From a6825a427a51da815a805d66ce65c98de282d89b Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 13 Sep 2018 12:24:03 +0100 Subject: COMPMID-1540 Implement YOLOLayer on NEON Change-Id: Ice28996959dc666fff5e8ae486c1ff8093db083f Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/148367 Reviewed-by: Georgios Pinitas Tested-by: bsgcomp --- arm_compute/core/NEON/kernels/NEYOLOLayerKernel.h | 116 +++++++++++++++++++++ .../kernels/detail/NEActivationFunctionDetail.h | 91 ++++++++++++++++ 2 files changed, 207 insertions(+) create mode 100644 arm_compute/core/NEON/kernels/NEYOLOLayerKernel.h (limited to 'arm_compute/core/NEON/kernels') diff --git a/arm_compute/core/NEON/kernels/NEYOLOLayerKernel.h b/arm_compute/core/NEON/kernels/NEYOLOLayerKernel.h new file mode 100644 index 0000000000..c0cfcc049e --- /dev/null +++ b/arm_compute/core/NEON/kernels/NEYOLOLayerKernel.h @@ -0,0 +1,116 @@ +/* + * 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_NEYOLOLAYERKERNEL_H__ +#define __ARM_COMPUTE_NEYOLOLAYERKERNEL_H__ + +#include "arm_compute/core/NEON/INEKernel.h" + +namespace arm_compute +{ +class ITensor; + +/** Interface for the YOLO layer kernel. */ +class NEYOLOLayerKernel : public INEKernel +{ +public: + const char *name() const override + { + return "NEYOLOLayerKernel"; + } + /** Constructor */ + NEYOLOLayerKernel(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEYOLOLayerKernel(const NEYOLOLayerKernel &) = delete; + /** Default move constructor */ + NEYOLOLayerKernel(NEYOLOLayerKernel &&) = default; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEYOLOLayerKernel &operator=(const NEYOLOLayerKernel &) = delete; + /** Default move assignment operator */ + NEYOLOLayerKernel &operator=(NEYOLOLayerKernel &&) = default; + /** Default destructor */ + ~NEYOLOLayerKernel() = default; + /** Set the input and output tensor. + * + * @note If the output tensor is a nullptr or is equal to the input, the activation 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 + * of the activation function. Data types supported: F16/F32. + * @param[out] output Destination tensor. Data type supported: same as @p input + * @param[in] act_info Activation layer parameters. + * @param[in] num_classes Number of classes to activate (must be submultiple of @p input channels) + */ + void configure(ITensor *input, ITensor *output, const ActivationLayerInfo &act_info, int32_t num_classes); + /** Static function to check if given info will lead to a valid configuration of @ref NEYOLOLayerKernel + * + * @param[in] input Source tensor info. In case of @p output tensor info = nullptr, this tensor will store the result + * of the activation function. Data types supported: F16/F32. + * @param[in] output Destination tensor info. Data type supported: same as @p input + * @param[in] act_info Activation layer information. + * @param[in] num_classes Number of classes to activate (must be submultiple of @p input channels) + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info, int32_t num_classes); + + // Inherited methods overridden: + void run(const Window &window, const ThreadInfo &info) override; + +private: +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + /** Function to run YOLO layer on fp16 + * + * @param[in] window Region on which to execute the kernel. + */ + void yolo_layer_fp16_nchw(const Window &window); + /** Function to run batch normalization on fp16 on tensors with NHWC format + * + * @param[in] window Region on which to execute the kernel. + */ + void yolo_layer_fp16_nhwc(const Window &window); +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + /** Function to run YOLO layer on fp32 + * + * @param[in] window Region on which to execute the kernel. + */ + void yolo_layer_fp32_nchw(const Window &window); + /** Function to run YOLO layer on fp32 on tensors with NHWC format + * + * @param[in] window Region on which to execute the kernel. + */ + void yolo_layer_fp32_nhwc(const Window &window); + /** Common signature for all the yolo layer functions + * + * @param[in] window Region on which to execute the kernel. + */ + using YOLOFunctionPtr = void (NEYOLOLayerKernel::*)(const Window &window); + +private: + YOLOFunctionPtr _func; + ITensor *_input; + ITensor *_output; + ActivationLayerInfo _act_info; + int32_t _num_classes; +}; +} // namespace arm_compute +#endif /*__ARM_COMPUTE_NEYOLOLAYERKERNEL_H__ */ diff --git a/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h b/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h index 71d5a9eef7..9344235d09 100644 --- a/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h +++ b/arm_compute/core/NEON/kernels/detail/NEActivationFunctionDetail.h @@ -54,6 +54,97 @@ struct dummy ARM_COMPUTE_UNUSED(vval); } }; +/** Linear activation object */ +template +struct linear +{ + /** NEON vector type. */ + using ExactType = typename wrapper::traits::neon_vector::type; + /** NEON vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_vector::tag_type; + + /** Construct a Linear activation object. + * + * @param[in] act_info Activation layer information. + */ + explicit linear(ActivationLayerInfo act_info) + : valpha(wrapper::vdup_n(static_cast(act_info.a()), ExactTagType{})), + vbeta(wrapper::vdup_n(static_cast(act_info.b()), ExactTagType{})) + { + } + + /** Run activation function. + * + * @param[in] vval Vector of values. + */ + void operator()(ExactType &vval) + { + vval = wrapper::vmla(vval, valpha, vbeta); + } + + /** Vector of alphas. */ + const ExactType valpha; + /** Vector of betas. */ + const ExactType vbeta; +}; +/** Square activation object */ +template +struct square +{ + /** NEON vector type. */ + using ExactType = typename wrapper::traits::neon_vector::type; + /** NEON vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_vector::tag_type; + + /** Construct a Square activation object. + * + * @param[in] act_info Activation layer information. + */ + explicit square(ActivationLayerInfo act_info) + { + ARM_COMPUTE_UNUSED(act_info); + } + + /** Run activation function. + * + * @param[in] vval Vector of values. + */ + void operator()(ExactType &vval) + { + vval = wrapper::vmul(vval, vval); + } +}; +/** Logistic activation object */ +template +struct logistic +{ + /** NEON vector type. */ + using ExactType = typename wrapper::traits::neon_vector::type; + /** NEON vector tag type. */ + using ExactTagType = typename wrapper::traits::neon_vector::tag_type; + + /** Construct a Logistic activation object. + * + * @param[in] act_info Activation layer information. + */ + explicit logistic(ActivationLayerInfo act_info) + : vone(wrapper::vdup_n(static_cast(1.f), ExactTagType{})) + { + ARM_COMPUTE_UNUSED(act_info); + } + + /** Run activation function. + * + * @param[in] vval Vector of values. + */ + void operator()(ExactType &vval) + { + vval = wrapper::vinv(wrapper::vadd(vone, wrapper::vexpq(wrapper::vnegq(vval)))); + } + + /** Vector of ones. */ + const ExactType vone; +}; /** RELU activation object */ template struct relu -- cgit v1.2.1