On Thu, 1 Sep 2005, Tim Jenness wrote:
> >> So I'm not sure whether this is an issue for Pat or just a flaky operating
> >> system compiler combination...
> >
> > Well if you believe it, Intel claims to be the more accurate floating
> > point representation, so that would tie in with this being a case of
> > the extra precision leading to an odd result. See the section "Current IEEE
> > 754 Implementations" in:
>
> Well, given that changing the last digit in E causes intel to fail as well
> something is clearly marginal. I have no idea whether the intel answer is
> the correct answer mind (it just converges and looks plausible).
>
> > to appreciate how such extra precision, and how it is used, could lead to
> > very different results. BTW, if I compile SLALIB with -ffloat-store (not
> > recommended as this is supposed give slower code) and run it under
> > Linux/g77 I do get the same results and status=-5.
> >
>
> So it fails if you are more careful?
By more careful, I mean to make the arithmetic work the same across
platforms.
As the article I pointed you at claims, the use of extended precision can
be both benefical and a potential problem. In this case it might be the
former, but you shouldn't be surprised if the results of a calculation
change a lot when you use it, especially when you have iterative behaviour
based on the comparison of double precision values.
Peter.
|