Re: Alternate floating-point results under directed rounding
John Pryce schrieb:
Arnold Neumaier wrote:
R. Baker Kearfott schrieb:
However, if we do supply strong arguments, particularly
on the fundamental or mathematical level, this may convince
floating point committee members to make a change in
the next 754 revision. (Note, though, that it is uncertain
when such a subsequent revision will take place.)
The final
Part 6: Useful directed rounding properties
of my proposal describes desirable deviations from IEEE 754 ...
This may be a valuable addition. But note
- To make cset implementations easier requires a different set
of changes to P754 behaviour, e.g.
+oo /dn +oo = 0, +oo /up +oo = +oo,
Yes, for infinity-as-number. But in my proposal, I mentioned that
this proposal was based on rejecting infinity-as-number.
where /dn, /up mean division with rounding down and up
respectively.
This change-set also should be included in such a section
of the text.
If cset arithmetic will become the standard, then yes.
Otherwise, one might add another section that describes
how to implement nonstandard arithmeitcs.
- Ensure any such requests are for a mathematical reason,
rather than to aid a particular algorithmic way of doing
things, which might be superseded by a better one.
I indicated the possible value for a particular scheme, and
vaiations work here as well. Since this is a recommendation only,
people catering for (to me unknown) algorithms which need
different choices are still free to do so.
Arnold Neumaier