diff options
| author | Pierre Blanchard <pierre.blanchard@arm.com> | 2020-10-29 15:50:19 +0000 |
|---|---|---|
| committer | Szabolcs Nagy <szabolcs.nagy@arm.com> | 2020-10-29 16:36:28 +0000 |
| commit | 1a79237df2fc1020ccce8528ed54977681302752 (patch) | |
| tree | ecc2bbb51711e7cd12e36658e9f3c47d401b8139 /math/test/runulp.sh | |
| parent | 0f4ae0c5b561de25acb10130fd5e473ec038f89d (diff) | |
| download | arm-optimized-routines-1a79237df2fc1020ccce8528ed54977681302752.tar.gz | |
math: add scalar erff.
In round-to-nearest mode the maximum error is 1.09 ULP.
Compared to glibc-2.28 erff: throughput is about 2.2x better,
latency is about 1.5x better on some AArch64 cores (on random
input in [-4,4]).
There are further optimization and quality improvement opportunities.
Diffstat (limited to 'math/test/runulp.sh')
| -rwxr-xr-x | math/test/runulp.sh | 11 |
1 files changed, 11 insertions, 0 deletions
diff --git a/math/test/runulp.sh b/math/test/runulp.sh index a8c391b..283145e 100755 --- a/math/test/runulp.sh +++ b/math/test/runulp.sh @@ -119,6 +119,17 @@ t powf 0x1p-70 0x1p70 x 0x1p-1 0x1p1 50000 t powf 0x1p-70 0x1p70 x -0x1p-1 -0x1p1 50000 t powf 0x1.ep-1 0x1.1p0 x 0x1p8 0x1p14 50000 t powf 0x1.ep-1 0x1.1p0 x -0x1p8 -0x1p14 50000 + +L=0.6 +Ldir=0.9 +t erff 0 0xffff0000 10000 +t erff 0x1p-127 0x1p-26 40000 +t erff -0x1p-127 -0x1p-26 40000 +t erff 0x1p-26 0x1p3 40000 +t erff -0x1p-26 -0x1p3 40000 +t erff 0 inf 40000 +Ldir=0.5 + done # vector functions |