| Re: More on Exp bugs [message #7678 is a reply to message #7665] |
Fri, 13 December 1996 00:00  |
Peter Mason
Messages: 145
Registered: June 1996
|
Senior Member |
|
|
On 11 Dec 1996, David Siskind wrote:
> Anyway, regarding the Crimson user who found wierd things with
> Exp(-710.something), it must be an SGI/IDL thing. I use
> an Indigo and IDL 4.0.1 and I once spent a day (plus a post to this
> group) trying to figure out what appears to be the exact same thing.
> Now I'm real careful about floating under or overflows when doing Exp.
> This did not seem to be a problem with 3.6.1.
Here I go holding forth and stirring once again.
I'm convinced that this problem is due to the introduction of "NaN" and
"Infinity" support since IDL 4.0. I also think that the addition of
these features is what caused the floating-point slowdown observed on at
least some platforms, when comparing IDL 3.6.1 to IDL 4.0.x.
When IDL4.0 first came out, I must say that I didn't really notice the
slowdown on the ALPHA/OSF system I use. (Perhaps this O/S supports
denormals by default, and RSI's enhancements didn't add much extra overhead?)
But I didn't like the way I could segfault IDL with just exp() calls.
(By the way, I STILL can. On a DEC 3000/500 with exp(-90.0) followed by
exp(-9000.0). (Single precision.) Can anyone else do this?)
Anyway, I think that RSI was faced with a tough task in implementing
these FP features on all the IDL platforms; I imagine that each one has
its qirks and peculiarities when it comes down to FP denormals and
exceptions. And some take a bigger performance hit than others. (ALPHA/NT
certainly takes a major performance hit. Doing FP the lean way it likes to,
it makes ALPHA/OSF look tired. Doing it with denormals enabled and
various exceptions changed takes away its edge.)
If anyone's still reading this...
I realise that this is a MAJOR stir, but I was wondering what people's views
are on IDL's "NaN" and "Infinity" support?
Personally: I haven't yet implemented "Infinity" in my IDL programs,
and I haven't used "NaN" much at all. I like the idea of "NaN", but I
started many of my programs before it was around in IDL, and so I found other
ways to cope with "bad values" and the like. I can't be bothered with
FP underflows (just give me 0). Overall, I actually prefer FP support in
IDL the way it was in 3.6.1.
Peter Mason
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
