IEEE 754R notes from December 15, 2005

Notes taken by Dan Zuras.

The December meeting of the 754 group was hosted by David Hough at Sun in Menlo Park on Wednesday 12/14 & Thursday 12/15 from 1:00 to 5:00.

Here is the attendence list from Thursday. In future, it will be used to keep track of voting members.

Name754 MemberAffiliation
Steven CarloughXIBM (phone)
Marius CorneaXIntel (phone)
Mike CowlishawXIBM (phone)
John CrawfordXIntel
Bob DavisSummit
Dick DelpXSelf
Mark ErleXIBM (phone)
J.P. FasanoIBM
Alex Fit-FloreaAMD
Ivan GoddardXOOTBC
Roger GolliverXIntel
Michel HackXIBM (phone)
John HarrisonXIntel (phone)
David HoughXSun
William KahanXU.C.Berkeley
Jeff KidderXIntel (phone)
Alex LiuXSun
Peter MarksteinXHP
Nobuyoshi MoriXSAP.COM (phone)
Stuart ObermanXNvidia
Jon OkadaXHP
Ian OllmannApple
Eric PostpischilApple
Eric SchwarzXIBM (phone)
Ilya SharapovSun
Jim ShearerIBM
Michael SiuNvidia
Peter TangXIntel
Jim ThomasXHP
Dan ZurasXGroup70

On Wednesday, Paul Zimmermann presented a result on a strict bound on roundoff error for complex multiply. It has some interesting & useful implications.

On Thursday, Paul talked about Correctly Rounded Transcendental Functions & Arbitrary Precision. The test for correct rounding in the fast case can be done with one FMA. MPFR was developed at LORIA. He urges us to adopt both transcendentals & arbitrary precision (as well as arbitrary precision transcendentals). We have already done so, at least in part & optionally.

We took our 3:00 break more or less on time.

After the break we began with the new IEEE operating rules. There were sufficient objections to details that it was felt that it needed editing before we consider it again. Jeff will head a PnP subcommittee together with Jim & Mike to edit the rules for presentation at the January meeting.

The Decimal Issue. Dave wanted to know if there had been any progress in the past few months. There has been a compromise proposed & Jeff & Mike are charged with writing it up for us. (All fields same width & significand different format) Dave pointed out that there is a third party on this issue, the users. Nobuyoshi supported the opinion that the users want only one format citing endianness as an example of a needless variation. He also supports database applications for which he already has a format & in which only comparison is needed. (Mike said there exists a decimal implementation in gcc.) Nobu said SAP supports Java, ABAP (an SAP language), & Cobol. While many people in the room agreed with his conclusion in principle, some felt his arguments lacked demonstrable support.

Numerical Debugging Horror Stories by Prof Kahan. We ran too long to get to this but we will put it on the January agenda.

754 | revision | FAQ | references