[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

754r Thursday July Meeting minutes



[I think I am back on the list.
If you don't get this raise your hand ;->]

IEEE 754r Meeting minutes July 22, 2004
Sun, Santa Clara

Attendees:      Affiliation
Dan Zuras       Self
Joe Darcy       Sun
David Hough     Sun
Jeff Kidder     Intel (minute taker)
Jim Thomas      HP
Alex Lieu       Sun (phone)
Peter Markstein HP
Jon Okada       HP
Michael Parks   Sun (host)
Mike Cowlishaw  IBM
Dick Delp       Self
Ivan Godard OOBC
Note taker: Jeff Kidder

Next meetings:
David Hough is hosting 18-19 August at Sun in Menlo Park.

Extended Precision: (Foils went out on list)
[Discussion notes for extended precision]
Most HP quad transcendental libraries are 4-6 times slower than double.
Some interest in 754r containing at least all the information from 754.
Might be useful to define an "IEEE 754-esque"
Short and long formats tend to be application specific.
The purpose of a standard is to influence future behavior.
There is a tension between double-double and quad.
If we make it clear that the target is quad.
Fast almost always wins over good.
Extended isn't going to go away whatever we do.
Extended types could exist in a nonstandard way for library support.
There was disagreement. There were questions about the
use of double double. Prof Kahan has used double double in
Matlab by using double extended.
We could indicate that if your are going to implement
extended, the best one would be quad.
Could canonize existing practice.
Could limit extended to be less than the next larger format.
Some people use double double form legacy code from
power or Apple. Possibly because of cray's double double.
SW concern is it is important is that if you use whatever
extended is there can use environmental enquiries to
use it. This requires more sophistication of the use.
Also want to implement language standards, compiler
writers, and library writers.
We have to provide things like libraries to support
the various widths, so the compiler writers only
need to supply the calls and support but not the
runtime libraries.
We don't need to write them, but should encourage
the appearance. Could make sure gcc had them.
Papers could be published about how to do this in quad.
There is a gamma in fdlibm. Peter Macleroy (sp?).
Apple's SANE provided reasonable support which also
included support for 65-bit integer format.
Option C, to remove extended all together, had
some former supporters claim that they have been
convinced against going this way.

Proposal:
If the primary datatype is 32-bits then
shall implement single and should implement double
(as your single extended).
If the primary datatype is 64-bits then
shall implement double and should implement quad
(as your double extended).

What happens for a language implementer on IA32 what should long
double be mapped to.

Let's try to be the ones to push it and encourage it.
Quad gets used often enough for people to complain
about the performance. Rf: Doug Priest wrote a
paper on using arbitrary precision in geometry.

It is very different to have a trick

Alternative:
If the primary datatype is 64-bits then
shall implement double and should implement
a double extended format, of which quad is preferred.

The fast case and common case aren't independent variables.

People like to have standard types.

ISV's (such as finite element methods).

If quad is too slow people will just use double double.
I appreciate the idea of moving the whole industry toward
quad. It would make things simpler.

It boils down to how good is your data, how wide is it,
and how much are you doing with it.

Should we provide different advice to HW for Language/Compiler.

If you have too many different floating point types they number of
conversions grows as the square. But using long double the is
different in different contexts.

If you are going to start from scratch, quad is better than double
extended. For languages they should support.

What about the interoperation of modules that interpret long double in
different ways?

Please meditated on this for next month.

Agreement in principle: there should be at least one double extended
listed in the standard.

Future meetings:
Aug 18-19, David Hough, Sun in Menlo Park.
Sep 15-16, Peter Markstein, HP Palo Alto
Oct 13-14, Jim Thomas, HP Cupertino
Nov 17-18, Jeff Kidder, Intel SC
Dec 15-16, David Hough, Sun Menlo Park
Jan 19-20, Joe Darcy, Sun SC

Focused meetings aiming at Tue / Thurs 12:30-7:00.
with lunch before and supper after. Want to have
a document with rational for what we do.

754 | revision | FAQ | references | list archive