diff options
author | Sang-Hoon Park <sang-hoon.park@arm.com> | 2020-10-13 23:34:09 +0100 |
---|---|---|
committer | Sang-Hoon Park <sang-hoon.park@arm.com> | 2020-10-16 08:43:24 +0000 |
commit | e7280585b317f695a932be5981895377e174946d (patch) | |
tree | 3d5662a779d2c3bee9e45a2db27a27cd0275c9d5 /src/core/NEON/wrapper/intrinsics/intrinsics.h | |
parent | cf343e3798d2a8c2ad2fcac488e4b78e2b5c968d (diff) | |
download | ComputeLibrary-e7280585b317f695a932be5981895377e174946d.tar.gz |
COMPMID-3805: Fix SQRT non-zero output for zero input
- For AArch64, NEActivationLayerKernel uses vsqrt rather than
vinvsqrt.
- For non-AArch64, it masks values to ensure zero input
results in zero output without producing NaN.
- Test cases for FP16 and FP32's positive boundary values
are added.
Change-Id: Ic0104ee5d7045059c2e9bd052616a4a3b43a315d
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4150
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/NEON/wrapper/intrinsics/intrinsics.h')
-rw-r--r-- | src/core/NEON/wrapper/intrinsics/intrinsics.h | 1 |
1 files changed, 1 insertions, 0 deletions
diff --git a/src/core/NEON/wrapper/intrinsics/intrinsics.h b/src/core/NEON/wrapper/intrinsics/intrinsics.h index 495321a6a1..070f3c7065 100644 --- a/src/core/NEON/wrapper/intrinsics/intrinsics.h +++ b/src/core/NEON/wrapper/intrinsics/intrinsics.h @@ -66,6 +66,7 @@ #include "src/core/NEON/wrapper/intrinsics/round.h" #include "src/core/NEON/wrapper/intrinsics/setlane.h" #include "src/core/NEON/wrapper/intrinsics/sin.h" +#include "src/core/NEON/wrapper/intrinsics/sqrt.h" #include "src/core/NEON/wrapper/intrinsics/store.h" #include "src/core/NEON/wrapper/intrinsics/sub.h" #include "src/core/NEON/wrapper/intrinsics/tanh.h" |