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

Re: Edge case conversions, exceptions to IEEE FPA



Van Snyder schrieb:
On Tue, 2008-11-11 at 13:32 -0800, R. Baker Kearfott wrote:
Van et al,

What's sacred about 1ULP? Might the programmer know the accuracy, which
isn't necessarily 1ULP?  Such a case would be an argument in favor of
perhaps supplying a conversion of two floats, one representing the
midpoint, and the other representing the accuracy (in the form of
a radius, say).  Thus, we WOULD have two conversion routines, one
for midpoint-radius, and one for lower bound and upper bound.

For an FP number not known to be exact, the closest to exact it can be
is 1/2 ulp each way.

Connecting these conversions to programming languages is a separate
issue.  Do we want the two routines defined at the 1788 operation
level, or at the programming language level?

I think conversions beyond what P754 specifies belong at the programming
language level.

The P1788 standard should say "An interval is represented by two
floating-point numbers; the first is the lower bound of the interval and
the second is the upper bound."

This is part of Section 1.2 of my proposal (v2.0 or v2.11).

But surely the standard should say more than that.


Arnold Neumaier