Pablo Tello | 8f43d74 | 2019-03-27 09:28:32 +0000 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2019 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | |
| 25 | #pragma once |
| 26 | |
| 27 | #include <utility> |
| 28 | |
| 29 | #include "arm_gemm_local.hpp" |
| 30 | #include "arm_gemm.hpp" |
| 31 | #include "winograd.hpp" |
| 32 | |
| 33 | namespace winograd |
| 34 | { |
| 35 | |
| 36 | |
| 37 | class IWinogradConvolutionLayer |
| 38 | { |
| 39 | public: |
| 40 | virtual ~IWinogradConvolutionLayer() = default; |
| 41 | |
| 42 | virtual unsigned int weight_transform_get_window(void) const = 0; |
| 43 | virtual void weight_transform_run(unsigned int start, unsigned int stop) = 0; |
| 44 | |
| 45 | virtual ITransform& input_transform(void) = 0; // Expose the input transform |
| 46 | virtual ITransform& output_transform(void) = 0; // Expose the output transform |
| 47 | virtual arm_gemm::IGemmCommon *gemm(void) = 0; // Expose the underlying GEMM |
| 48 | }; |
| 49 | |
| 50 | /** Example of how to construct an ACL-like interface. |
| 51 | * |
| 52 | * Use `get_weight_storage_size`, `get_input_storage_size` and |
| 53 | * `get_output_storage_size` to allocate memory for the convolution engine. |
| 54 | * Then create a `WinogradConvolutionLayer`. |
| 55 | * |
| 56 | * Initialise the weights using `weights_transform.run(...)`. |
| 57 | * |
| 58 | * For each inference: |
| 59 | * 1. Transform the inputs to the Winograd domain using `input_transform.run(...)` |
| 60 | * 2. Perform a number of GEMMs using `gemms.run(...)` |
| 61 | * 3. Transform the output to the spatial domain using `output_transform.run(...)` |
| 62 | */ |
| 63 | template <int OutputTileRows, int OutputTileCols, int KernelRows, int KernelCols, |
| 64 | typename TIn, typename TInGEMM, typename TOutGEMM, typename TOut, |
| 65 | WinogradRoots Roots> |
| 66 | class WinogradConvolutionLayer : public IWinogradConvolutionLayer |
| 67 | { |
| 68 | private: |
| 69 | static constexpr int InnerTileRows = OutputTileRows + KernelRows - 1; |
| 70 | static constexpr int InnerTileCols = OutputTileCols + KernelCols - 1; |
| 71 | static constexpr int N_GEMMS = InnerTileRows * InnerTileCols; |
| 72 | |
| 73 | const KernelShape _kernel_shape; |
| 74 | const Tensor4DShape _input_shape; |
| 75 | const PaddingType _padding; |
| 76 | const Tensor4DShape _output_shape; |
| 77 | const int _n_output_rows, _n_output_cols; |
| 78 | const int _kernel_matrix_stride, _kernel_matrix_row_stride; |
| 79 | const int _input_matrix_stride, _input_matrix_row_stride; |
| 80 | const int _output_matrix_stride, _output_matrix_row_stride; |
| 81 | const int _tile_rows, _tile_cols; |
| 82 | const int _m, _k, _n; |
| 83 | |
| 84 | public: |
| 85 | using WinogradBase = winograd::WinogradGEMM<OutputTileRows, OutputTileCols, KernelRows, KernelCols, Roots>; |
| 86 | using WeightsTransform = typename WinogradBase::template WeightsTransform<TIn, TInGEMM>; |
| 87 | using InputTransform = typename WinogradBase::template InputTransform<TIn, TInGEMM>; |
| 88 | using WinogradConv = typename WinogradBase::template Convolution<TOut, TIn, TInGEMM, TOutGEMM>; |
| 89 | using OutputTransform = typename WinogradBase::template OutputTransform<TOutGEMM, TOut>; |
| 90 | |
| 91 | /* Public member variables. */ |
| 92 | WeightsTransform weights_transform; /** Operator to transform weights to Winograd domain. */ |
| 93 | InputTransform _input_transform; /** Operator to transform input to Winograd domain. */ |
| 94 | arm_gemm::UniqueGemmCommon<TInGEMM, TOutGEMM> gemms; /** Operator to perform multiple GEMMs. */ |
| 95 | OutputTransform _output_transform; /** Operator to transform output from Winograd domain. */ |
| 96 | |
| 97 | /** Determine how much memory (in units of TIn) to allocate for the |
| 98 | * transformed weights. |
| 99 | */ |
| 100 | static unsigned int get_weight_storage_size( |
| 101 | const int n_output_channels, /** Number of output feature maps. */ |
| 102 | const int n_input_channels /** Number of input feature maps. */ |
| 103 | ); |
| 104 | |
| 105 | static unsigned int get_weight_stride( |
| 106 | const int n_output_channels, /** Number of output feature maps. */ |
| 107 | const int n_input_channels /** Number of input feature maps. */ |
| 108 | ); |
| 109 | |
| 110 | static unsigned int get_weight_multi_stride( |
| 111 | const int n_output_channels, /** Number of output feature maps. */ |
| 112 | const int n_input_channels /** Number of input feature maps. */ |
| 113 | ); |
| 114 | |
| 115 | /** Determine how much memory (in units of TIn) to allocate for the |
| 116 | * transformed input. |
| 117 | */ |
| 118 | static unsigned int get_input_storage_size( |
| 119 | const int n_batches, /** Number of batches in the input tensor. */ |
| 120 | const int n_channels, /** Number of feature maps in the input tensor. */ |
| 121 | const int n_rows, /** Number of rows in each feature map. */ |
| 122 | const int n_cols, /** Number of columns in each feature map. */ |
| 123 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 124 | ); |
| 125 | |
| 126 | /** Get the row stride for the A matrix in the Winograd domain. */ |
| 127 | static unsigned int get_input_stride( |
| 128 | const int n_batches, /** Number of batches in the input tensor. */ |
| 129 | const int n_channels, /** Number of feature maps in the input tensor. */ |
| 130 | const int n_rows, /** Number of rows in each feature map. */ |
| 131 | const int n_cols, /** Number of columns in each feature map. */ |
| 132 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 133 | ); |
| 134 | |
| 135 | /** Get the stride between A matrices in the Winograd domain. */ |
| 136 | static unsigned int get_input_multi_stride( |
| 137 | const int n_batches, /** Number of batches in the input tensor. */ |
| 138 | const int n_channels, /** Number of feature maps in the input tensor. */ |
| 139 | const int n_rows, /** Number of rows in each feature map. */ |
| 140 | const int n_cols, /** Number of columns in each feature map. */ |
| 141 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 142 | ); |
| 143 | |
| 144 | /** Determine how much memory (in units of TOut) to allocate for the |
| 145 | * (Winograd domain) output. |
| 146 | */ |
| 147 | static unsigned int get_output_storage_size( |
| 148 | const int n_batches, /** Number of batches in the output tensor. */ |
| 149 | const int n_rows, /** Number of rows in each feature map of the input tensor. */ |
| 150 | const int n_cols, /** Number of columns in each feature map of the input tensor. */ |
| 151 | const int n_output_channels, /** Number of feature maps in the output tensor. */ |
| 152 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 153 | ); |
| 154 | |
| 155 | static unsigned int get_output_stride( |
| 156 | const int n_batches, /** Number of batches in the output tensor. */ |
| 157 | const int n_rows, /** Number of rows in each feature map of the input tensor. */ |
| 158 | const int n_cols, /** Number of columns in each feature map of the input tensor. */ |
| 159 | const int n_output_channels, /** Number of feature maps in the output tensor. */ |
| 160 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 161 | ); |
| 162 | |
| 163 | static unsigned int get_output_multi_stride( |
| 164 | const int n_batches, /** Number of batches in the output tensor. */ |
| 165 | const int n_rows, /** Number of rows in each feature map of the input tensor. */ |
| 166 | const int n_cols, /** Number of columns in each feature map of the input tensor. */ |
| 167 | const int n_output_channels, /** Number of feature maps in the output tensor. */ |
| 168 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 169 | ); |
| 170 | |
| 171 | /** Get the shape (rows, cols) of a feature map of the output tensor. */ |
| 172 | static std::pair<int, int> get_output_feature_map_shape( |
| 173 | const int n_input_rows, /** Number of rows in the input feature map. */ |
| 174 | const int n_input_cols, /** Number of columns in the input feature map. */ |
| 175 | const bool same_padding /** Use "SAME" padding, otherwise use "VALID". */ |
| 176 | ); |
| 177 | |
| 178 | /** Create a new Winograd convolution layer. |
| 179 | */ |
| 180 | WinogradConvolutionLayer( |
| 181 | const arm_gemm::CPUInfo &cpuinfo, /** Describes CPU properties. */ |
| 182 | const int n_threads, /** Maximum number of threads used to execute the convolution. */ |
| 183 | const int n_batches, /** Number of batches in the input and output tensors. */ |
| 184 | const int n_input_channels, /** Number of feature maps in a batch of the input tensor. */ |
| 185 | const int n_input_rows, /** Number of rows in a feature map of the input tensor. */ |
| 186 | const int n_input_cols, /** Number of columns in a feature map of the input tensor. */ |
| 187 | const int n_output_channels, /** Number of feature maps in the output tensor. */ |
| 188 | const bool same_padding, /** Use "SAME" padding, otherwise use "VALID". */ |
| 189 | const TIn* const weights, /** Pointer to weight tensor in spatial domain. Must be ordered as "Height x Rows x Input Feature Maps x Output Feature Maps. */ |
| 190 | TInGEMM* const weights_storage, /** Pointer to storage for weight tensor in the Winograd domain. Must be at least the size returned by `get_weight_storage_size`. */ |
| 191 | const TIn* const input, /** Pointer to NHWC ordered input tensor, in the spatial domain. */ |
| 192 | TInGEMM* const winograd_input, /** Pointer to working space for the input tensor in the Winograd domain. Must be at least the size returned by `get_input_storage_size`. */ |
| 193 | const TOut* const biases, /** Pointer to biases vector. Pass nullptr if no bias is provided. */ |
| 194 | TOut* const output, /** Pointer to NHWC ordered output tensor, in the spatial domain. */ |
| 195 | TOutGEMM* const winograd_output, /** Pointer to working space for the output tensor in the Winograd domain. Must be at least the size returned by `get_output_storage_size`. */ |
| 196 | const bool pretranspose_B=true, /** Hint that the B matrix can be pretransposed. */ |
| 197 | arm_gemm::GemmConfig *gemm_cfg=nullptr /** Pointer to GEMM configuration. */ |
| 198 | ); |
| 199 | |
| 200 | /* Utility methods for interacting with the layer. */ |
| 201 | unsigned int weight_transform_get_window(void) const; |
| 202 | void weight_transform_run(const unsigned int start, const unsigned int stop); |
| 203 | |
| 204 | ITransform& input_transform(void); |
| 205 | ITransform& output_transform(void); |
| 206 | |
| 207 | /* Get a pointer to the GEMM underlying the Winograd transform. */ |
| 208 | arm_gemm::IGemmCommon *gemm(void); |
| 209 | }; |
| 210 | |
| 211 | } |