The eighth meeting of the IEEE 754R revision group was held Wednesday 15 August 2001 at 3:00 pm at University of California, Berkeley, Bob Davis chair. Attending were David Bailey, David Bindel, Joe Darcy, Bob Davis, Dick Delp, Jim Demmel, David Hough, David James, Rick James, W Kahan, Ren-Cang Li, Alex Liu, Peter Markstein, Michael Parks, Jason Riedy, David Scott, Jim Thomas, Brandon Thompson, Neil Toda, Dan Zuras.
A mailing list has been established for this work. Send a message "subscribe stds-754" to firstname.lastname@example.org to join. Knowledgeable persons with a desire to contribute positively toward a substantially upward-compatible revision are encouraged to subscribe and participate. An official website contains some public material; current drafts require a password. Jason Riedy's site contains additional information.
The next meeting is scheduled for Thursday September 13 at 1:00 in conference room 101 at Intel SC-4 in Santa Clara. Further meeting dates are reserved for October 18, November 15, December 13, January 14, at Network Appliance, Santa Cruz conference room.
First order of business was a talk by David Bailey on multiple precision floating-point arithmetic needs: postscript. PDF. A new IEEE-based implementation is under development with Brandon Thompson and Sherry Li. Applications: discovering integer relations - Ferguson PSLQ algorithm; recognizing numeric constants; computing individual digits of constants like log 2 or pi.
Implications for 754R:
They haven't needed more than double precision exponent range so far, but a whole word exponent is allocated in the new package.
Short term, quad in hardware is the most valuable thing. Nobody knows how much more will be required, or when. Doubled-quad and Quadrupled-quad will suffice for quite some time. How should arbitrary precision be specified? Eventually some kind of standard would be desirable.
Kahan: after quad, one could provide the lower parts of quad + - * operations, much like old CDC computers did. The upper plus the lower part would be the exact doubled-quad result. A fused quad multiply-add would provide the lower part of *; a fused add-add with three operand and one rounding error would provide the lower part of +-. Lots of old knowledge is getting lost - and reinvented and patented via patent examiners too young to remember.
The draft minutes of the previous meeting were approved.
We discussed the current draft 0.25 prepared by David James on August 13. A number of issues were raised:
Tables are hard to get right; development in parallel with a text standard is difficult. Zuras - over the years, many implementation arguments were resolved by reference to specific sentences in 754. Delp - doesn't want to postpone tables to the end of the process for fear they might be wrong.
Kahan - mixing up 754 and 854 and 1596.5 at the outset make everything harder. Riedy - onus of change wordings is on proposers of changes - they should obtain approval first for the concept and again for the specific wording. Zuras - that provides an incentive for brief minimal changes.
Hough (afterthought) - "Rounding" needs to be defined at the outset as a mapping from the reals to a set of representable numbers, or from one set of representable numbers to another.
Process: for an audit trail, the following steps
The first vote was whether to continue with 0.25 as base document or start over with either 754 or 854. The vote was 11-1-4 to start over.
The second vote was whether to start over with 754 or 854. The vote was 9-1 to start with 754. Again this provides better auditability from the existing standard, even though in most ways 854 is an improvement.
The third vote was whether the evolving draft should be maintained in an open text-source format such as ASCII or HTML, viewable on any browser, or in one of the proprietary binary-source formats that IEEE accepts draft standards in (Adobe Framemaker or Microsoft Word). The vote was 10-2 in favor of an open text-source format. The second vote was between ASCII (.txt) and HTML formats, and ASCII prevailed 7-6. The tables will be added later, perhaps not in the same format.
Under these circumstances, David James resigned as document editor, and David Hough volunteered to take his place, and lacking any other candidate, was elected 14-0. After further discussion it was decided to maintain the evolving draft in HTML with some kind of change bars notation.
To relieve Hough, David Bindel agreed to serve as secretary and minute taker.
Given all the foregoing, it was agreed by consensus that all previous "agreed" technical changes bore re-examination at the next meeting. There was a list of these in the minutes from July.
Bob Davis commented that standard IEEE boilerplate must appear on EVERY page of EVERY draft. He will supply that.
Change list: it was agreed that a list of changes should be maintained with the document, an editor's responsibility. Bob Davis wants to keep an archival copy of each major step in the draft, for auditing purposes.
Todo list: it was agreed that we needed a todo list of proposed changes on the web site. Since these are agenda items, that appears to be the chair's responsibility.