diff options
author | Joe Ramsay <Joe.Ramsay@arm.com> | 2022-10-19 14:51:12 +0100 |
---|---|---|
committer | Joe Ramsay <joe.ramsay@arm.com> | 2022-10-19 14:51:12 +0100 |
commit | 23f1039a7a4d43c0b45991374e99f1cf757fc1fe (patch) | |
tree | 5fd90ff177cee9a98b9a8f8d35e31ace58466356 /pl/math/test/runulp.sh | |
parent | 222183caf938d3b9626623b0a4949dc77abd5599 (diff) | |
download | arm-optimized-routines-23f1039a7a4d43c0b45991374e99f1cf757fc1fe.tar.gz |
pl/math: Fix subnormal handling in atan2
Previously subnormals were handled specially when determining from the
exponent whether to shortcut calculating x^8. This was not the correct
approach - to compare with the threshold in such a way that matches
the reference they should be treated the name as normal exponents.
Larger errors than previously thought were found in the subnormal
region - a new test has been added in runulp and the threshold has
been updated.
Diffstat (limited to 'pl/math/test/runulp.sh')
-rwxr-xr-x | pl/math/test/runulp.sh | 13 |
1 files changed, 7 insertions, 6 deletions
diff --git a/pl/math/test/runulp.sh b/pl/math/test/runulp.sh index 2a79a27..059be60 100755 --- a/pl/math/test/runulp.sh +++ b/pl/math/test/runulp.sh @@ -74,12 +74,13 @@ t erfcf 0x1p-26 0x1p5 40000 t erfcf -0x1p-26 -0x1p3 40000 t erfcf 0 inf 40000 -L=1.5 -t atan2 -10.0 10.0 50000 -t atan2 -1.0 1.0 40000 -t atan2 0.0 1.0 40000 -t atan2 1.0 100.0 40000 -t atan2 1e6 1e32 40000 +L=1.74 +t atan2 -10.0 10.0 50000 +t atan2 -1.0 1.0 40000 +t atan2 0.0 1.0 40000 +t atan2 1.0 100.0 40000 +t atan2 1e6 1e32 40000 +t atan2 0x1p-1022 0x1p-1000 x 0 0x1p-1022 40000 # Regression-test for correct NaN handling check atan2 0x1.7887a0a717aefp+1017 0x1.7887a0a717aefp+1017 x -nan -nan check atan2 nan nan x -nan -nan |