P1788
I read your article -- but then I don't need to be convinced of its merit.
Indeed, I think several members of the P1788 wg would agree, though some
think various different approaches not requiring fixed-point accumulators
can deliver equivalent results.
The issue for P1788 is timely delivery of a complete document, and that
is not easy. Had we kept an early title of "Dependable Arithmetic", the
case for including a complete description of Complete Arithmetic would
have been stronger. There are however unresolved technical issues in my
mind in DotProdP1788.pdf, with issues of language binding and exceptional
cases. These could certainly be resolved by further discussion, but not
in the little time we have to bring P1788 to the Sponsor Ballot stage.
That's why some of us suggested that work be started on a separate
standard, which the next revision of P1788 would gladly reference.
What we don't want to do is to include a requirement consisting of an
incomplete portion of Complete Arithmetic just for the purpose of a
correctly-rounded overflow-proof dot product. That's why the best we
could do was to require this dot product, and to recommend that it be
provided by means of Complete Arithmetic, with a reference.
> -- In December 2013 Intel published its new architecture:
> /Intel Architecture, Instruction Set Extensions Programming Reference/.
> See [36] in attachment 1. It provides register space of 16 K bits.
Unfortunately that reference leads nowhere; it just displays
789D5C2E-8816-4269-9E75-0842FA10B650-imageId=B40BECD0-6009-4455-9808-644186E21CA1.jpg
Searching on the Intel Developer Zone site for 319433-017.pdf or for
the title "Intel Architecture Instruction Set Extensions Programming
Reference" yields "Whoops! We're sorry".
It would have been nice to see how Intel would have been able to accomodate
some form of Complete Arithmetic.
I'm indeed familiar with the old S/370 implementation, but that benefited
from the narrow exponent range of the old IBM hexadecimal format. But as
you point out, IEEE Binary64 still leads to reasonable requirements on
today's machines. Binary128 might be a stretch however, never mind DFP128
or IBM's old Fortran XEXP, which would require 100 times wider registers.
That's why I would recommend support of Complete Arithmetic for at least
one supported number format, instead of for all of them.
Michel.
---Sent: 2014-04-30 19:01:58 UTC