| Re: exp function bug [message #7574] |
Fri, 06 December 1996 00:00  |
Peter Mason
Messages: 145
Registered: June 1996
|
Senior Member |
|
|
On Thu, 5 Dec 1996, Robert Cannon wrote:
> Can anyone tell me what is going on here, or if it happens on other
> platforms?
> The first time I call exp (-710.72026d0) after starting idl I get:
> hera:~> idl
> IDL. Version 4.0.1 (IRIX mipseb).
> IDL> print, exp (-710.72026d0)
> 2.8462073e+134
> IDL> print, exp (-710.72026d0)
> 2.1782760e-309
> This is on an sgi crimson. I get the impression the same may happen on
> an alpha, but for a different value - it has to be -710.72026d0 on the
> sgi: -710.72025d0 works fine.....
I think that exp(-710.72026d0) is on the borderline of giving a
floating-point underflow on your platform. Here are some approximate
values of "a" in EXP(a) beyond which a floating-point underflow occurs on
IDL platforms I have available here. (I haven't observed inconsistencies
like you did.)
PLATFORM A
Intel Pentium, NT3.51 -708.D
DEC ALPHASTATION 200 4/233, NT3.51 -708.D
DEC ALPHA/AXP 3000/500, OSF -744.D
(|A| is much smaller for single precision, of course.)
I don't really understand what could cause those inconsistent answers.
Here's, well, a stream of consciousness...
EXP(large neg value) used to silently return 0.0 in older versions of IDL (the
way I prefer it to do), and when RSI introduced support for some floating
point "denormals" (Infinity and NaN), things were quite strange initially:
On my DEC OSF platform, EXP(-x) would generate an underflow error for
moderately large x, but would cause IDL to crash with a segfault given a
suitably large (much larger) x.
I think that when you're working right at the edge of FP precision like this,
you can easily run into platform-specific differences. (For example, ALPHA/
NT doesn't handle denormals in hardware.)
But I also think that EXP(-710.72026d0) == 2.8462073e+134 is unacceptable at
any time, and that you should send in a bug report. (Perhaps there is some
bug in IDL's FP exception-handling code on the SGI.)
Peter Mason
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
