Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 1 | /* |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 2 | * Copyright (c) 2021-2022 Arm Limited. |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 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 | |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "depthfirst_driver.hpp" |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 29 | #include "utils.hpp" |
Pablo Marquez Tello | 84a0941 | 2022-06-30 17:00:09 +0100 | [diff] [blame] | 30 | #if !defined(_WIN64) && !defined(__OpenBSD__) |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 31 | #include <alloca.h> |
Pablo Marquez Tello | 84a0941 | 2022-06-30 17:00:09 +0100 | [diff] [blame] | 32 | #endif /* !defined(_WIN64) && !defined(__OpenBSD__) */ |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 33 | |
| 34 | namespace arm_conv { |
| 35 | namespace pooling { |
| 36 | |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 37 | template <typename TInput, typename TOutput, typename OutputStage = Nothing> |
| 38 | class IGenericDepthfirstStrategy; |
| 39 | |
| 40 | template <typename TInput, typename TOutput> |
| 41 | class IGenericDepthfirstStrategy<TInput, TOutput, Nothing> |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 42 | { |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 43 | public: |
| 44 | virtual ~IGenericDepthfirstStrategy() = default; |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 45 | |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 46 | typedef void (*KernelType)( |
| 47 | uint64_t window_cells, |
| 48 | uint64_t n_valid_cells, |
| 49 | uint64_t n_channels, |
| 50 | const TInput *const *, |
| 51 | TOutput * |
| 52 | ); |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 53 | |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 54 | virtual KernelType get_kernel(void) const = 0; |
| 55 | }; |
| 56 | |
| 57 | template <typename TInput, typename TOutput> |
| 58 | class IGenericDepthfirstStrategy<TInput, TOutput, Requantize32> |
| 59 | { |
| 60 | public: |
| 61 | virtual ~IGenericDepthfirstStrategy() = default; |
| 62 | |
| 63 | typedef void (*KernelType)( |
| 64 | uint64_t window_cells, |
| 65 | uint64_t n_valid_cells, |
| 66 | uint64_t n_channels, |
| 67 | const TInput *const *, |
| 68 | TOutput *, |
| 69 | const Requantize32 & |
| 70 | ); |
| 71 | |
| 72 | virtual KernelType get_kernel(void) const = 0; |
| 73 | }; |
| 74 | |
| 75 | template <typename TInput, typename TOutput, typename OutputStage> |
| 76 | struct Invoker; |
| 77 | |
| 78 | template <typename TInput, typename TOutput> |
| 79 | struct Invoker<TInput, TOutput, Nothing> |
| 80 | { |
| 81 | static inline void invoke( |
| 82 | const typename IGenericDepthfirstStrategy<TInput, TOutput, Nothing>::KernelType kern, |
| 83 | uint64_t window_cells, |
| 84 | uint64_t n_valid_cells, |
| 85 | uint64_t n_channels, |
| 86 | const TInput *const *inptrs, |
| 87 | TOutput *outptr, |
| 88 | const Nothing & |
| 89 | ) |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 90 | { |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 91 | kern(window_cells, n_valid_cells, n_channels, inptrs, outptr); |
| 92 | } |
| 93 | }; |
| 94 | |
| 95 | template <typename TInput, typename TOutput> |
| 96 | struct Invoker<TInput, TOutput, Requantize32> |
| 97 | { |
| 98 | static inline void invoke( |
| 99 | const typename IGenericDepthfirstStrategy<TInput, TOutput, Requantize32>::KernelType kern, |
| 100 | uint64_t window_cells, |
| 101 | uint64_t n_valid_cells, |
| 102 | uint64_t n_channels, |
| 103 | const TInput *const *inptrs, |
| 104 | TOutput *outptr, |
| 105 | const Requantize32 &qp |
| 106 | ) |
| 107 | { |
| 108 | kern(window_cells, n_valid_cells, n_channels, inptrs, outptr, qp); |
| 109 | } |
| 110 | }; |
| 111 | |
| 112 | template <typename TInput, typename TOutput, typename OutputStage> |
| 113 | class GenericDepthfirstWrapper : public IDepthfirstStrategy |
| 114 | { |
| 115 | using StratType = IGenericDepthfirstStrategy<TInput, TOutput, OutputStage>; |
| 116 | |
| 117 | std::unique_ptr<const StratType> m_strat; |
| 118 | const unsigned int window_rows, window_cols; |
| 119 | |
| 120 | public: |
| 121 | GenericDepthfirstWrapper(const StratType *strat, const PoolingArgs &args) |
| 122 | : m_strat(strat), window_rows(args.pool_window.rows), window_cols(args.pool_window.cols) |
| 123 | { |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 124 | } |
| 125 | |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 126 | unsigned int get_input_rows(void) const override { return window_rows; } |
| 127 | unsigned int get_input_cols(void) const override { return window_cols; } |
| 128 | unsigned int get_output_rows(void) const override { return 1; } |
| 129 | unsigned int get_output_cols(void) const override { return 1; } |
| 130 | |
| 131 | typename StratType::KernelType get_kernel(void) const { return m_strat->get_kernel(); } |
| 132 | }; |
| 133 | |
| 134 | template <typename TInput, typename TOutput=TInput, typename OutputStage=Nothing> |
| 135 | class PoolingDepthfirstGeneric : public DepthfirstDriver<TInput, TOutput> |
| 136 | { |
| 137 | const OutputStage m_os; |
| 138 | |
| 139 | protected: |
| 140 | size_t get_working_size_per_thread(unsigned int) const override { return 0; } |
| 141 | void initialise_working_space(void *, unsigned int) const override { /* Nothing */ } |
| 142 | |
| 143 | /* Compute a portion of the output tensor with padding. */ |
| 144 | void compute_tile_padded( |
| 145 | unsigned int output_i, unsigned int output_j, |
| 146 | unsigned int channel_start, unsigned int channel_end, |
| 147 | const TensorSpec<const TInput *> &input, |
| 148 | const TensorSpec<TOutput *> &output, |
| 149 | void * |
| 150 | ) const override |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 151 | { |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 152 | // Determine start position and padding |
| 153 | const int start_i = static_cast<int>(output_i * this->m_args.pool_stride.rows) - this->m_args.padding.top; |
| 154 | const auto input_i = static_cast<unsigned int>(start_i < 0 ? 0 : start_i); |
| 155 | const auto pad_top = static_cast<unsigned int>(start_i < 0 ? -start_i : 0); |
| 156 | const int end_i = start_i + this->m_args.pool_window.rows; |
| 157 | const auto pad_bottom = static_cast<unsigned int>((unsigned int) end_i < this->m_args.input_rows ? 0 : end_i - this->m_args.input_rows); |
| 158 | const auto valid_rows = this->m_args.pool_window.rows - (pad_top + pad_bottom); |
| 159 | |
| 160 | const int start_j = static_cast<int>(output_j * this->m_args.pool_stride.cols) - this->m_args.padding.left; |
| 161 | const auto input_j = static_cast<unsigned int>(start_j < 0 ? 0 : start_j); |
| 162 | const auto pad_left = static_cast<unsigned int>(start_j < 0 ? -start_j : 0); |
| 163 | const int end_j = start_j + this->m_args.pool_window.cols; |
| 164 | const auto pad_right = static_cast<unsigned int>((unsigned int) end_j < this->m_args.input_cols ? 0 : end_j - this->m_args.input_cols); |
| 165 | const auto valid_cols = this->m_args.pool_window.cols - (pad_left + pad_right); |
| 166 | |
| 167 | // Determine the number of valid cells and prepare the pointers |
| 168 | const auto n_valid_cells = valid_rows * valid_cols; |
| 169 | auto inptrs = reinterpret_cast<const TInput **>(alloca(n_valid_cells * sizeof(TInput *))); |
| 170 | { |
| 171 | auto my_ptr = inptrs; |
| 172 | auto row_ptr = input.base + input_i*input.ld_row + input_j*input.ld_col + channel_start; |
| 173 | for (auto i = valid_rows; i; i--) |
| 174 | { |
| 175 | auto ptr = row_ptr; |
| 176 | row_ptr += input.ld_row; |
| 177 | |
| 178 | for (auto j = valid_cols; j; j--) |
| 179 | { |
| 180 | *(my_ptr++) = ptr; |
| 181 | ptr += input.ld_col; |
| 182 | } |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | auto outptr = output.base + output_i*output.ld_row + output_j*output.ld_col + channel_start; |
| 187 | |
| 188 | // Some padding variants include (or exclude) the padding values; we handle |
| 189 | // this by computing the extent of the padded input tensor and hence |
| 190 | // computing the total number of cells captured in the pooling window. |
| 191 | const auto bottom_padded_height = this->m_args.input_rows + this->m_args.padding.bottom; |
| 192 | const auto captured_rows = std::min<int>(end_i, bottom_padded_height) - start_i; |
| 193 | const auto right_padded_width = this->m_args.input_cols + this->m_args.padding.right; |
| 194 | const auto captured_cols = std::min<int>(end_j, right_padded_width) - start_j; |
| 195 | const auto captured_cells = captured_rows * captured_cols; |
| 196 | const auto window_cells = this->m_args.exclude_padding ? n_valid_cells : captured_cells; |
| 197 | |
| 198 | // Execute the kernel |
| 199 | Invoker<TInput, TOutput, OutputStage>::invoke( |
| 200 | reinterpret_cast<const GenericDepthfirstWrapper<TInput, TOutput, OutputStage> *>(this->m_strat.get())->get_kernel(), |
| 201 | window_cells, n_valid_cells, channel_end - channel_start, inptrs, outptr, m_os |
| 202 | ); |
| 203 | } |
| 204 | |
| 205 | // Compute a portion of the work with only top/bottom padding. |
| 206 | void compute_row_padded_tile_row( |
| 207 | const unsigned int output_i, unsigned int output_j, unsigned int n_tile_cols, |
| 208 | const unsigned int channel_start, const unsigned int channel_end, |
| 209 | const TensorSpec<const TInput *> &input, |
| 210 | const TensorSpec<TOutput *> &output, |
| 211 | void *working_space |
| 212 | ) const override |
| 213 | { |
| 214 | ARM_COMPUTE_UNUSED(working_space); |
| 215 | // Determine start position and padding |
| 216 | const int start_i = static_cast<int>(output_i * this->m_args.pool_stride.rows) - this->m_args.padding.top; |
| 217 | const auto input_i = static_cast<unsigned int>(start_i < 0 ? 0 : start_i); |
| 218 | const auto pad_top = static_cast<unsigned int>(start_i < 0 ? -start_i : 0); |
| 219 | const int end_i = start_i + this->m_args.pool_window.rows; |
| 220 | const auto pad_bottom = static_cast<unsigned int>((unsigned int) end_i < this->m_args.input_rows ? 0 : end_i - this->m_args.input_rows); |
| 221 | const auto valid_rows = this->m_args.pool_window.rows - (pad_top + pad_bottom); |
| 222 | |
| 223 | const int start_j = static_cast<int>(output_j * this->m_args.pool_stride.cols) - this->m_args.padding.left; |
| 224 | const auto input_j = static_cast<unsigned int>(start_j < 0 ? 0 : start_j); |
| 225 | const auto valid_cols = this->m_args.pool_window.cols; |
| 226 | |
| 227 | // Determine the number of valid cells and prepare the pointers |
| 228 | const auto n_valid_cells = valid_rows * valid_cols; |
| 229 | auto inptrs = reinterpret_cast<const TInput **>(alloca(n_valid_cells * sizeof(TInput *))); |
| 230 | { |
| 231 | auto my_ptr = inptrs; |
| 232 | auto row_ptr = input.base + input_i*input.ld_row + input_j*input.ld_col + channel_start; |
| 233 | for (auto i = valid_rows; i; i--) |
| 234 | { |
| 235 | auto ptr = row_ptr; |
| 236 | row_ptr += input.ld_row; |
| 237 | |
| 238 | for (auto j = valid_cols; j; j--) |
| 239 | { |
| 240 | *(my_ptr++) = ptr; |
| 241 | ptr += input.ld_col; |
| 242 | } |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | auto outptr = output.base + output_i*output.ld_row + output_j*output.ld_col + channel_start; |
| 247 | |
| 248 | // Some padding variants include (or exclude) the padding values; we handle |
| 249 | // this by computing the extent of the padded input tensor and hence |
| 250 | // computing the total number of cells captured in the pooling window. |
| 251 | const auto bottom_padded_height = this->m_args.input_rows + this->m_args.padding.bottom; |
| 252 | const auto captured_rows = std::min<int>(end_i, bottom_padded_height) - start_i; |
| 253 | const auto captured_cells = captured_rows * valid_cols; |
| 254 | const auto window_cells = this->m_args.exclude_padding ? n_valid_cells : captured_cells; |
| 255 | |
| 256 | for (; n_tile_cols; n_tile_cols--) |
| 257 | { |
| 258 | // Execute the kernel |
| 259 | Invoker<TInput, TOutput, OutputStage>::invoke( |
| 260 | reinterpret_cast<const GenericDepthfirstWrapper<TInput, TOutput, OutputStage> *>(this->m_strat.get())->get_kernel(), |
| 261 | window_cells, n_valid_cells, channel_end - channel_start, inptrs, outptr, m_os |
| 262 | ); |
| 263 | |
| 264 | // Update the pointers; the output strides by a column and the inputs |
| 265 | // stride by a number of columns. |
| 266 | outptr += output.ld_col; |
| 267 | for (auto n = 0u; n < n_valid_cells; n++) |
| 268 | { |
| 269 | inptrs[n] += this->m_args.pool_stride.cols * input.ld_col; |
| 270 | } |
| 271 | } |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 272 | } |
| 273 | |
| 274 | public: |
ramelg01 | c827e99 | 2022-04-08 03:52:28 +0100 | [diff] [blame] | 275 | PoolingDepthfirstGeneric( |
| 276 | const IGenericDepthfirstStrategy<TInput, TOutput, OutputStage> *strat, |
| 277 | const PoolingArgs &args, |
| 278 | const OutputStage &os = {} |
| 279 | ) |
| 280 | : DepthfirstDriver<TInput, TOutput>( |
| 281 | new GenericDepthfirstWrapper<TInput, TOutput, OutputStage>(strat, args), |
| 282 | args |
| 283 | ), |
| 284 | m_os(os) |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 285 | { |
| 286 | } |
Michele Di Giorgio | d556d7b | 2020-10-27 10:56:31 +0000 | [diff] [blame] | 287 | }; |
| 288 | |
| 289 | } // namespace pooling |
| 290 | } // namespace arm_conv |