On Wed, 3 Feb 2016 11:07:22 +0100, Ulrich Kulisch wrote:
I would vote YES if in Table 4.1 under /Basic operations/ after
fma(x,y,z) the /dot product of two vectors/ would be added.
Interesting turn of events. In the context of 1788.1 being a subset
of the full standard, I assume this refers to the *correctly rounded*
dot product that is required in the full standard, and not the EDP.
I'm sorry that this possibility never came up during the discussion
period, because it raises several issues.
First, specifically with respect to Table 4.1: this table lists the
Required forward elementary functions, and corresponds to Table 9.1 in
the full standard. This is not the place for mentioning the reduction
operations; the required correctly-rounded reduction operations of the
full standard are not mentioned in Table 9.1.
So if Prof. Kulisch wants to require support for those reduction operations
in 1788.1 it would have to be in a different place.
Second -- when looking this up in P1788_1_MAIN.pdf, I noticed a glaring
omission in Clause 8, "Not Required": The reduction operations! Indeed,
they are not mentioned at all in the document. This we should fix.
Then comes the interesting question, which would require a new Motion:
Should the correctly-rounded reduction operations of Clause 12.12.2 of
the full standard be included in the reduced standard? Or should they
perhaps be "strongly recommended"?
This boils down to whether there are sufficiently many situations where
the lack of a correctly-rounded dot product would lead to useless results,
e.g. solutions to an ill-conditioned problem being Entire, whereas if the
dot product were correctly rounded, a reasonable result would come out.
Of course, with EDP, a reasonable result could be expected in almost all
cases except for true singularities.