[Swarm-Modelling] Floating point arithmetic

[Gary Polhill]

Way back in October 2002, I mailed this list about some problems I was
having
with floating point arithmetic. With colleagues, I have written a
couple of 
publications on this in recent issues of JASSS, and one forthcoming in

Environmental Modelling and Software. One thing I have learned from all
this 
that is particularly relevant to Swarm (and rival platforms) is that
Swarm's 
Objective-C base allows access to the IEEE 754 stipulated routines (in
ieeefp.h) 
that allow programmers to check for floating point errors and control
the 
rounding direction. Java, by contrast, does not -- something for which
it is 
criticised by floating point experts such as William Kahan 
(http://www.cs.berkeley.edu/~wkahan/JAVAhurt.pdf).

For Java to be taken seriously as a language for scientific
programming, it 
needs to treat floating point arithmetic with the respect it deserves,
and 
implement the IEEE 754 stipulated functionality to achieve full
compliance with 
the standard. I've logged some requests for enhancement with Sun about
this, and 
they will only get implemented if enough people support them, by voting
for 
them. I have a page on them at 
http://www.macaulay.ac.uk/fearlus/floating-point/javabugs.html,
but if you want to access the RFEs directly (you will need to join the
sun 
developer network community to do so...), you can do this at:

http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=5095286
http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=5095600

Even if you don't think you'll ever use the routines yourself, I hope
you'll agree 
that you should have the right to inspect floating point exceptions in
Java, 
just in case you do come across a problem with your Java code that
might be 
attributable to fp issues.

Until Sun implement the changes, Obj-C Swarm users can at least enjoy a
small 
feeling of smugness that they are using a more numerically rigorous
platform... 
(There is one exception -- Cygwin doesn't implement the routines in
ieeefp.h 
(you get linking errors if you try). I have partially addressed this
issue by 
implementing some of them for Intel chips. See the CIieeefp package at

http://www.macaulay.ac.uk/fearlus/floating-point/download.html.)


Gary Polhill
Research Scientist
The Macaulay Institute
Craigiebuckler
Aberdeen AB15 8QH
UK
Tel: +(44) (0)1224 498200 Ext 2238
Fax: +(44) (0)1224 311556
e-mail: address@hidden
http://www.macaulay.ac.uk/
http://www.macaulay.ac.uk/fearlus/