From: Larry Estep <ljeste01 AT msuacad DOT morehead-st DOT edu>
Subject: Re: Math/rounding errors?
To: eliz AT is DOT elta DOT co DOT il (Eli Zaretskii) (Eli Zaretskii)
Date: Wed, 31 May 95 13:05:31 EDT
Cc: djgpp AT sun DOT soe DOT clarkson DOT edu
Mailer: Elm [revision: 70.85.2.1]

> 
> >     For instance, passing 5.0 to the function should result in a value
> > of 120.0.  Instead I get 118.401 from the djgpp version.  This is just
> > not acceptable.
> >     The code for the function is fairly simple.  It involves a few
> > multiplications and divisions, and one call to exp().  There are no
> > iterative loops to build up roundoff errors.  And even if there were,
> > I wouldn't expect it to be this severe.
> 
> Why not post the code?  It's usually very hard (at least for some of us)
> to second-guess a code without seeing it.

    Because I suspected it was a configuration problem.  I was expecting
someone to recommend a compiler option, an upgraded library, etc.
    As it turns out, the problem was that I was using the floating point 
emulator, which leaves something to be desired as far as precision goes.
I think I remember reading something about this when I installed djgpp
some time ago on my old 386.  Evidently I'd never done anything that was
math-intensive enough to show the limits of the emulator.  Fortunately,
I upgraded to a top-end 486 a couple months ago, so I no longer need the
emulator.  
    Thanks to all the folks who took the time to reply.  Especially
Morten Welinder, who correctly identified the problem.    

Larry

P.S.
    No, I wasn't trying this on a Pentium.  Although I was amused at
how many responses I got asking about this.  :)