diff options
| author | SiCong Li <sicong.li@arm.com> | 2020-06-08 17:30:51 +0100 |
|---|---|---|
| committer | SiCong Li <sicong.li@arm.com> | 2020-06-10 17:33:25 +0000 |
| commit | a32e2aef81cfcba9f5ae1770ceeb4a8d26fdc1f4 (patch) | |
| tree | b791e466df9dff4032be409d6765cb8a9af9319d /src/core/NEON/kernels/NEActivationLayerKernel.cpp | |
| parent | 8aa8764982d23ed8b8c8810bbfda30542f21e034 (diff) | |
| download | ComputeLibrary-a32e2aef81cfcba9f5ae1770ceeb4a8d26fdc1f4.tar.gz | |
COMPMID-3523: Fix validation fails on armv8.2-a
* Fix neon sqrt activation delta(epsilon)
* Fix NEON Hard Swish validation tolerance
* Fix NEON FP16 LogSoftmaxLayer validation test typo
* Raise NEON reduction (sum) f16 tolerance
Change-Id: Ia33d69ce5f0b78be1893fb8e13d2761a8e7fceff
Signed-off-by: SiCong Li <sicong.li@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3318
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/NEActivationLayerKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEActivationLayerKernel.cpp | 9 |
1 files changed, 7 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/NEActivationLayerKernel.cpp b/src/core/NEON/kernels/NEActivationLayerKernel.cpp index 8e91e6b4d1..ffbfd710f9 100644 --- a/src/core/NEON/kernels/NEActivationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEActivationLayerKernel.cpp @@ -265,7 +265,12 @@ NEActivationLayerKernel::activation(const Window &window) Iterator input(_input, win_collapsed); Iterator output(_output, win_collapsed); - const auto epsilon = wrapper::vdup_n(static_cast<T>(1e-24), ExactTagType{}); + // A small delta added to the input to prevent NAN values caused by zeros in inputs to SQRT +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + const auto delta = wrapper::vdup_n(static_cast<T>(1e-7), ExactTagType{}); +#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + const auto delta = wrapper::vdup_n(static_cast<T>(1e-24), ExactTagType{}); +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const auto const_1 = wrapper::vdup_n(static_cast<T>(1.f), ExactTagType{}); const auto const_0 = wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{}); const auto const_6 = wrapper::vdup_n(static_cast<T>(6.f), ExactTagType{}); @@ -318,7 +323,7 @@ NEActivationLayerKernel::activation(const Window &window) tmp = wrapper::vbsl(wrapper::vcge(vin, const_0), vin, wrapper::vmul(va, wrapper::vsub(wrapper::vexpq(vin), const_1))); break; case ActivationFunction::SQRT: - tmp = wrapper::vinv(wrapper::vinvsqrt(vin + epsilon)); + tmp = wrapper::vinv(wrapper::vinvsqrt(wrapper::vadd(vin, delta))); break; case ActivationFunction::SQUARE: tmp = wrapper::vmul(vin, vin); |