RE: [Stds-754] Motion: Delete 5.3.2 Scaled-product operation

["James, David V" <david.v.james@xxxxxxxxx>]

David,

After the last meeting, Prof Kahan expressed concern that
the baby had been discarded along with the bathwater.
After hearing his explanation, I agree, and will assist.

The plan-of-action was to iterate a revision, where the
results were clearly specified (via tables, of course),
so that the details could be clearly understood. After
initial review, that would be posted as a motion.

The intent was to foster cooperation through preliminary
reviews, while relieving the WG of the responsibility
of tracking these intermediate steps.

The rough idea is to catch the baby and leave the bathwater.

Rather than specifying a vector, specify a single step.
Then, the programmer/hardware only has to shove out the
arguments for each step, w/o monitoring the intermediate
values. This would cover your concern about add/sub overflow.

Basic primitives would be:
  ScaledProduct:
    {value3, scale3} = {value1, scale1} * value2
  ScaledSum:
    {value3, scale3} = {value1, scale1} * (value2a + value2b)
  ScaledDiff:
    {value3, scale3} = {value1, scale1} * (value2a - value2b)

I'll be working with key parties to develop a detailed
proposal (including tables, of course) within the next weeks.
If you could spare the time to review initial proposals,
perhaps we can deliver a mutually satisfactory solution.

Respectfully,
DVJ


-----Original Message-----
From: stds-754-bounces@xxxxxxxxxxxx
[mailto:stds-754-bounces@xxxxxxxxxxxx] On Behalf Of David Hough 754R
work
Sent: Sunday, June 25, 2006 1:58 PM
To: stds-754@xxxxxxxx; stds-754@xxxxxxxxxxx
Subject: [Stds-754] Motion: Delete 5.3.2 Scaled-product operation


Rationale:
The current specification is too limited to be worth standardizing.
Actual applications are just as likely to need over/underflow
protection for products of sums and differences.     And the array
versions would not likely be implemented upon a primitive with the
prescribed
semantics (a more probable implementation primitive would return zero sf
if
no over/underflow intervenes).    Since the whole point is performance,
the user feature (acting on arrays) would be implemented upon a
performant
primitive.
_______________________________________________
Stds-754 mailing list
Stds-754@xxxxxxxxxxxx
http://mailman.oakapple.net/mailman/listinfo/stds-754