diff options
| author | Michalis Spyrou <michalis.spyrou@arm.com> | 2020-11-22 00:49:42 +0000 |
|---|---|---|
| committer | Michalis Spyrou <michalis.spyrou@arm.com> | 2020-12-14 16:02:26 +0000 |
| commit | aa51a5ba9a3f05be08b94859b53c398edee5d2e3 (patch) | |
| tree | b28829b483421b210cd7c8a256c7feafed736b36 /src/core/NEON/kernels/NEActivationLayerKernel.cpp | |
| parent | 3737c7934da929003bda446291489cf352e43751 (diff) | |
| download | ComputeLibrary-aa51a5ba9a3f05be08b94859b53c398edee5d2e3.tar.gz | |
COMPMID-3870: Create ActivationLayer SVE/SVE2
Adds support for ActivationLayer for SVE and SVE2.
Datatypes supported:
*FP32
*FP16
*QASYMM8
*QASYMM8_SIGNED
*QSYMM16
Change-Id: Ia3583891795cda4ca2f9fa27c440731a5c27710d
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4566
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/NEActivationLayerKernel.cpp')
| -rw-r--r-- | src/core/NEON/kernels/NEActivationLayerKernel.cpp | 33 |
1 files changed, 32 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/NEActivationLayerKernel.cpp b/src/core/NEON/kernels/NEActivationLayerKernel.cpp index d969fd8e38..f215787bf6 100644 --- a/src/core/NEON/kernels/NEActivationLayerKernel.cpp +++ b/src/core/NEON/kernels/NEActivationLayerKernel.cpp @@ -56,6 +56,18 @@ struct ActivationKernel static const ActivationKernel available_kernels[] = { +#if defined(__ARM_FEATURE_SVE) + { + "fp16_sve_activation", + [](const ActivationSelectorData & data) { return data.dt == DataType::F16; }, + REGISTER_FP16_SVE(arm_compute::cpu::fp16_sve_activation) + }, + { + "fp32_sve_activation", + [](const ActivationSelectorData & data) { return data.dt == DataType::F32; }, + REGISTER_FP32_SVE(arm_compute::cpu::fp32_sve_activation) + }, +#else /* !defined(__ARM_FEATURE_SVE) */ { "fp16_neon_activation", [](const ActivationSelectorData & data) { return data.dt == DataType::F16; }, @@ -66,6 +78,25 @@ static const ActivationKernel available_kernels[] = [](const ActivationSelectorData & data) { return data.dt == DataType::F32; }, REGISTER_FP32_NEON(arm_compute::cpu::fp32_neon_activation) }, +#endif /* defined(__ARM_FEATURE_SVE) */ + +#if defined(__ARM_FEATURE_SVE2) /* defined(__ARM_FEATURE_SVE2) */ + { + "qasymm8_sve_activation", + [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8; }, + REGISTER_QASYMM8_SVE(arm_compute::cpu::qasymm8_sve_activation) + }, + { + "qasymm8_signed_sve_activation", + [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8_SIGNED; }, + REGISTER_QASYMM8_SIGNED_SVE(arm_compute::cpu::qasymm8_signed_sve_activation) + }, + { + "qsymm16_sve_activation", + [](const ActivationSelectorData & data) { return data.dt == DataType::QSYMM16; }, + REGISTER_QSYMM16_SVE(arm_compute::cpu::qsymm16_sve_activation) + }, +#else /* !defined(__ARM_FEATURE_SVE2) */ { "qasymm8_neon_activation", [](const ActivationSelectorData & data) { return data.dt == DataType::QASYMM8; }, @@ -81,6 +112,7 @@ static const ActivationKernel available_kernels[] = [](const ActivationSelectorData & data) { return data.dt == DataType::QSYMM16; }, REGISTER_QSYMM16_NEON(arm_compute::cpu::qsymm16_neon_activation) }, +#endif /* defined(__ARM_FEATURE_SVE2) */ }; const ActivationKernel *get_implementation(const ActivationSelectorData &data) @@ -159,7 +191,6 @@ std::pair<Status, Window> validate_and_configure_window(const ITensorInfo *input // Output auto inizialitation if not yet initialized auto_init_if_empty(*output, *input->clone()); - // NEActivationLayerKernel doesn't need padding so update_window_and_padding() can be skipped Coordinates coord; coord.set_num_dimensions(output->num_dimensions()); output->set_valid_region(ValidRegion(coord, output->tensor_shape())); |