[
Date Prev][
Date Next][
Thread Prev][
Thread Next][
Date Index][
Thread Index]
RE: Meeting the Scope and Purpose of P754
Michel,
I agree with your general scoping, but believe that one
has to address some of the basics also, such as:
before/after tiny rounding
FMA exeception scanning, before/after
etc.
If you wise to work together on this, I would be happy
to work on a joint (not in the 60's sense) proposal.
Busy next week, though...
Cheers,
DVJ
-----Original Message-----
From: stds-754@xxxxxxxx [mailto:stds-754@xxxxxxxx]On Behalf Of Michel
Hack (1-914-784-7648)
Sent: Friday, July 13, 2007 12:07 AM
To: Bob Davis
Cc: stds-754
Subject: Meeting the Scope and Purpose of P754
Rather than chasing the impossible dream, perhaps we should put the
stated "1.4 Purpose" in perspective, as I suggested in this forum
on May 30 in my reply to Dave James with regard to
RE: rogue comment about reproducibility ("Identical results")
Here are some extracts:
As I said, it *is* reasonable to require control over reproducibility,
but at the same time the context in which reproducibility applies has
to be (a) feasible and (b) describable.
[brief discussion on effects of different integer sizes]
The 754 standard has gone a long way towards bounded reproducibility,
and in my opion the "Purpose" states an overall goal to which 754 is
expected to contribute substantially. Perhaps part of the solution is
to say this explicitly, instead of giving the wrong impression that
compliance with 754 will be sufficient to achieve that goal.
[mention of rules used to define compatibility across generations]
We can do something similar. Exactly reproducible results can be
obtained by adhering to the following guidelines:
(a) No timing dependencies
(b) No peeking at internal formats
(c) No dependency on aspects documented to have a platform dependency
(d) Explicit request of order-preserving compilation or interpretation
(e) Staying within minimal required bounds (e.g. fewer than M+3 digits
in conversions between decimal strings and binary)
(f) No use of reduction operations
(g) Avoidance of operations not documented to be correctly rounding
(h) Awareness of the variability of aspects not covered by
the standard
Some of the above overlap, and perhaps there are a few others. This
may seem awfully restrictive, but it does leave room for a very large
class of useful programs, AND it is achievable without lying.
As Dave James said later (June 1st):
DJ> So, one either has to change the purpose or meet the purpose...
I was suggesting a flexible interpretation of the purpose instead,
and Mark Erle concurred (same day):
ME> I interpret the purpose statement as pertaining to the operations
ME> defined in the standard, not all conceivable operations.
There is of course the issue that the current standard does not fully
describe the operations defined in the standard, e.g. with respect to
NaN-related properties. Several people have requested that those be
listed explicitly, and I concur. This has however to some extent
been discouraged precisely because it *appears* to conflict with the
stated Purpose. So we have to make clear that it does not conflict,
i.e. that the Purpose addresses *reasonable* expectations of repro-
ducibility, and that the details of applicability will be made clear.
Michel.
Sent: 2007-07-13 23:47:03 UTC