Re: Please listen to Ulrich here...
Ian McIntosh wrote:
> I understand, but System/370 had a 7 bit exponent for all precisions.
Actually, 9-bit binary equivalent (7-bit hex exponent), i.e. intermediate
exponent range between those of IEEE binary32 and binary64. S/370 also
had an extended exponent format with a 14-bit hex exponent, whose range
actually exceeded those of IEEE binary128 and decimal128 -- but there
was no ACRITH (IBM's name for Complete Arithmetic) for that; indeed, it
was a software-only format used by Fortran.
With respect to Ulrich's DotProdP1788.pdf -- that still exposes the CA
implementation of EDP (Exact Dot Product). This kind of detail belongs
in a standard for Complete Arithmetic, and is not relevant here. If we
were to require support for an EDP, we would specify an abstract format
capable of representing an EDP, which could be an array or list of FP
numbers, or something else. A detailed format would be relevant only
if we also required an interchange format for this type-- but that again
would only be relevant to a standard for, effectively, complete arithmetic.
There is another issue with the operations described in DotProdP1788 -- but
that one could be treated as an implementation detail -- namely the fact that
such complete formats typically cannot practically be passed or returned by
value -- except in environments that indeed have multiple complete registers.
In a C-like environment the actual programming interface would most likely be
CompleteOp(CAtype *result, CAtype *arg1, CAtype *arg2)
instead of
CAtype CompleteFormatOf_op(CAtype arg1, CAtype arg2).
Michel.
---Sent: 2013-08-26 18:48:27 UTC