diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-05-12 21:03:56 +0100 |
---|---|---|
committer | TeresaARM <teresa.charlinreyes@arm.com> | 2020-05-13 08:52:14 +0000 |
commit | 2cd7a37af612723b712ca37d5fa99e17a84d7425 (patch) | |
tree | a50372dc04aef20d49afffe027b9a6364031b8d2 /src/core | |
parent | a208a808363195978188803b27a0de2a57d3e77d (diff) | |
download | ComputeLibrary-2cd7a37af612723b712ca37d5fa99e17a84d7425.tar.gz |
COMPMID-3464: Address NESoftmaxLayer failures for QASYMM8_SIGNED
Normalization with the maximum value was causing results to wrap-around
As a work-around we use saturating intrinsics to perform the operation
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I719b7ac7ad274dc2ae339bc4a055f9200134ed97
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3184
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core')
-rw-r--r-- | src/core/NEON/kernels/NESoftmaxLayerKernel.cpp | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp index 790c8bacc5..41bf03ad1d 100644 --- a/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp +++ b/src/core/NEON/kernels/NESoftmaxLayerKernel.cpp @@ -311,7 +311,7 @@ void logits_1d_softmax_qasymm8(const ITensor &in, const ITensor &max, void *cons for(; x <= (input_width - vec_size); x += vec_size) { auto vec_elements = wrapper::vloadq(in_ptr + x); - vec_elements = wrapper::vsub(vec_max, vec_elements); + vec_elements = wrapper::vqsub(vec_max, vec_elements); auto vec_elements_flt = convert_int_to_float<float32x4x4_t>(vec_elements); if(is_log) |