This is the mail archive of the
mailing list for the glibc project.
Re: [PATCH] New numbers in the benchtests.
On Sat, 16 Dec 2017, Siddhesh Poyarekar wrote:
> You're right in that specific inputs (or rather results) are more
> important than others due to the number of bits required to ensure a
> correct rounding regardless of the algorithm used. However, they assume
And such a correct rounding is not part of the accuracy goals for
functions such as exp (only for a few functions such as sqrt and fma that
are directly bound to IEEE 754 operations).
This does not rule out support for TS 18661-4 reserved function names such
as crexp, though that was not part of my TS 18661-4 proposal. If based on
exhaustive searches for worst cases for correct rounding, such
implementations would not actually need anywhere near the number of bits
used by the old exp implementation in glibc.
Joseph S. Myers