wendyt@pyramid.UUCP (Wendy Thrash) (04/22/86)
One thing that has struck me about this discussion is that none of the
participants (including yours truly) seem to have been on the IEEE 754
committee, which was not that small a group. Could someone over at Sun
kick David Hough and point him this way?
[So tell us, David, (or Profs. Kahan, Cody, et.al.) is there any chance that
a future IEEE committee might establish some sort of standards for library
routines? Sounds like you could have some pretty interesting meetings;
any way to get invited if it happens?]
I see now that I should have studied numerical analysis rather than
algebraic topology, but I didn't. Consequently, I am too stupid to know
how to implement a full-precision routine without having extended-precision
arithmetic available. Are you guys saying this is possible? Dik Winter
says that "Given enough time, it is quite simple to implement," but perhaps
the reference to enough time implies the use of extended precision. I'd love
some references. I tend to believe Koenig's article, which seems to
imply that even in the presence of arbitrary-precision arithmetic it
may be impossible to write a full-precision routine.
Matt Wiener suggests writing routines such as sinpi, in order to solve
certain accuracy problems. I see this as a step away from standardization
and portability, therefore not in keeping with the spirit of IEEE.
(Yes, I recognize that too much standardization can stifle creativity and
prevent good ideas from being born -- sincos, for example.) Moreover,
this is moot in my case, for I have been given the task of writing a standard
math library -- you know, sin, cos, exp, sqrt, and so on. No frills, no
sinpi, just the things that standard F77 wants.
There are issues in an IEEE library that have nothing to do with rounding.
For instance, what is atan2(0.,0.)? The only thing that makes sense to me
is to return a NaN (perhaps issuing an error message and/or setting an
error flag as well). What should C's printf and f77's print routines do
when confronted with Inf or NaN? These are the real puzzles for me in
writing a library. I would like to do them in some standard way, but there
seems to be no standard to follow. I can either buy a Sun-3 and copy
David Hough's solutions, or make up my own.
Summary: HELP! Get a committee together to answer some of these questions!
If there is one already, give me some hints about what's going on.
--
--
Wendy Thrash {allegra,cmcl2,decwrl,hplabs,topaz,ut-sally}!pyramid!wendyt
Pyramid Technology Corp, Mountain View, CA +1 415 965 7200 ext. 3001