Re: Please listen to Ulrich here...
Vincent Lefèvre, replying to Baker's question:
> However CA is intended to be implemented in hardware. In software,
> you would need to extract the sign, exponent and significand from
> each FP datum... This should be compared to pure FP algorithms
> (e.g. based on EFT's like TwoSum).
Any standard for CA would describe *behaviour*, not implementation.
One person's "hardware" is another's microcode, or even library function.
Moreover, a software implementation has many options, including deferred
unpacking and rebalancing, depending on the operation. It may well use
TwoSum, even FastTwoSum by keeping an ordered tree of components... It
all depends on whether the internal structure would have to be exposed.
Perhaps such exposure could be defined to be procedural, using opaque
objects with an operation to produce the standard layout as a bit vector.
I still believe CA deserves its own standard; unfortunately that takes a
lot of work, and it certainly couldn't happen in time for 1788 to refer
to it. Surely we don't want to drag P1788 out for another five years!
Michel.
---Sent: 2013-08-09 03:41:52 UTC