diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2018-02-15 12:29:44 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:47:18 +0000 |
commit | 57c033bb5400ef19e5952f191da3e878e21bba91 (patch) | |
tree | b325e4a0beba35bcdf29c4ae6dea874d7cd26b9f /arm_compute/core/NEON/kernels | |
parent | 02ee4291795f64fb510a71c6c754671438635186 (diff) | |
download | ComputeLibrary-57c033bb5400ef19e5952f191da3e878e21bba91.tar.gz |
COMPMID-906: Use fused activation in NEON Batch normalization
Change-Id: I5a6413548b2c9b8972c91ddba57395509dffd87e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/120656
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Diffstat (limited to 'arm_compute/core/NEON/kernels')
-rw-r--r-- | arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h | 107 | ||||
-rw-r--r-- | arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h | 113 |
2 files changed, 191 insertions, 29 deletions
diff --git a/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h b/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h index f748830b81..63eb739487 100644 --- a/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h +++ b/arm_compute/core/NEON/kernels/NEBatchNormalizationLayerKernel.h @@ -55,49 +55,98 @@ public: * * @note If the output tensor is a nullptr, the batch normalization function will be performed in-place * - * @param[in, out] input Source tensor. In case of @p output tensor = nullptr, this tensor will store the result. - * 3 lower dimensions represent a single input with dimensions [width, height, FM]. - * The rest are optional and used for representing batches. Data types supported: QS8/QS16/F16/F32. - * @param[out] output Destination tensor. Output will have the same number of dimensions as input. Data type supported: same as @p input - * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] beta Beta values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] gamma Gamma values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] epsilon Small value to avoid division with zero. + * @param[in, out] input Source tensor. In case of @p output tensor = nullptr, this tensor will store the result. + * 3 lower dimensions represent a single input with dimensions [width, height, FM]. + * The rest are optional and used for representing batches. Data types supported: QS8/QS16/F16/F32. + * @param[out] output Destination tensor. Output will have the same number of dimensions as input. Data type supported: same as @p input + * @param[in] mean Mean values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] var Variance values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] beta Beta values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] gamma Gamma values tensor. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] epsilon Small value to avoid division with zero. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. + * Data types supported: F32 */ - void configure(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon); + void configure(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon, + ActivationLayerInfo act_info = ActivationLayerInfo()); /** Static function to check if given info will lead to a valid configuration of @ref NEBatchNormalizationLayerKernel * - * @param[in] input Source tensor info. In case of @p output tensor = nullptr, this tensor will store the result. - * 3 lower dimensions represent a single input with dimensions [width, height, FM]. - * The rest are optional and used for representing batches. Data types supported: QS8/QS16/F16/F32. - * @param[in] output Destination tensor info. Output will have the same number of dimensions as input. Data type supported: same as @p input - * @param[in] mean Mean values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] var Variance values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] beta Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] gamma Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input - * @param[in] epsilon Small value to avoid division with zero. + * @param[in] input Source tensor info. In case of @p output tensor = nullptr, this tensor will store the result. + * 3 lower dimensions represent a single input with dimensions [width, height, FM]. + * The rest are optional and used for representing batches. Data types supported: QS8/QS16/F16/F32. + * @param[in] output Destination tensor info. Output will have the same number of dimensions as input. Data type supported: same as @p input + * @param[in] mean Mean values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] var Variance values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] beta Beta values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] gamma Gamma values tensor info. 1 dimension with size equal to the feature maps [FM]. Data types supported: Same as @p input + * @param[in] epsilon Small value to avoid division with zero. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. + * Data types supported: F32 * * @return a status */ static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *mean, const ITensorInfo *var, const ITensorInfo *beta, const ITensorInfo *gamma, - float epsilon); + float epsilon, ActivationLayerInfo act_info); // Inherited methods overridden: void run(const Window &window, const ThreadInfo &info) override; private: - using BatchNormFunction = void(ITensor *input, ITensor *output, const ITensor *mean, const ITensor *var, const ITensor *beta, const ITensor *gamma, float epsilon, const Window &window); - BatchNormFunction *_func; - ITensor *_input; - ITensor *_output; - const ITensor *_mean; - const ITensor *_var; - const ITensor *_gamma; - const ITensor *_beta; - float _epsilon; + /** Configure execution function in case of non-fused activation **/ + void configure_non_fused(); + /** Configure execution function in case of fused activation **/ + void configure_fused(); + /** Template function to run batch normalization on 8-bit fixed point + * + * @tparam fused_activation Boolean that flags if its a fused activation or not + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + template <bool fused_activation> + void batch_normalization_qs8(const Window &window); + /** Template function to run batch normalization on 16-bit fixed point + * + * @tparam fused_activation Boolean that flags if its a fused activation or not + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + template <bool fused_activation> + void batch_normalization_qs16(const Window &window); + /** Template function to run batch normalization on fp16 + * + * @tparam fused_activation Boolean that flags if its a fused activation or not + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + template <bool fused_activation> + void batch_normalization_fp16(const Window &window); + /** Template function to run batch normalization on fp32 + * + * @tparam fused_activation Boolean that flags if its a fused activation or not + * @tparam F Activation function functor to run + * + * @param[in] window Region on which to execute the kernel. (Must be a valid region of the window returned by window()). + */ + template <bool fused_activation, typename F> + void batch_normalization_fp32(const Window &window); + /** Common signature for all the batch normalization functions + * + * @param[in] window Region on which to execute the kernel. + */ + using BatchNormFunctionPtr = void (NEBatchNormalizationLayerKernel::*)(const Window &window); + +private: + BatchNormFunctionPtr _func; + ITensor *_input; + ITensor *_output; + const ITensor *_mean; + const ITensor *_var; + const ITensor *_gamma; + const ITensor *_beta; + float _epsilon; + ActivationLayerInfo _act_info; }; } // namespace arm_compute #endif /*__ARM_COMPUTE_NEBATCHNORMALIZATIONLAYERKERNEL_H__ */ diff --git a/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h b/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h new file mode 100644 index 0000000000..e4d3f54943 --- /dev/null +++ b/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h @@ -0,0 +1,113 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_DETAIL_NEACTIVATION_FUNCTION_DETAIL_H__ +#define __ARM_COMPUTE_DETAIL_NEACTIVATION_FUNCTION_DETAIL_H__ + +#include "arm_compute/core/NEON/wrapper/wrapper.h" + +namespace arm_compute +{ +namespace detail +{ +// Dummy activation object +/** Dummy activation object */ +template <typename T, int S> +struct dummy +{ + using ExactType = typename wrapper::traits::neon_vector<T, S>::type; + + explicit dummy(ActivationLayerInfo act_info) + { + ARM_COMPUTE_UNUSED(act_info); + } + void operator()(ExactType &vval) + { + ARM_COMPUTE_UNUSED(vval); + } +}; +/** RELU activation object */ +template <typename T, int S> +struct relu +{ + using ExactType = typename wrapper::traits::neon_vector<T, S>::type; + using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type; + + explicit relu(ActivationLayerInfo act_info) + : vzero(wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{})) + { + ARM_COMPUTE_UNUSED(act_info); + } + + void operator()(ExactType &vval) + { + vval = wrapper::vmax(vzero, vval); + } + + const ExactType vzero; +}; +/** Bounded RELU activation object */ +template <typename T, int S> +struct brelu +{ + using ExactType = typename wrapper::traits::neon_vector<T, S>::type; + using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type; + + explicit brelu(ActivationLayerInfo act_info) + : vzero(wrapper::vdup_n(static_cast<T>(0.f), ExactTagType{})), + valpha(wrapper::vdup_n(static_cast<T>(act_info.a()), ExactTagType{})) + { + } + + void operator()(ExactType &vval) + { + vval = wrapper::vmin(valpha, wrapper::vmax(vzero, vval)); + } + + const ExactType vzero; + const ExactType valpha; +}; +/** Lower-Upper Bounded RELU activation object */ +template <typename T, int S> +struct lubrelu +{ + using ExactType = typename wrapper::traits::neon_vector<T, S>::type; + using ExactTagType = typename wrapper::traits::neon_vector<T, S>::tag_type; + + explicit lubrelu(ActivationLayerInfo act_info) + : valpha(wrapper::vdup_n(static_cast<T>(act_info.a()), ExactTagType{})), + vbeta(wrapper::vdup_n(static_cast<T>(act_info.b()), ExactTagType{})) + { + } + + void operator()(ExactType &vval) + { + vval = wrapper::vmin(valpha, wrapper::vmax(vbeta, vval)); + } + + const ExactType valpha; + const ExactType vbeta; +}; +} // namespace detail +} // namespace arm_compute +#endif /* __ARM_COMPUTE_DETAIL_NEACTIVATION_FUNCTION_DETAIL_H__ */ |