the help some freinds, perform some simple bench marks 
>on three development systems.

I played with floating point on MegaMax a while ago...  The problem I have
with them is that they decided that compatibility with the world of 8-byte
doubles is important.  This results in a conversion from 8 to 10 byte format
and back for every operation, three traps where one should do.  Perhaps I am
naive, but I would implement the default floating point in C (double) using
the default floating point format (10 byte IEEE).  Compilers should at least
allow the use of the 10 byte format, perhaps via an "extended" type (I think
Consulair does this).

Any program which relies on "double" being 10 bytes or exactly two "float"s
is BROKEN.

Anyway, to the point -- coding my own (very simple) fp routines using

typedef struct { char b[10]; } extended;

sped up my crude benchmark by a factor of two.  It is not clear that this
is worth losing infix equations and normal assignments and comparisons.

Stew