diff options
| author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-01-14 16:17:48 +0000 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2021-01-15 19:32:14 +0000 |
| commit | 6c25aad49787676c8dcaab74efd536690184ed04 (patch) | |
| tree | 3034bb24554d1c52077fcf20e38bd655e230bbe2 /arm_compute/core/ITensorPack.h | |
| parent | ddd79f5a1e92bfea85ad2104f15c991952f65ef8 (diff) | |
| download | ComputeLibrary-6c25aad49787676c8dcaab74efd536690184ed04.tar.gz | |
Introduce threshold for NEON floating-point SOFT_RELU activation
Large inputs produce incorrect results of Soft ReLu activations where
the output saturates at around 88.72283. This is due to the
approximation algorithm used for the logarithm.
For this reason, we introduce a threshold such that with x > threshold,
Soft ReLu activation will return the value x itself.
SVE does not seem to suffer from the same issue, hence NEON kernels only
are modified.
Resolves COMPMID-4091
Change-Id: I357883deed4e4aba571a1d3163267772096c0412
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4865
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'arm_compute/core/ITensorPack.h')
0 files changed, 0 insertions, 0 deletions