math: Improve hypot performance with FMA
commitccfa865a82c648fde56864ea094f70ee1a8a944b
authorWilco Dijkstra <Wilco.Dijkstra@arm.com>
Tue, 30 Nov 2021 19:29:25 +0000 (30 16:29 -0300)
committerAdhemerval Zanella <adhemerval.zanella@linaro.org>
Mon, 13 Dec 2021 12:02:34 +0000 (13 09:02 -0300)
tree5d05ad55e786295bb8c49afa244ac9646a8ee98a
parent6c848d70383e1dbe932ef41723ac0abfdeec7ca8
math: Improve hypot performance with FMA

Improve hypot performance significantly by using fma when available. The
fma version has twice the throughput of the previous version and 70% of
the latency.  The non-fma version has 30% higher throughput and 10%
higher latency.

Max ULP error is 0.949 with fma and 0.792 without fma.

Passes GLIBC testsuite.
sysdeps/ieee754/dbl-64/e_hypot.c