Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +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 "arm_compute/runtime/NEON/functions/assembly/NEGEMMInterleavedWrapper.h" |
| 26 | |
| 27 | #include "arm_compute/core/ITensor.h" |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/NEON/kernels/assembly/Helpers.h" |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/Utils.h" |
| 30 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 31 | |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 32 | #include "src/core/NEON/kernels/assembly/NEGEMMInterleavedStrategies.h" |
| 33 | |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 34 | #include <atomic> |
| 35 | #include <condition_variable> |
| 36 | #include <mutex> |
| 37 | |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 38 | namespace arm_compute |
| 39 | { |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 40 | #ifndef NO_MULTI_THREADING |
| 41 | class BufferManagerMultipleThreads final : public IBufferManager |
| 42 | { |
| 43 | public: |
| 44 | /** Number of buffers to ping pong between */ |
| 45 | static constexpr unsigned int NUM_BUFFERS = 3; |
| 46 | |
| 47 | explicit BufferManagerMultipleThreads(unsigned int max_num_users) |
| 48 | : _max_num_users(max_num_users) |
| 49 | { |
| 50 | } |
| 51 | unsigned int num_buffers() const override |
| 52 | { |
| 53 | return NUM_BUFFERS; |
| 54 | } |
| 55 | /* - Lock the requested index if it's free and return true if it needs reshaping. |
| 56 | * - Return false without acquiring the lock if the buffer at the index is already reshaped / being reshaped. |
| 57 | * - Block if the corresponding buffer for the given index is still being used by a different index. |
| 58 | */ |
| 59 | bool lock_to_reshape_if_needed(unsigned int index) override |
| 60 | { |
| 61 | Buffer &buf = get_buffer_from_index(index); |
| 62 | while(true) |
| 63 | { |
| 64 | if(buf.index == index && buf.state != State::FREE) |
| 65 | { |
| 66 | //Another thread already is reshaping / has reshaped this block: nothing to do |
| 67 | return false; |
| 68 | } |
| 69 | else |
| 70 | { |
| 71 | std::unique_lock<std::mutex> lock(buf.mutex); |
| 72 | //If the buffer is free then lock it for reshaping: |
| 73 | if(buf.state == State::FREE) |
| 74 | { |
| 75 | buf.index = index; |
| 76 | buf.state = State::BEING_RESHAPED; |
| 77 | return true; |
| 78 | } |
| 79 | // Check again just in case it changed while we were acquiring the lock: |
| 80 | if(buf.index == index) |
| 81 | { |
| 82 | //Another thread is reshaping this block already, nothing to do |
| 83 | return false; |
| 84 | } |
| 85 | // buf.index != index: Buffer still being used by another block, need to wait |
| 86 | buf.sem.wait(lock); |
| 87 | } |
| 88 | } |
| 89 | } |
| 90 | /* Mark the buffer at the given index as reshaped and release the lock acquired via lock_to_reshape_if_needed() */ |
| 91 | void mark_as_reshaped(unsigned int index) override |
| 92 | { |
| 93 | Buffer &buf = get_buffer_from_index(index); |
| 94 | { |
| 95 | std::lock_guard<std::mutex> lock(buf.mutex); |
| 96 | buf.users = _max_num_users; |
| 97 | buf.state = State::IN_USE; |
| 98 | } |
| 99 | buf.sem.notify_all(); |
| 100 | } |
| 101 | |
| 102 | /* Block until the buffer at the given index is reshaped */ |
| 103 | void wait_for_reshaping(unsigned int index) override |
| 104 | { |
| 105 | Buffer &buf = get_buffer_from_index(index); |
| 106 | ARM_COMPUTE_ERROR_ON(buf.index != index); // Should have blocked in lock_to_reshape_if_needed() |
| 107 | // Check if it's already ready to use: |
| 108 | if(buf.state == State::IN_USE) |
| 109 | return; |
| 110 | std::unique_lock<std::mutex> lock(buf.mutex); |
| 111 | //Double check it didn't change while we were acquiring the lock: |
| 112 | if(buf.state == State::IN_USE) |
| 113 | return; |
| 114 | buf.sem.wait(lock); |
| 115 | } |
| 116 | /* Mark the buffer at the given index as not used by this thread anymore. |
| 117 | * Once all the threads have called this method then the buffer is marked as free again. |
| 118 | */ |
| 119 | void mark_as_unused(unsigned int index) override |
| 120 | { |
| 121 | Buffer &buf = get_buffer_from_index(index); |
| 122 | ARM_COMPUTE_ERROR_ON(buf.index != index); // Should have blocked in lock_to_reshape_if_needed() |
| 123 | if(--buf.users == 0) |
| 124 | { |
| 125 | std::unique_lock<std::mutex> lock(buf.mutex); |
| 126 | buf.state = State::FREE; |
| 127 | lock.unlock(); |
| 128 | buf.sem.notify_all(); |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | private: |
| 133 | enum class State |
| 134 | { |
| 135 | FREE, |
| 136 | BEING_RESHAPED, |
| 137 | IN_USE |
| 138 | }; |
| 139 | struct Buffer |
| 140 | { |
| 141 | unsigned int index{}; |
| 142 | std::atomic_uint users{}; |
| 143 | State state{ State::FREE }; |
| 144 | std::mutex mutex{}; |
| 145 | std::condition_variable sem{}; |
| 146 | } _buffers[NUM_BUFFERS]; |
| 147 | Buffer &get_buffer_from_index(unsigned int index) |
| 148 | { |
| 149 | return _buffers[index % NUM_BUFFERS]; |
| 150 | } |
| 151 | unsigned int _max_num_users; |
| 152 | }; |
| 153 | #endif /* NO_MULTI_THREADING */ |
| 154 | |
| 155 | class BufferManagerSingleThread : public IBufferManager |
| 156 | { |
| 157 | public: |
| 158 | unsigned int num_buffers() const override |
| 159 | { |
| 160 | return 1; |
| 161 | } |
| 162 | bool lock_to_reshape_if_needed(unsigned int index) override |
| 163 | { |
| 164 | return true; |
| 165 | } |
| 166 | void mark_as_reshaped(unsigned int index) override |
| 167 | { |
| 168 | } |
| 169 | void wait_for_reshaping(unsigned int index) override |
| 170 | { |
| 171 | } |
| 172 | void mark_as_unused(unsigned int index) override |
| 173 | { |
| 174 | } |
| 175 | }; |
| 176 | |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 177 | NEGEMMInterleavedWrapper::NEGEMMInterleavedWrapper(std::shared_ptr<IMemoryManager> memory_manager) |
| 178 | : _memory_group(std::move(memory_manager)) |
| 179 | { |
| 180 | } |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 181 | |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 182 | void NEGEMMInterleavedWrapper::run() |
| 183 | { |
| 184 | prepare(); |
| 185 | |
| 186 | _memory_group.acquire(); |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 187 | NEScheduler::get().run_tagged_workloads(_workloads, _tag.c_str()); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 188 | _memory_group.release(); |
| 189 | } |
| 190 | |
| 191 | void NEGEMMInterleavedWrapper::prepare() |
| 192 | { |
| 193 | if(!_is_prepared) |
| 194 | { |
| 195 | if(_pretranspose_b) |
| 196 | { |
Georgios Pinitas | ca1250d | 2018-11-22 19:38:27 +0000 | [diff] [blame] | 197 | _transformed_b.allocator()->allocate(); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 198 | NEScheduler::get().schedule(_prepare_b.get(), Window::DimX); |
| 199 | _b->mark_as_unused(); |
| 200 | } |
| 201 | else |
| 202 | { |
| 203 | _prepare_b->create_workloads(_b_workloads); |
| 204 | } |
| 205 | _transform_a->create_workloads(_a_workloads); |
| 206 | _matrix_multiply->create_workloads(_mm_workloads); |
| 207 | |
| 208 | //Maximum number of workloads to create: |
| 209 | const unsigned int num_threads = NEScheduler::get().num_threads(); |
Georgios Pinitas | 1509e4b | 2019-01-28 10:01:50 +0000 | [diff] [blame] | 210 | const unsigned int max_iterations = std::max(num_threads, _num_windows); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 211 | //Maximum number of iterations the parameters allow: |
| 212 | const unsigned int num_iterations = _batch_window.num_iterations_total(); |
| 213 | // Keep the smallest of the two: |
| 214 | const unsigned int num_windows = std::min(num_iterations, max_iterations); |
| 215 | const TensorShape window_shape = _batch_window.shape(); |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 216 | const unsigned int num_x_blocks = _block_walker.num_iterations(Window::DimX); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 217 | |
| 218 | // Create a 1D window to dynamically split the batch window: |
| 219 | Window win_1D; |
| 220 | win_1D.set(0, Window::Dimension(0, num_iterations)); |
| 221 | |
| 222 | // Create one workload for each sub-window: |
| 223 | for(unsigned int w = 0; w < num_windows; w++) |
| 224 | { |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 225 | Window win = win_1D.split_window(0, w, num_windows); |
| 226 | const Coordinates start_offset = index2coords(window_shape, win.x().start()); |
| 227 | const Coordinates end_offset = index2coords(window_shape, win.x().end() - 1); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 228 | |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 229 | if(_pretranspose_b) |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 230 | { |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 231 | auto workload = [start_offset, end_offset, num_x_blocks, this](const ThreadInfo & info) |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 232 | { |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 233 | //For each block of rows in "M" |
| 234 | auto workload_mm = this->_mm_workloads.begin(); |
| 235 | for(auto workload_a = this->_a_workloads.begin(); workload_a != this->_a_workloads.end(); workload_a++) |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 236 | { |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 237 | // Transform one k_block from A: |
| 238 | this->_transform_a->transform(*workload_a, info, this->_batch_window, start_offset, end_offset); |
| 239 | // Then perform the matrix multiplication for each x block along N: |
| 240 | for(unsigned int i = 0; i < num_x_blocks; i++) |
| 241 | { |
| 242 | ARM_COMPUTE_ERROR_ON(workload_mm == this->_mm_workloads.end()); |
| 243 | this->_matrix_multiply->transform(*workload_mm++, info, this->_batch_window, start_offset, end_offset); |
| 244 | } |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 245 | } |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 246 | }; |
| 247 | _workloads.push_back(workload); |
| 248 | } |
| 249 | else |
| 250 | { |
| 251 | auto workload = [num_threads, start_offset, end_offset, num_x_blocks, this](const ThreadInfo & info) |
| 252 | { |
| 253 | //For each block of rows in "M" |
| 254 | auto workload_mm = this->_mm_workloads.begin(); |
| 255 | unsigned int workload_b = 0; |
| 256 | //If there is only one thread then only reshape the B blocks as you need them: |
| 257 | unsigned int workload_b_next = num_threads == 1 ? this->_b_workloads.size() : 1; |
| 258 | |
| 259 | for(auto workload_a = this->_a_workloads.begin(); workload_a != this->_a_workloads.end(); workload_a++) |
| 260 | { |
| 261 | // Transform one k_block from A: |
| 262 | this->_transform_a->transform(*workload_a, info, this->_batch_window, start_offset, end_offset); |
| 263 | // Then perform the matrix multiplication for each x block along N: |
| 264 | for(unsigned int i = 0; i < num_x_blocks; i++) |
| 265 | { |
| 266 | ARM_COMPUTE_ERROR_ON(workload_mm == this->_mm_workloads.end()); |
| 267 | if(workload_b_next < this->_b_workloads.size()) |
| 268 | { |
| 269 | //Lock on BufferManager: need to run it ? |
| 270 | if(this->_buffer_manager->lock_to_reshape_if_needed(workload_b_next)) |
| 271 | { |
| 272 | this->_prepare_b->transform(this->_b_workloads[workload_b_next], info); |
| 273 | this->_buffer_manager->mark_as_reshaped(workload_b_next); |
| 274 | } |
| 275 | workload_b_next++; |
| 276 | } |
| 277 | ARM_COMPUTE_ERROR_ON(workload_b >= this->_b_workloads.size()); |
| 278 | // Run if needed or wait |
| 279 | if(this->_buffer_manager->lock_to_reshape_if_needed(workload_b)) |
| 280 | { |
| 281 | this->_prepare_b->transform(this->_b_workloads[workload_b], info); |
| 282 | this->_buffer_manager->mark_as_reshaped(workload_b); |
| 283 | } |
| 284 | this->_buffer_manager->wait_for_reshaping(workload_b); |
| 285 | this->_matrix_multiply->transform(*workload_mm++, info, this->_batch_window, start_offset, end_offset); |
| 286 | this->_buffer_manager->mark_as_unused(workload_b); |
| 287 | workload_b++; |
| 288 | } |
| 289 | } |
| 290 | }; |
| 291 | _workloads.push_back(workload); |
| 292 | } |
| 293 | } |
| 294 | if(!_pretranspose_b && num_windows > 1 && num_windows % num_threads != 0) |
| 295 | { |
| 296 | //Make sure the number of workloads is a multiple of the number of threads to avoid dead locks: |
| 297 | for(unsigned int leftover = num_windows % num_threads; leftover != num_threads; leftover++) |
| 298 | { |
| 299 | auto workload = [this](const ThreadInfo & info) |
| 300 | { |
| 301 | unsigned int workload_b = 0; |
| 302 | //If there is only one thread then only reshape the B blocks as you need them: |
| 303 | unsigned int workload_b_next = 1; |
| 304 | |
| 305 | for(unsigned int iteration = 0; iteration < this->_mm_workloads.size(); iteration++) |
| 306 | { |
| 307 | if(workload_b_next < this->_b_workloads.size()) |
| 308 | { |
| 309 | //Lock on BufferManager: need to run it ? |
| 310 | if(this->_buffer_manager->lock_to_reshape_if_needed(workload_b_next)) |
| 311 | { |
| 312 | this->_prepare_b->transform(this->_b_workloads[workload_b_next], info); |
| 313 | this->_buffer_manager->mark_as_reshaped(workload_b_next); |
| 314 | } |
| 315 | workload_b_next++; |
| 316 | } |
| 317 | ARM_COMPUTE_ERROR_ON(workload_b >= this->_b_workloads.size()); |
| 318 | // Run if needed or wait |
| 319 | if(this->_buffer_manager->lock_to_reshape_if_needed(workload_b)) |
| 320 | { |
| 321 | this->_prepare_b->transform(this->_b_workloads[workload_b], info); |
| 322 | this->_buffer_manager->mark_as_reshaped(workload_b); |
| 323 | } |
| 324 | this->_buffer_manager->wait_for_reshaping(workload_b); |
| 325 | this->_buffer_manager->mark_as_unused(workload_b); |
| 326 | workload_b++; |
| 327 | } |
| 328 | }; |
| 329 | _workloads.push_back(workload); |
| 330 | } |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 331 | } |
| 332 | |
| 333 | _is_prepared = true; |
| 334 | } |
| 335 | } |
| 336 | |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 337 | void NEGEMMInterleavedWrapper::configure(const ITensor *a, const ITensor *b, ITensor *c, float alpha, float beta, bool pretranspose_b) |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 338 | { |
| 339 | _params = INEGEMMWrapperKernel::extract_parameters(a, b, c); |
| 340 | _a = a; |
| 341 | _b = b; |
| 342 | _c = c; |
| 343 | _pretranspose_b = pretranspose_b; |
| 344 | |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 345 | const DataType input_type = a->info()->data_type(); |
| 346 | const CPUInfo &ci = NEScheduler::get().cpu_info(); |
| 347 | const unsigned int num_threads = NEScheduler::get().num_threads(); |
| 348 | |
| 349 | const arm_gemm::KernelDescription gemm_kernel_info = get_gemm_info(input_type, ci, num_threads, _params, alpha, beta, pretranspose_b); |
| 350 | ARM_COMPUTE_ERROR_ON(gemm_kernel_info.method != arm_gemm::GemmMethod::GEMM_INTERLEAVED); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 351 | |
| 352 | // Forcing 128-byte alignment (required by 32-bit kernels) |
| 353 | const unsigned int alignment = 128; |
| 354 | _transformed_b.allocator()->init(TensorInfo{}, alignment); |
| 355 | _tmp_c.allocator()->init(TensorInfo{}, alignment); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 356 | _tag = "NEGEMMInterleaved_" + gemm_kernel_info.name; |
| 357 | |
| 358 | // Get strategy |
| 359 | std::unique_ptr<detail::IInterleavedStrategy> strategy = detail::create_strategy(gemm_kernel_info.name); |
Georgios Pinitas | 1509e4b | 2019-01-28 10:01:50 +0000 | [diff] [blame] | 360 | _num_windows = iceildiv(_params.M, strategy->out_height()) * _params.batches; |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 361 | ARM_COMPUTE_ERROR_ON(strategy == nullptr); |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 362 | |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 363 | if(!_pretranspose_b) |
| 364 | { |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 365 | _block_sizes = strategy->calculate_block_sizes_for_strategy(ci, _params); |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 366 | _batch_window.set(Window::DimX, Window::Dimension(0, ceil_to_multiple(_block_sizes.m_round, _block_sizes.strategy_out_height), _block_sizes.strategy_out_height)); |
| 367 | _batch_window.set(Window::DimY, Window::Dimension(0, _params.batches)); |
| 368 | // If the execution is single threaded or has only one window then the buffer manager only needs 1 buffer else we will use NUM_BUFFERS buffers and ping pong between them: |
| 369 | const unsigned int num_iterations = _batch_window.num_iterations_total(); |
| 370 | if(NEScheduler::get().num_threads() == 1 || num_iterations == 1) |
| 371 | { |
| 372 | _buffer_manager = support::cpp14::make_unique<BufferManagerSingleThread>(); |
| 373 | } |
| 374 | else |
| 375 | { |
| 376 | #ifdef NO_MULTI_THREADING |
| 377 | ARM_COMPUTE_ERROR("Can't have more than 1 buffer without multiple threads"); |
| 378 | #else /* NO_MULTI_THREADING */ |
| 379 | _buffer_manager = support::cpp14::make_unique<BufferManagerMultipleThreads>(NEScheduler::get().num_threads()); |
| 380 | #endif /* NO_MULTI_THREADING */ |
| 381 | } |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 382 | // If B is transposed at every iteration then transformed_B can be managed: |
| 383 | _memory_group.manage(&_transformed_b); |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 384 | auto_init_if_empty(*_transformed_b.info(), _b->info()->clone()->set_tensor_shape(TensorShape(_block_sizes.x_block * _block_sizes.k_block, _buffer_manager->num_buffers()))); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 385 | } |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 386 | else |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 387 | { |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 388 | _tag += "_preB"; |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 389 | } |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 390 | |
| 391 | _prepare_b = strategy->instantiate_prepareB(b, &_transformed_b, _params, ci); |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 392 | ARM_COMPUTE_ERROR_ON(_prepare_b == nullptr); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 393 | |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 394 | if(_pretranspose_b) |
| 395 | { |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 396 | _block_sizes = _prepare_b->block_sizes(); |
Anthony Barbier | ff0bccf | 2018-11-30 10:42:40 +0000 | [diff] [blame] | 397 | _batch_window.set(Window::DimX, Window::Dimension(0, ceil_to_multiple(_block_sizes.m_round, _block_sizes.strategy_out_height), _block_sizes.strategy_out_height)); |
| 398 | _batch_window.set(Window::DimY, Window::Dimension(0, _params.batches)); |
Anthony Barbier | ac314c2 | 2018-09-11 17:49:10 +0100 | [diff] [blame] | 399 | } |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 400 | |
| 401 | _block_walker.set(Window::DimX, Window::Dimension(0, ceil_to_multiple(_params.N, _block_sizes.x_block), _block_sizes.x_block)); |
| 402 | _block_walker.set(Window::DimY, Window::Dimension(0, ceil_to_multiple(_params.K, _block_sizes.k_block), _block_sizes.k_block)); |
| 403 | _block_walker.set(Window::DimZ, Window::Dimension(0, _params.multis)); |
| 404 | |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 405 | _transformed_a.allocator()->init(TensorInfo(TensorShape{ _block_sizes.k_block, _block_sizes.m_round, _params.batches }, 1, input_type), alignment); |
| 406 | _memory_group.manage(&_transformed_a); |
| 407 | _memory_group.manage(&_tmp_c); |
| 408 | |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 409 | _transform_a = strategy->instantiate_transformA(_a, &_transformed_a, _block_walker, _params); |
| 410 | _matrix_multiply = strategy->instantiate_matrix_multiply(&_transformed_a, &_transformed_b, &_tmp_c, c, _block_walker, _block_sizes, _params, alpha, beta, pretranspose_b, num_threads); |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 411 | ARM_COMPUTE_ERROR_ON(_transform_a == nullptr); |
| 412 | ARM_COMPUTE_ERROR_ON(_matrix_multiply == nullptr); |
Georgios Pinitas | 7cd26d4 | 2019-01-09 18:35:17 +0000 | [diff] [blame] | 413 | |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 414 | _transformed_a.allocator()->allocate(); |
| 415 | _tmp_c.allocator()->allocate(); |
Georgios Pinitas | ca1250d | 2018-11-22 19:38:27 +0000 | [diff] [blame] | 416 | if(!_pretranspose_b) |
| 417 | { |
| 418 | _transformed_b.allocator()->allocate(); |
| 419 | } |
Anthony Barbier | 3d677cc | 2018-07-23 16:42:59 +0100 | [diff] [blame] | 420 | } |
| 421 | } // namespace arm_compute |