ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 1 | /* |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 2 | * Copyright (c) 2022-2023 Arm Limited. |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [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 | #include "depthfirst_driver.hpp" |
| 26 | #include "interleaves/generic.hpp" |
| 27 | |
| 28 | namespace arm_conv { |
| 29 | namespace depthwise { |
| 30 | |
| 31 | template <typename OutputStage> |
| 32 | class IPlanarStrategy |
| 33 | { |
| 34 | public: |
| 35 | virtual ~IPlanarStrategy() = default; |
| 36 | virtual unsigned int get_output_rows(void) const = 0; |
| 37 | virtual arm_gemm::VLType get_vl_type(void) const = 0; |
| 38 | |
| 39 | virtual size_t get_storage_size(const DepthwiseArgs &) const = 0; |
| 40 | virtual void pack_parameters( |
| 41 | const DepthwiseArgs &args, void *buffer, |
| 42 | const void *biases, const OutputStage &, |
| 43 | const void *weights, size_t ld_weight_col, size_t ld_weight_row |
| 44 | ) const = 0; |
| 45 | }; |
| 46 | |
| 47 | |
| 48 | template <typename TInput, typename TWeight, typename TOutput, typename TAccum, |
| 49 | typename OutputStage> |
| 50 | struct PlanarKernelType; |
| 51 | |
| 52 | template <typename TInput, typename TWeight, typename TOutput, typename TAccum> |
| 53 | struct PlanarKernelType<TInput, TWeight, TOutput, TAccum, Nothing> |
| 54 | { |
| 55 | using Type = std::function<void( |
| 56 | const TInput *, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 57 | unsigned int pad_top, unsigned int valid_input_rows, |
| 58 | unsigned int pad_left, unsigned int valid_input_cols, |
| 59 | const TWeight *, const TAccum *, |
| 60 | TOutput **, const size_t *, const size_t *, unsigned int output_cols, |
| 61 | unsigned int start_channels, unsigned int valid_channels, |
| 62 | TAccum act_min, TAccum act_max |
| 63 | )>; |
| 64 | |
| 65 | template <typename WorkspaceType> |
| 66 | static inline void execute( |
| 67 | const Type fn, |
| 68 | const TInput *inptr, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 69 | unsigned int pad_top, unsigned int valid_input_rows, |
| 70 | unsigned int pad_left, unsigned int valid_input_cols, |
| 71 | const TWeight *weights, const TAccum *bias, |
| 72 | TOutput **outptrs, const size_t *outlds, const size_t *outvllds, unsigned int output_cols, |
| 73 | unsigned int start_channel, unsigned int valid_channels, |
| 74 | const Nothing &, const WorkspaceType *ws |
| 75 | ) |
| 76 | { |
| 77 | fn( |
| 78 | inptr, ld_in_row, ld_in_col, ld_in_vl, |
| 79 | pad_top, valid_input_rows, |
| 80 | pad_left, valid_input_cols, |
| 81 | weights, bias, |
| 82 | outptrs, outlds, outvllds, output_cols, |
| 83 | start_channel, valid_channels, |
| 84 | ws->activation_min, ws->activation_max |
| 85 | ); |
| 86 | } |
| 87 | }; |
| 88 | |
| 89 | template <typename TInput, typename TWeight, typename TOutput> |
| 90 | struct PlanarKernelType<TInput, TWeight, TOutput, int32_t, arm_gemm::Requantize32> |
| 91 | { |
| 92 | using Type = std::function<void( |
| 93 | const TInput *, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 94 | unsigned int pad_top, unsigned int valid_input_rows, |
| 95 | unsigned int pad_left, unsigned int valid_input_cols, |
| 96 | const TWeight *, |
| 97 | TOutput **, const size_t *, const size_t *, unsigned int output_cols, |
| 98 | unsigned int start_channel, unsigned int valid_channels, |
| 99 | const arm_gemm::Requantize32 & |
| 100 | )>; |
| 101 | |
| 102 | template <typename WorkspaceType> |
| 103 | static inline void execute( |
| 104 | const Type fn, |
| 105 | const TInput *inptr, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 106 | unsigned int pad_top, unsigned int valid_input_rows, |
| 107 | unsigned int pad_left, unsigned int valid_input_cols, |
| 108 | const TWeight *weights, const int32_t *, |
| 109 | TOutput **outptrs, const size_t *outlds, const size_t *outldvls, unsigned int output_cols, |
| 110 | unsigned int first_channel, unsigned int valid_channels, |
| 111 | const arm_gemm::Requantize32 &qp, const WorkspaceType * |
| 112 | ) |
| 113 | { |
| 114 | fn( |
| 115 | inptr, ld_in_row, ld_in_col, ld_in_vl, |
| 116 | pad_top, valid_input_rows, |
| 117 | pad_left, valid_input_cols, |
| 118 | weights, |
| 119 | outptrs, outlds, outldvls, output_cols, |
| 120 | first_channel, valid_channels, |
| 121 | qp |
| 122 | ); |
| 123 | } |
| 124 | }; |
| 125 | |
| 126 | |
| 127 | template <typename TInput, typename TWeight=TInput, typename TOutput=TInput, |
| 128 | typename TAccum=typename DefaultTAccum<TOutput>::Type, |
| 129 | typename OutputStage=typename DefaultOutputStage<TOutput>::Type> |
| 130 | class PlanarStrategy : public IPlanarStrategy<OutputStage> |
| 131 | { |
| 132 | unsigned int m_kernel_rows, m_kernel_cols; |
| 133 | unsigned int m_stride_rows, m_stride_cols; |
| 134 | unsigned int m_output_rows; |
| 135 | arm_gemm::VLType m_vl_type; |
| 136 | |
| 137 | protected: |
| 138 | virtual bool get_kernel_packing_point(const unsigned int index, unsigned int &x, unsigned int &y) const |
| 139 | { |
| 140 | // Get the kernel point to pack at the given index; return false to |
| 141 | // indicate that this index (and all greater indices) is out of range. |
| 142 | if (m_kernel_rows * m_kernel_cols <= index) |
| 143 | return false; |
| 144 | |
| 145 | y = index % m_kernel_cols; |
| 146 | x = index / m_kernel_cols; |
| 147 | return true; |
| 148 | } |
| 149 | |
| 150 | virtual interleaves::PackingArguments get_kernel_packing_arguments(void) const |
| 151 | { |
| 152 | return interleaves::PackingArguments( |
| 153 | m_kernel_rows, m_kernel_cols, sizeof(TWeight), |
| 154 | false, sizeof(TAccum), // Don't pack the bias |
| 155 | m_vl_type, sizeof(TAccum), 1, // Accumulator depth of 1 TODO |
| 156 | [this] (unsigned int idx, unsigned int &x, unsigned int &y) -> bool |
| 157 | { return this->get_kernel_packing_point(idx, x, y); } |
| 158 | ); |
| 159 | } |
| 160 | |
| 161 | public: |
| 162 | PlanarStrategy( |
| 163 | unsigned int kernel_rows, unsigned int kernel_cols, |
| 164 | unsigned int stride_rows, unsigned int stride_cols, |
| 165 | unsigned int output_rows, |
| 166 | arm_gemm::VLType vl_type |
| 167 | ) : m_kernel_rows(kernel_rows), m_kernel_cols(kernel_cols), |
| 168 | m_stride_rows(stride_rows), m_stride_cols(stride_cols), |
| 169 | m_output_rows(output_rows), m_vl_type(vl_type) |
| 170 | { |
| 171 | } |
| 172 | |
| 173 | unsigned int get_output_rows(void) const override { return m_output_rows; } |
| 174 | arm_gemm::VLType get_vl_type(void) const override { return m_vl_type; } |
| 175 | |
| 176 | size_t get_storage_size(const DepthwiseArgs &args) const override |
| 177 | { |
| 178 | return interleaves::get_storage_size_generic(this->get_kernel_packing_arguments(), args); |
| 179 | } |
| 180 | |
| 181 | void pack_parameters( |
| 182 | const DepthwiseArgs &args, void *buffer, |
| 183 | const void *biases, const OutputStage &, |
| 184 | const void *weights, size_t ld_weight_col, size_t ld_weight_row |
| 185 | ) const override |
| 186 | { |
| 187 | interleaves::pack_parameters_generic( |
| 188 | this->get_kernel_packing_arguments(), args, |
| 189 | buffer, biases, weights, ld_weight_col, ld_weight_row |
| 190 | ); |
| 191 | } |
| 192 | |
| 193 | using KernelType = typename PlanarKernelType<TInput, TWeight, TOutput, TAccum, OutputStage>::Type; |
| 194 | virtual KernelType get_kernel(void) const = 0; |
| 195 | }; |
| 196 | |
| 197 | |
| 198 | namespace { |
| 199 | |
| 200 | template <typename T> |
| 201 | struct OutputRowPtrsElement |
| 202 | { |
| 203 | struct Workspace |
| 204 | { |
| 205 | T **output_row_ptrs; |
| 206 | size_t *output_ld_cols; |
| 207 | size_t *output_ld_vls; // Stride between vectors of channels |
| 208 | T *output_padding_buffer; |
| 209 | }; |
| 210 | |
| 211 | template <typename OutputStage> |
| 212 | static size_t get_element_size(const WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage> &args) |
| 213 | { |
| 214 | // We need one pointer and stride for each row of output, and an additional |
| 215 | // blob of memory into which padded stores can go. |
| 216 | return args.strategy->get_output_rows() * (sizeof(T *) + 2*sizeof(size_t)) + |
| 217 | get_vector_length<char>(args.strategy->get_vl_type()); |
| 218 | } |
| 219 | |
| 220 | template <typename WorkspaceType, typename OutputStage> |
| 221 | static void *initialise(WorkspaceType *ws, void *buffer, |
| 222 | const WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage> &args) |
| 223 | { |
| 224 | const auto n_rows = args.strategy->get_output_rows(); |
| 225 | ws->output_row_ptrs = reinterpret_cast<T **>(buffer); |
| 226 | ws->output_ld_cols = reinterpret_cast<size_t *>(ws->output_row_ptrs + n_rows); |
| 227 | ws->output_ld_vls = ws->output_ld_cols + n_rows; |
| 228 | ws->output_padding_buffer = reinterpret_cast<T *>(ws->output_ld_vls + n_rows); |
| 229 | return ws->output_padding_buffer + get_vector_length<T>(args.strategy->get_vl_type()); |
| 230 | } |
| 231 | }; |
| 232 | |
| 233 | } // namespace {anonymous} |
| 234 | |
| 235 | |
| 236 | template <typename TInput, typename TWeight=TInput, typename TOutput=TInput, |
| 237 | typename TAccum=typename DefaultTAccum<TOutput>::Type, |
| 238 | typename OutputStage=typename DefaultOutputStage<TOutput>::Type> |
| 239 | class DepthwisePlanar : public DepthwiseCommon<TInput, TWeight, TOutput> |
| 240 | { |
| 241 | using Parent = DepthwiseCommon<TInput, TWeight, TOutput>; |
| 242 | using StrategyType = IPlanarStrategy<OutputStage>; |
| 243 | using WorkspaceManager = Workspace< |
| 244 | OutputRowPtrsElement<TOutput>, |
| 245 | ActivationsElement<TAccum, OutputStage> |
| 246 | >; |
| 247 | using WorkspaceType = typename WorkspaceManager::WorkspaceType; |
| 248 | |
| 249 | std::unique_ptr<StrategyType> m_strat; |
| 250 | const TAccum *m_bias; |
| 251 | OutputStage m_os; |
| 252 | |
| 253 | public: |
| 254 | DepthwisePlanar(StrategyType *const strat, const DepthwiseArgs &args, const OutputStage &os = {}) |
| 255 | : Parent(args), m_strat(strat), m_bias(nullptr), m_os(os) |
| 256 | { |
| 257 | } |
| 258 | |
Viet-Hoa Do | 03b2971 | 2022-06-01 11:47:14 +0100 | [diff] [blame] | 259 | DepthwisePlanar(DepthwisePlanar &) = delete; |
| 260 | DepthwisePlanar &operator=(DepthwisePlanar &) = delete; |
| 261 | |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 262 | size_t get_storage_size(void) const override |
| 263 | { |
| 264 | return m_strat->get_storage_size(this->m_args); |
| 265 | } |
| 266 | |
| 267 | void pack_parameters( |
| 268 | void *buffer, const void *biases, |
| 269 | const void *weights, size_t ld_weight_col, size_t ld_weight_row |
| 270 | ) override |
| 271 | { |
| 272 | m_strat->pack_parameters(this->m_args, buffer, biases, {}, weights, ld_weight_col, ld_weight_row); |
| 273 | this->m_bias = reinterpret_cast<const TAccum *>(biases); |
| 274 | depthwise_depthfirst::stash_bias(this->m_os, biases); |
| 275 | } |
| 276 | |
| 277 | size_t get_working_size(unsigned int n_threads, unsigned int) const override |
| 278 | { |
| 279 | return this->get_working_size_per_thread() * n_threads; |
| 280 | } |
| 281 | |
| 282 | protected: |
| 283 | /* Compute the amount of working space required for a single thread. */ |
| 284 | virtual size_t get_working_size_per_thread(void) const |
| 285 | { |
| 286 | return WorkspaceManager::get_sizeof_workspace( |
| 287 | WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage>(m_strat.get(), this->m_args, m_os)); |
| 288 | } |
| 289 | |
| 290 | /* Initialise the working space for a thread. */ |
| 291 | virtual void initialise_working_space(void *buffer) const |
| 292 | { |
| 293 | WorkspaceManager::initialise( |
| 294 | buffer, |
| 295 | WorkspaceArgs<IPlanarStrategy<OutputStage>, OutputStage>(m_strat.get(), this->m_args, m_os) |
| 296 | ); |
| 297 | } |
| 298 | |
| 299 | /* Execute the kernel for a given chunk of work. */ |
| 300 | virtual void execute_kernel( |
| 301 | const TInput *inptr, size_t ld_in_row, size_t ld_in_col, size_t ld_in_vl, |
| 302 | unsigned int pad_top, unsigned int valid_input_rows, |
| 303 | unsigned int pad_left, unsigned int valid_input_cols, |
| 304 | const TWeight *weights, const TAccum *bias, |
| 305 | TOutput *outptr, size_t ld_out_row, size_t ld_out_col, size_t ld_out_vl, |
| 306 | unsigned int valid_output_rows, unsigned int valid_output_cols, |
| 307 | unsigned int first_channel, unsigned int valid_channels, |
| 308 | WorkspaceType *ws |
| 309 | ) const |
| 310 | { |
| 311 | // Initialise the output pointers |
| 312 | for (auto i = 0u; i < m_strat->get_output_rows(); i++) |
| 313 | { |
| 314 | // Point at the output tensor for all valid rows; otherwise point at the |
| 315 | // padding buffer. |
| 316 | ws->output_row_ptrs[i] = i < valid_output_rows ? outptr : ws->output_padding_buffer; |
| 317 | ws->output_ld_cols[i] = i < valid_output_rows ? ld_out_col : 0; |
| 318 | ws->output_ld_vls[i] = i < valid_output_rows ? ld_out_vl : 0; |
| 319 | outptr += ld_out_row; |
| 320 | } |
| 321 | |
| 322 | // Execute the kernel |
| 323 | PlanarKernelType<TInput, TWeight, TOutput, TAccum, OutputStage>::template execute<WorkspaceType>( |
| 324 | reinterpret_cast<const PlanarStrategy<TInput, TWeight, TOutput, TAccum, OutputStage> *>(m_strat.get())->get_kernel(), |
| 325 | inptr, ld_in_row, ld_in_col, ld_in_vl, |
| 326 | pad_top, valid_input_rows, pad_left, valid_input_cols, |
| 327 | weights, bias, |
| 328 | ws->output_row_ptrs, ws->output_ld_cols, ws->output_ld_vls, |
| 329 | valid_output_cols, first_channel, valid_channels, |
| 330 | this->m_os, ws |
| 331 | ); |
| 332 | } |
| 333 | |
| 334 | void execute_internal( |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 335 | const DepthwiseArgs &args, |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 336 | const void *input, |
| 337 | size_t ld_input_col, |
| 338 | size_t ld_input_row, |
| 339 | size_t ld_input_batch, |
| 340 | const void *parameters, |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 341 | void *output, |
| 342 | size_t ld_output_col, |
| 343 | size_t ld_output_row, |
| 344 | size_t ld_output_batch, |
| 345 | void *working_space, |
| 346 | unsigned int thread_id, |
| 347 | unsigned int n_threads |
| 348 | ) const override |
| 349 | { |
| 350 | // Get and initialise the working space for this thread. |
| 351 | void *thread_working_space = |
| 352 | static_cast<uint8_t *>(working_space) + thread_id * this->get_working_size_per_thread(); |
| 353 | this->initialise_working_space(thread_working_space); |
| 354 | auto ws = reinterpret_cast<WorkspaceType *>(thread_working_space); |
| 355 | |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 356 | const auto n_output_channels = args.input_channels * args.channel_multiplier; |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 357 | const auto vl = get_vector_length<TAccum>(m_strat->get_vl_type()); |
| 358 | |
| 359 | // Get typed pointers |
| 360 | auto input_batch = reinterpret_cast<const TInput *>(input); |
| 361 | auto output_batch = reinterpret_cast<TOutput *>(output); |
| 362 | auto weights = reinterpret_cast<const TWeight *>(parameters); |
| 363 | |
| 364 | // Iterate over batches |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 365 | for (auto batches = args.n_batches; batches; batches--) |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 366 | { |
| 367 | // NOTE: Other loop orderings are possible and it would be worth |
| 368 | // investigating them. |
| 369 | |
| 370 | // Within a batch, stripe threads across rows. |
| 371 | for (auto start_output_i = thread_id * m_strat->get_output_rows(); |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 372 | start_output_i < args.output_rows; |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 373 | start_output_i += n_threads * m_strat->get_output_rows()) |
| 374 | { |
| 375 | // Determine what (if any padding) is required on the top/bottom of |
| 376 | // this row of the convolution. |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 377 | const int start_input_i = start_output_i * args.stride_rows - args.padding.top; |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 378 | const unsigned int input_pad_top = start_input_i < 0 ? -start_input_i : 0; |
| 379 | const unsigned int input_i = start_input_i < 0 ? 0 : start_input_i; |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 380 | const unsigned int valid_input_rows = input_i > args.input_rows ? 0 : args.input_rows - input_i; |
| 381 | const unsigned int valid_output_rows = args.output_rows - start_output_i; |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 382 | |
| 383 | auto inptr_row = input_batch + input_i*ld_input_row; |
| 384 | auto outptr_row = output_batch + start_output_i * ld_output_row; |
| 385 | |
| 386 | // Execute the kernel |
| 387 | this->execute_kernel( |
| 388 | inptr_row, ld_input_row, ld_input_col, vl, |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 389 | input_pad_top, valid_input_rows, args.padding.left, args.input_cols, |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 390 | weights, this->m_bias, |
| 391 | outptr_row, ld_output_row, ld_output_col, vl, |
Pablo Marquez Tello | 4e2bbbb | 2023-01-09 17:21:01 +0000 | [diff] [blame] | 392 | valid_output_rows, args.output_cols, |
ramelg01 | 8a16488 | 2022-04-07 02:42:52 +0100 | [diff] [blame] | 393 | 0 /* first channel */, n_output_channels, |
| 394 | ws |
| 395 | ); |
| 396 | } |
| 397 | |
| 398 | // Update the input and output pointers to account for batch |
| 399 | input_batch += ld_input_batch; |
| 400 | output_batch += ld_output_batch; |
| 401 | } |
| 402 | } |
| 403 | }; |
| 404 | |
| 405 | } // namespace depthwise |
| 406 | } // namespace arm_conv |