Re: IEEE Floating point

["Marc Snir" <snir@us.ibm.com>]

What I had in mind is (ISO/IEC TR 15580) Fortran floating point exceptions
which will be part of Fortran 2000. But I should have left these dogs
sleep, for the time being.

Marc Snir
IBM T.J. Watson Research Center
P.O. Box 218, Yorktown Heights, NY 10598
Tel: 914-945-3204 (8-862-3204)
Fax: 914-945-4425 (8-862-4425)
URL: http://www.research.ibm.com/people/s/snir


James Cownie <jcownie@etnus.com>@mpi-forum.org on 05/15/2001 05:04:06 AM

Please respond to mpi-21@mpi-forum.org

Sent by:  owner-mpi-21@mpi-forum.org


To:   mpi-21@mpi-forum.org
cc:
Subject:  Re: IEEE Floating point




Marc Snir wrote :-
> Another issue: I expect that most Fortran bindings of MPI are
> layered atop a C binding. This implies that the floating point
> operations done in the fortran MPI library may not be fully
> compliant to the Fortran floating point spec (when and whether one
> generates exceptions, how one handles NANs, the handling of
> denormalized numbers, etc.). I is worthwhile to add a clarification
> to that extent.

While it is entirely possible that the numerics inside the library
will not be the same as those which would occur were functionally
equivalent code implemented in Fortran, this is not a standards issue.

The Fortran standards say nothing about NaNs, exceptions or denorms,
all of which are IEEE floating point arithmetic issues. Fortran
standards certainly do not require that compliant Fortran systems use
IEEE arithmetic. (And neither do any C standards).

I don't really think that there's much to say here. Whenever you call
a library from Fortran you (should) already expect that the exception
behaviour and numerics will not be influenced by the way you compiled
your code, and that's all that is happening here.

-- Jim

James Cownie   <jcownie@etnus.com>
Etnus, LLC.     +44 117 9071438
http://www.etnus.com