Does Java's floating point implementation still have the problems identified by Kahan?

views:

473

answers:

+5 Q:

Does Java's floating point implementation still have the problems identified by Kahan?

I've read a few papers from Kahan tonight, and his famous rant against Java. Before I dive into the JVM spec, did anything change since the initial rant on this front? For example:

setting rounding mode
accessing the flags
getting more precision for free
... ?

Thanks,

Nico.

My guess is they are still valid as I have not seen much change in this area since '98.

However, I am not sure they would really be used much even if they were added today. Many languages support variable precision arithmetic (as does Java) which appears to be a better solution for this type of problem. e.g. BigInteger, BigDecimal.

Peter Lawrey 2009-03-30 05:58:15

"Arbitrary precision", not "variable precision". Together with changes to the floating point semantics in Java 2.0 (1.2) the arbitrary precision types give a good answer to Kahan's complaints. Not that I'm a numerical analyst myself.

Guss 2009-03-30 06:42:32

I should confess that I've never seen a matrix inversion or a Kalman predictor in arbitrary precision. I'm not sure this is the right tool for numerical simulation.

nraynaud 2009-03-30 23:52:18

+2 A:

Prof. Kahan's student, Joe Darcy, became Sun's "floating point czar". His blog entry: http://blogs.sun.com/darcy/entry/everything_old_is_new_again is an entry point for learning more about work he did to correct the problems. There have been great improvements.

Tod Landis 2009-03-30 07:08:25

Thank you, but this doesn't adress the rant. But it hints at the fact that there is no change in the fp system.

nraynaud 2009-03-30 13:00:44

ansaurus

tags:

views:

answers:

Does Java's floating point implementation still have the problems identified by Kahan?

related questions