diff options
| author | Viet-Hoa Do <viet-hoa.do@arm.com> | 2022-06-21 15:56:15 +0100 |
|---|---|---|
| committer | Viet-Hoa Do <viet-hoa.do@arm.com> | 2022-06-29 13:30:06 +0000 |
| commit | b042e39060901b44e615b923b5723c04d9b42a95 (patch) | |
| tree | e23fd9b89c753f9731e1e8ec4a0d9ca468f9f683 /src/cpu/kernels/CpuActivationKernel.cpp | |
| parent | 13f96d0a5efc140785a6de58bff9b24b80dd0cfd (diff) | |
| download | ComputeLibrary-b042e39060901b44e615b923b5723c04d9b42a95.tar.gz | |
Add LUT-based leaky relu for QASYMM8 on CPU
* Add LUT generation function for Leaky ReLU.
* Some additional changes in the existing LUT implementation:
+ Bring back the NEON implementation of hard swish for 32-bit
build. Library size of 64-bit build is not affected.
+ Add some extra #ifdef to remove unnecessary code in 32-bit
build.
Resolves: COMPMID-5386
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I1ea49611cc922765ee741e31138c888401d33e9b
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7845
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/cpu/kernels/CpuActivationKernel.cpp')
| -rw-r--r-- | src/cpu/kernels/CpuActivationKernel.cpp | 16 |
1 files changed, 10 insertions, 6 deletions
diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp index 330de1ece2..9eaf44af51 100644 --- a/src/cpu/kernels/CpuActivationKernel.cpp +++ b/src/cpu/kernels/CpuActivationKernel.cpp @@ -45,11 +45,13 @@ namespace { static const std::vector<CpuActivationKernel::ActivationKernel> available_kernels = { - { // neon LUT implementantion of HARD_SWISH takes precedence - "neon_qu8_activation_hardswish_lut", - [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.f == ActivationLayerInfo::ActivationFunction::HARD_SWISH; }, - REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_hardswish_lut) +#ifdef __aarch64__ + { // Neon LUT implementantion takes precedence + "neon_qu8_activation_lut", + [](const ActivationDataTypeISASelectorData & data) { return ActivationLayerInfo::is_lut_supported(data.f, data.dt); }, + REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation_lut) }, +#endif // __aarch64__ { "sve2_qu8_activation", [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.isa.sve2; }, @@ -87,7 +89,7 @@ static const std::vector<CpuActivationKernel::ActivationKernel> available_kernel }, { "neon_qu8_activation", - [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8 && data.f != ActivationLayerInfo::ActivationFunction::HARD_SWISH; }, + [](const ActivationDataTypeISASelectorData & data) { return data.dt == DataType::QASYMM8; }, REGISTER_QASYMM8_NEON(arm_compute::cpu::neon_qasymm8_activation) }, { @@ -188,10 +190,12 @@ void CpuActivationKernel::configure(const ITensorInfo *src, ITensorInfo *dst, Ac _run_method = uk->ukernel; _name = std::string("CpuActivationKernel").append("/").append(uk->name); - if(activation_info.activation() == ActivationLayerInfo::ActivationFunction::HARD_SWISH && src->data_type() == DataType::QASYMM8) +#ifdef __aarch64__ + if(ActivationLayerInfo::is_lut_supported(activation_info.activation(), src->data_type())) { activation_info.init_lut(src->quantization_info().uniform(),(dst)?dst->quantization_info().uniform():src->quantization_info().uniform()); } +#endif // __aarch64__ _act_info = activation_info; // Configure kernel window |