From 47d39dc615d1dee2482bc84699802165a9778ac8 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 11 Mar 2019 14:03:23 +0000 Subject: COMPMID-1975: Update depthwise convolution. Change-Id: Iad58672be35710a7ec2e918653d6d529709387e8 Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/898 Tested-by: Arm Jenkins Reviewed-by: Giuseppe Rossini Comments-Addressed: Arm Jenkins Reviewed-by: Gian Marco Iodice --- .../NEON/functions/NEDepthwiseConvolutionLayer.h | 52 ++++++++- .../NEDepthwiseConvolutionAssemblyDispatch.h | 117 +++++++++++++++++++++ 2 files changed, 165 insertions(+), 4 deletions(-) create mode 100644 arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h (limited to 'arm_compute/runtime') diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h index e2fe11ea7f..28f0560e93 100644 --- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -37,6 +37,7 @@ #include "arm_compute/runtime/MemoryGroup.h" #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" #include "arm_compute/runtime/NEON/functions/NEPermute.h" +#include "arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h" #include "arm_compute/runtime/Tensor.h" namespace arm_compute @@ -53,7 +54,15 @@ class NEDepthwiseConvolutionLayer3x3 : public IFunction { public: /** Default constructor */ - NEDepthwiseConvolutionLayer3x3(); + NEDepthwiseConvolutionLayer3x3(std::shared_ptr memory_manager = nullptr); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDepthwiseConvolutionLayer3x3(const NEDepthwiseConvolutionLayer3x3 &) = delete; + /** Default move constructor */ + NEDepthwiseConvolutionLayer3x3(NEDepthwiseConvolutionLayer3x3 &&) = default; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDepthwiseConvolutionLayer3x3 &operator=(const NEDepthwiseConvolutionLayer3x3 &) = delete; + /** Default move assignment operator */ + NEDepthwiseConvolutionLayer3x3 &operator=(NEDepthwiseConvolutionLayer3x3 &&) = default; /** Initialize the function's source, destination, kernels and border_size. * * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling). @@ -86,9 +95,44 @@ public: // Inherited methods overriden: void run() override; + void prepare() override; private: + /** Configure the kernels/functions for the generic pipeline. + * + * @param[in, out] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling). + * @param[in] weights Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] biases (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input. + * @param[out] output Destination tensor. Data type supported: same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] act_info Activation layer information in case of a fused activation. + */ + void configure_generic(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, const ActivationLayerInfo &act_info); + /** Configure the kernels/functions for the optimized pipeline. + * + * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling). + * @param[in] weights Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] biases (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input. + * @param[out] output Destination tensor. Data type supported: same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] act_info Activation layer information in case of a fused activation. + */ + void configure_optimized(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, + unsigned int depth_multiplier, const ActivationLayerInfo &act_info); + /** Run generic kernel */ + void run_generic(); + /** Run optimized function */ + void run_optimized(); + +private: + MemoryGroup _memory_group; NEDepthwiseConvolutionLayer3x3Kernel _dwc_kernel; + NEDepthwiseConvolutionAssemblyDispatch _dwc_optimized_func; NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel; NEFillBorderKernel _border_handler; NEPermute _permute_input; @@ -99,14 +143,14 @@ private: Tensor _permuted_input; Tensor _permuted_weights; Tensor _permuted_output; + const ITensor *_original_weights; bool _has_bias; bool _is_quantized; bool _is_optimized; - bool _are_weights_reshaped; bool _is_nchw; - bool _is_first_run; bool _permute; bool _is_activationlayer_enabled; + bool _is_prepared; }; /** Basic function to execute a generic depthwise convolution. This function calls the following NEON kernels: diff --git a/arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h b/arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h new file mode 100644 index 0000000000..df8f29d2c7 --- /dev/null +++ b/arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h @@ -0,0 +1,117 @@ +/* + * Copyright (c) 2019 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_NEDEPTHWISECONVOLUTIONASSEMBLYDISPATCH_H__ +#define __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONASSEMBLYDISPATCH_H__ + +#include "arm_compute/runtime/IFunction.h" + +#include "arm_compute/runtime/IMemoryManager.h" +#include "arm_compute/runtime/MemoryGroup.h" +#include "arm_compute/runtime/Tensor.h" + +#include "arm_compute/core/NEON/kernels/assembly/NEDepthwiseConvolutionAssemblyKernelWrapper.h" +#include "arm_compute/core/NEON/kernels/convolution/depthwise/depthwise.hpp" + +namespace arm_compute +{ +/** Depthwise convolution assembly kernel glue */ +class NEDepthwiseConvolutionAssemblyDispatch : public IFunction +{ +public: + /** Default constructor + * + * @param[in,out] memory_manager Memory manager to use + */ + NEDepthwiseConvolutionAssemblyDispatch(std::shared_ptr memory_manager = nullptr); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDepthwiseConvolutionAssemblyDispatch(const NEDepthwiseConvolutionAssemblyDispatch &) = delete; + /** Default move constructor */ + NEDepthwiseConvolutionAssemblyDispatch(NEDepthwiseConvolutionAssemblyDispatch &&) = default; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + NEDepthwiseConvolutionAssemblyDispatch &operator=(const NEDepthwiseConvolutionAssemblyDispatch &) = delete; + /** Default move assignment operator */ + NEDepthwiseConvolutionAssemblyDispatch &operator=(NEDepthwiseConvolutionAssemblyDispatch &&) = default; + /** Initialize the function's source, destination, kernels and border_size. + * + * @note Supports only NHWC format + * + * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling). + * @param[in] weights Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] bias (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input. + * @param[out] output Destination tensor. Data type supported: same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + */ + void configure(const ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, + const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionAssemblyDispatch + * + * @note Supports only NHWC format + * + * @param[in] input Source tensor. Data type supported: QASYMM8/F16/F32. (Written to only for border filling). + * @param[in] weights Weights tensor. These are 3D tensors with shape [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] bias (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input. + * @param[out] output Destination tensor. Data type supported: same as @p input. + * @param[in] conv_info Padding and stride information to use for the convolution. + * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + * + * @return An error status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, + const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + /** Check if the optimized kernel can be used for the given kernel sizes and strides + * + * @warning Even if this return true the inputs and outputs might need to get permuted as the only layout supported is NHWC + * + * @param[in] input Input tensor info. + * @param[in] weights Weights tensor info. + * @param[in] conv_info Convolution layer metadata. + * @param[in] depth_multiplier (Optional) Depth multiplier to be used. + * + * @return True if the assembly kernel could be used else false. Note that transformations of input/output could be needed. + */ + static bool is_optimized_supported(const ITensorInfo *input, const ITensorInfo *weights, PadStrideInfo conv_info, unsigned int depth_multiplier = 1); + + // Inherited methods overridden: + void run() override; + void prepare() override; + +private: + MemoryGroup _memory_group; + const ITensor *_input; + const ITensor *_weights; + const ITensor *_bias; + ITensor *_output; + Tensor _packed_weights; + Tensor _workspace; + bool _is_prepared; + std::unique_ptr _dwc_assembly_kernel; + NEDepthwiseConvolutionAssemblyKernelWrapper _dwc_acl_kernel; +}; +} // namespace arm_compute +#endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONASSEMBLYDISPATCH_H__ */ -- cgit v1.2.1