Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: Happiness



Replying to Ulrich Kulisch, Dan Zuras wrote:
> > So mandating the operations with directed roundings
> > cannot be a barrier to acceptance of 1788.
>
>  Perhaps.  Perhaps not.
>
>  It depends on what people like those at Intel think you are
>  mandating.

The mandate would be at the IA runtime level, not at the hardware level.

In a 754 system it is trivial to provide the functions, though perhaps
they would not perform well in isolation in the absence of hardware
support.  I'll note that I have heard that modern x86 processors have
the ability to cache the last two rounding modes used to permit switching
among them with little penalty, so Intel must have thought about this
already.  (The IBM/Sony Cell BE has separate modes for the two halves
of the vector pipeline, and DFP on IBM's Z already has individual modes
per operation; other platforms may also.  For type conversions explicit
directed roundings are even more common; IBM have them on both P and Z,
for BFP as well as for DFP.)

The issue is what a user can expect to find in an IA runtime environment.

If the user wants point-function directed rounding, she could dig through
blogs (where else do you find compiler information these days?) relevant
to the platform, write her own (perhaps by converting to an interval and
extracting a bound, using standard features, or using the underlying 754
dynamic modes, and give up performance potential), or use the standard
functions proposed by Motion 24.03.  Which sounds more attractive?

As I pointed out before, end users might not need this, but providers of
IA runtime libraries would...  and the functions proposed by this Motion
are at a low-enough level that library writers could reasonably expect to
find them even if the full 1788 standard is not present yet (in case the
library is being written to complete conformance to the full standard).

Michel.
---Sent: 2011-06-11 09:42:49 UTC