IEEE 754R minutes from June 19, 2002

Attendance and minutes review

The note taker was a bit too active to take comprehensive notes. These minutes are very abbreviated.

The IEEE 754R committee met June 19, 2002 at Network Appliance. In attendance were:

Draft Review

Back to Section 5.7, Comparisons. Kahan felt the important point made in the last line was hidden by the editing. It should emphasize that the equivalence between negation after testing and the negated predicate does not occur for predicates other than equality. The text will be restored.

There was great consternation about the unreferenced predicates that exist in the table. Hough will add an explanation for the people who feel they must all be referenced somewhere. Riedy asked why we shouldn't just require the entire table. Kahan and Darcy pointed to namespace polution.

Kahan prefers ! in the negated predicates (!>), but Darcy and Riedy prefer to work with positive varieties (?≤). This would matter if a short form is ever picked up by languages, so it won't matter.

Liu pointed out a somewhat typo, != v. ≠.

Kahan would like a rationale added for signalling predicates. They originally existed because systems (esp. the language portions) had no other way of testing for NaNs.

Alternate exception handling

No one liked Hough's proposal. (Well, someone may have, but the person taking notes had such an aversion that he didn't notice.) No one likes flush to zero. Hough intends on making hardware faster but less useful.

Repeat many of the conversations from the previous meeting here.

Interlude: Scheduling problems

The August meeting was rescheduled for 22 August.

Interlude: Transcendentals

Markstein: Paul Zimmerman provided an example argument where rounding arctan to +∞ on one platform produced a result two ulps lower than rounding to -∞ on another.

One discovered problem with implementations is that directed rounding and transcendentals don't mix. You can get a computed result that is on the wrong side of the actual result, and this breaks interval arithmetic. Markstein will try implementing a few quality transcendentals that round in the correct direction and report back.

More on alternate exception handling

On the product special operation: Fewer people detested it. It has actual uses (pown, determinants), but it misses some of the benefits of counting mode.

Kahan listed the three general uses of counting mode:

The discussion turned to the overhead of trapping. Schwarz pointed out that the P-series penalizes precise interrupts heavily. The six-stage pipeline only allows one insn to pass through it. Riedy asked if memory exceptions are precise. They are, but presumably they're handled by some piece that's designed to work well.

Kahan pointed out that we don't really need precise interrupts, except we do need interrupt boundaries at scope changes.

Benchmarks were declared hopeless, and many people are interested in knowing how to implement gradual underflow well.

Hough's proposal to make all NaNs possibly signalling was mentioned. We would gain something under-specified, but we would lose the ability for signalling NaNs to serve as uninitialized data.

The End.

754 | revision | FAQ | references