benchtests: Add fmaxf/fminf benchmarks
commitda16c9b524908fe0353efc4af8eab6a5ad5ea50b
authorAdhemerval Zanella <adhemerval.zanella@linaro.org>
Mon, 19 Dec 2016 13:00:45 +0000 (19 11:00 -0200)
committerAdhemerval Zanella <adhemerval.zanella@linaro.org>
Mon, 19 Dec 2016 18:04:16 +0000 (19 16:04 -0200)
tree880d9f65a8cc14dd76bca5d0a1e613f9a16f6db9
parent5d1f604a87214e133d0adf9dbd53891138b5276c
benchtests: Add fmaxf/fminf benchmarks

This patch adds fmaxf and fminf benchtests.  It is based on
math/s_fmax_template.c implementation which checks for basically four
different classes:

  1. if x is greater or equal than y.
  2. if x is less than y.
  3. if x or y is signaling.
  4. if y is nan.

Cases 1 and 2 are used for default input number (by mixing normal double
numbers and infinity), while case 3 and 4 are used each for on for a
benchmark class.

Checked on x86_64-linux-gnu and powerpc64-linux-gnu.

* benchtests/Makefile (bench-math): Add fminf and fmaxf.
(CFLAGS-bench-fmaxf.c): New rule.
(CFLAGS-bench-fminf.c): Likewise.
        * benchtests/fmaxf-inputs: New file.
        * benchtests/fminf-inputs: Likewise.
ChangeLog
benchtests/Makefile
benchtests/fmaxf-inputs [new file with mode: 0644]
benchtests/fminf-inputs [new file with mode: 0644]