Vincent Lefèvre's objections to "exact dot product" in P1788
(Was: Re: Motion 31: V04.2 Revision of proposed Level 1 text)
The discussions of exact dot product and Complete Arithmetic are not Level 1
issues. I think John and Christian were trying to introduce each Level 2
issue at Level 1 first -- but that can be done better by adding a paragraph
the mentions by name several concepts that are only relevant at Level 2
because they involve rounding, or avoidance thereof in a specific manner.
Note that fma falls into this category too.
Well, a level 1 definition can be given: all arithmetic is complete, fma
is just x*y+z, and dot product is a sum of products.
A separate issue is Complete Arithmetic. At one point P1788 was going to be
a "Standard for Interval and Complete Arithmetic for Dependable Computing",
but it ended being just "Interval Arithmetic".
Ulrich Kulisch and the numerical IFIP WG 2.5 want to find a place for the
exact dot product and Complete Arithmetic. They could not persuade P754R,
and P1788 (with its very early provisional scope) would have been another
possibility. Are we really committed at this point to restrict ourselves
to Interval Arithmetic? We did pass Motion 9 after all!
Michel.
---Sent: 2012-03-05 17:20:57 UTC