1 /* 2 * Copyright (c) 2017-2020 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 #ifndef ARM_COMPUTE_NEPOOLINGLAYERKERNEL_H 25 #define ARM_COMPUTE_NEPOOLINGLAYERKERNEL_H 26 27 #include "src/core/NEON/INEKernel.h" 28 29 namespace arm_compute 30 { 31 class ITensor; 32 33 /** Interface for the pooling layer kernel */ 34 class NEPoolingLayerKernel : public INEKernel 35 { 36 public: name()37 const char *name() const override 38 { 39 return "NEPoolingLayerKernel"; 40 } 41 /** Default constructor */ 42 NEPoolingLayerKernel(); 43 /** Prevent instances of this class from being copied (As this class contains pointers) */ 44 NEPoolingLayerKernel(const NEPoolingLayerKernel &) = delete; 45 /** Prevent instances of this class from being copied (As this class contains pointers) */ 46 NEPoolingLayerKernel &operator=(const NEPoolingLayerKernel &) = delete; 47 /** Allow instances of this class to be moved */ 48 NEPoolingLayerKernel(NEPoolingLayerKernel &&) = default; 49 /** Allow instances of this class to be moved */ 50 NEPoolingLayerKernel &operator=(NEPoolingLayerKernel &&) = default; 51 /** Default destructor */ 52 ~NEPoolingLayerKernel() = default; 53 /** Set the input and output tensors. 54 * 55 * @note F16 are supported for pool sizes 2 and 3 only 56 * 57 * @param[in] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. 58 * @param[out] output Destination tensor. Data types supported: Same as @p input. 59 * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. 60 * @param[out] indices (optional) The indices of the maximal values. Data type supported: U32. 61 */ 62 void configure(const ITensor *input, ITensor *output, const PoolingLayerInfo &pool_info, ITensor *indices = nullptr); 63 /** Static function to check if given info will lead to a valid configuration of @ref NEPoolingLayerKernel 64 * 65 * @note F16 are supported for pool sizes 2 and 3 only 66 * 67 * @param[in] input Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. 68 * @param[in] output Destination tensor. Data types supported: Same as @p input. 69 * @param[in] pool_info Contains pooling operation information described in @ref PoolingLayerInfo. 70 * @param[in] indices (optional) The indices of the maximal values. Data type supported: U32. 71 * 72 * @return a status 73 */ 74 static Status validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info, const ITensorInfo *indices = nullptr); 75 76 // Inherited methods overridden: 77 void run(const Window &window, const ThreadInfo &info) override; 78 BorderSize border_size() const override; 79 80 private: 81 /** Function to perform 2x2 pooling. 82 * 83 * @param[in] window_input Input region on which to execute the kernel. 84 * @param[in] window Output region on which to execute the kernel. 85 * @param[in] pooling_type Pooling operation to be computed. 86 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 87 */ 88 void pooling2_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 89 /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool. 90 * 91 * @param[in] window_input Input region on which to execute the kernel. 92 * @param[in] window Output region on which to execute the kernel. 93 */ 94 void pooling2_f32_nhwc_maxpool_indices(const Window &window_input, const Window &window); 95 /** Function to perform MxN pooling for 32-bit floating point values. 96 * 97 * @param[in] window_input Input region on which to execute the kernel. 98 * @param[in] window Output region on which to execute the kernel. 99 * @param[in] pooling_type Pooling operation to be computed. 100 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 101 */ 102 void poolingMxN_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 103 /** Function to perform MxN pooling for 32-bit floating point values (NHWC). 104 * 105 * @param[in] window_input Input region on which to execute the kernel. 106 * @param[in] window Output region on which to execute the kernel. 107 * @param[in] pooling_type Pooling operation to be computed. 108 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 109 */ 110 void poolingMxN_f32_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 111 /** Function to perform 7x7 pooling. 112 * 113 * @param[in] window_input Input region on which to execute the kernel. 114 * @param[in] window Output region on which to execute the kernel. 115 * @param[in] pooling_type Pooling operation to be computed. 116 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 117 */ 118 void pooling7_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 119 /** Function to perform 3x3 pooling. 120 * 121 * @param[in] window_input Input region on which to execute the kernel. 122 * @param[in] window Output region on which to execute the kernel. 123 * @param[in] pooling_type Pooling operation to be computed. 124 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 125 */ 126 void pooling3_f32_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 127 /** Function to perform 2x2 pooling for float16_t. 128 * 129 * @param[in] window_input Input region on which to execute the kernel. 130 * @param[in] window Output region on which to execute the kernel. 131 * @param[in] pooling_type Pooling operation to be computed. 132 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 133 */ 134 void pooling2_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 135 /** Function to perform 2x2 pooling and compute the pooling indices for FP32/FP16. The indices can be used for max unpool. 136 * 137 * @param[in] window_input Input region on which to execute the kernel. 138 * @param[in] window Output region on which to execute the kernel. 139 */ 140 template <typename T> 141 void pooling2_nchw_maxpool_indices(const Window &window_input, const Window &window); 142 /** Function to perform 2x2 pooling and compute the pooling indices. The indices can be used for max unpool. 143 * 144 * @param[in] window_input Input region on which to execute the kernel. 145 * @param[in] window Output region on which to execute the kernel. 146 */ 147 void pooling2_f16_nhwc_maxpool_indices(const Window &window_input, const Window &window); 148 /** Function to perform 3x3 pooling. 149 * 150 * @param[in] window_input Input region on which to execute the kernel. 151 * @param[in] window Output region on which to execute the kernel. 152 * @param[in] pooling_type Pooling operation to be computed. 153 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 154 */ 155 void pooling3_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 156 /** Function to perform MxN pooling for 16-bit floating point values. 157 * 158 * @param[in] window_input Input region on which to execute the kernel. 159 * @param[in] window Output region on which to execute the kernel. 160 * @param[in] pooling_type Pooling operation to be computed. 161 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 162 */ 163 void poolingMxN_f16_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 164 /** Function to perform MxN pooling for 16-bit floating point values. (NHWC) 165 * 166 * @param[in] window_input Input region on which to execute the kernel. 167 * @param[in] window Output region on which to execute the kernel. 168 * @param[in] pooling_type Pooling operation to be computed. 169 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 170 */ 171 void poolingMxN_f16_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 172 /** Template function to perform 2x2 pooling for 8bit quantized fixed point. (NCHW) 173 * 174 * @param[in] window_input Input region on which to execute the kernel. 175 * @param[in] window Output region on which to execute the kernel. 176 * @param[in] pooling_type Pooling operation to be computed. 177 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 178 */ 179 template <typename T> 180 void pooling2_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 181 /** Template function to perform 3x3 pooling for 8bit quantized fixed point. (NCHW) 182 * 183 * @param[in] window_input Input region on which to execute the kernel. 184 * @param[in] window Output region on which to execute the kernel. 185 * @param[in] pooling_type Pooling operation to be computed. 186 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 187 */ 188 template <typename T> 189 void pooling3_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 190 /** Template function to perform MxN pooling for 8-bit quantized. (NCHW) 191 * 192 * @param[in] window_input Input region on which to execute the kernel. 193 * @param[in] window Output region on which to execute the kernel. 194 * @param[in] pooling_type Pooling operation to be computed. 195 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 196 */ 197 template <typename T> 198 void poolingMxN_q8_nchw(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 199 /** Template function to perform MxN pooling for 8-bit quantized. (NHWC) 200 * 201 * @param[in] window_input Input region on which to execute the kernel. 202 * @param[in] window Output region on which to execute the kernel. 203 * @param[in] pooling_type Pooling operation to be computed. 204 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 205 */ 206 template <typename T> 207 void poolingMxN_q8_nhwc(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding = false); 208 /** Common signature for all the specialised Pooling functions 209 * 210 * @param[in] window_input Input region on which to execute the kernel. 211 * @param[in] window Output region on which to execute the kernel. 212 * @param[in] pooling_type Pooling operation to be computed. 213 * @param[in] exclude_padding Flag to specify exclusion of padding from the operation. 214 */ 215 using PoolingFunction = void (NEPoolingLayerKernel::*)(const Window &window_input, const Window &window, PoolingType pooling_type, bool exclude_padding); 216 217 private: 218 PoolingFunction _func; 219 const ITensor *_input; 220 ITensor *_output; 221 ITensor *_indices; 222 PoolingLayerInfo _pool_info; 223 DataLayout _data_layout; 224 unsigned int _num_elems_processed_per_iteration; 225 BorderSize _border_size; 226 bool _is_square; 227 }; 228 } // namespace arm_compute 229 #endif /*ARM_COMPUTE_NEPOOLINGLAYERKERNEL_H */ 230