1 // SPDX-License-Identifier: Apache-2.0 2 // ---------------------------------------------------------------------------- 3 // Copyright 2011-2024 Arm Limited 4 // 5 // Licensed under the Apache License, Version 2.0 (the "License"); you may not 6 // use this file except in compliance with the License. You may obtain a copy 7 // of the License at: 8 // 9 // http://www.apache.org/licenses/LICENSE-2.0 10 // 11 // Unless required by applicable law or agreed to in writing, software 12 // distributed under the License is distributed on an "AS IS" BASIS, WITHOUT 13 // WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the 14 // License for the specific language governing permissions and limitations 15 // under the License. 16 // ---------------------------------------------------------------------------- 17 18 /** 19 * @brief Functions and data declarations for the outer context. 20 * 21 * The outer context includes thread-pool management, which is slower to 22 * compile due to increased use of C++ stdlib. The inner context used in the 23 * majority of the codec library does not include this. 24 */ 25 26 #ifndef ASTCENC_INTERNAL_ENTRY_INCLUDED 27 #define ASTCENC_INTERNAL_ENTRY_INCLUDED 28 29 #include <atomic> 30 #include <condition_variable> 31 #include <functional> 32 #include <mutex> 33 34 #include "astcenc_internal.h" 35 36 /* ============================================================================ 37 Parallel execution control 38 ============================================================================ */ 39 40 /** 41 * @brief A simple counter-based manager for parallel task execution. 42 * 43 * The task processing execution consists of: 44 * 45 * * A single-threaded init stage. 46 * * A multi-threaded processing stage. 47 * * A condition variable so threads can wait for processing completion. 48 * 49 * The init stage will be executed by the first thread to arrive in the critical section, there is 50 * no main thread in the thread pool. 51 * 52 * The processing stage uses dynamic dispatch to assign task tickets to threads on an on-demand 53 * basis. Threads may each therefore executed different numbers of tasks, depending on their 54 * processing complexity. The task queue and the task tickets are just counters; the caller must map 55 * these integers to an actual processing partition in a specific problem domain. 56 * 57 * The exit wait condition is needed to ensure processing has finished before a worker thread can 58 * progress to the next stage of the pipeline. Specifically a worker may exit the processing stage 59 * because there are no new tasks to assign to it while other worker threads are still processing. 60 * Calling @c wait() will ensure that all other worker have finished before the thread can proceed. 61 * 62 * The basic usage model: 63 * 64 * // --------- From single-threaded code --------- 65 * 66 * // Reset the tracker state 67 * manager->reset() 68 * 69 * // --------- From multi-threaded code --------- 70 * 71 * // Run the stage init; only first thread actually runs the lambda 72 * manager->init(<lambda>) 73 * 74 * do 75 * { 76 * // Request a task assignment 77 * uint task_count; 78 * uint base_index = manager->get_tasks(<granule>, task_count); 79 * 80 * // Process any tasks we were given (task_count <= granule size) 81 * if (task_count) 82 * { 83 * // Run the user task processing code for N tasks here 84 * ... 85 * 86 * // Flag these tasks as complete 87 * manager->complete_tasks(task_count); 88 * } 89 * } while (task_count); 90 * 91 * // Wait for all threads to complete tasks before progressing 92 * manager->wait() 93 * 94 * // Run the stage term; only first thread actually runs the lambda 95 * manager->term(<lambda>) 96 */ 97 class ParallelManager 98 { 99 private: 100 /** @brief Lock used for critical section and condition synchronization. */ 101 std::mutex m_lock; 102 103 /** @brief True if the stage init() step has been executed. */ 104 bool m_init_done; 105 106 /** @brief True if the stage term() step has been executed. */ 107 bool m_term_done; 108 109 /** @brief Condition variable for tracking stage processing completion. */ 110 std::condition_variable m_complete; 111 112 /** @brief Number of tasks started, but not necessarily finished. */ 113 std::atomic<unsigned int> m_start_count; 114 115 /** @brief Number of tasks finished. */ 116 unsigned int m_done_count; 117 118 /** @brief Number of tasks that need to be processed. */ 119 unsigned int m_task_count; 120 121 /** @brief Progress callback (optional). */ 122 astcenc_progress_callback m_callback; 123 124 /** @brief Lock used for callback synchronization. */ 125 std::mutex m_callback_lock; 126 127 /** @brief Minimum progress before making a callback. */ 128 float m_callback_min_diff; 129 130 /** @brief Last progress callback value. */ 131 float m_callback_last_value; 132 133 public: 134 /** @brief Create a new ParallelManager. */ ParallelManager()135 ParallelManager() 136 { 137 reset(); 138 } 139 140 /** 141 * @brief Reset the tracker for a new processing batch. 142 * 143 * This must be called from single-threaded code before starting the multi-threaded processing 144 * operations. 145 */ reset()146 void reset() 147 { 148 m_init_done = false; 149 m_term_done = false; 150 m_start_count = 0; 151 m_done_count = 0; 152 m_task_count = 0; 153 m_callback_last_value = 0.0f; 154 m_callback_min_diff = 1.0f; 155 } 156 157 /** 158 * @brief Trigger the pipeline stage init step. 159 * 160 * This can be called from multi-threaded code. The first thread to hit this will process the 161 * initialization. Other threads will block and wait for it to complete. 162 * 163 * @param init_func Callable which executes the stage initialization. It must return the 164 * total number of tasks in the stage. 165 */ init(std::function<unsigned int (void)> init_func)166 void init(std::function<unsigned int(void)> init_func) 167 { 168 std::lock_guard<std::mutex> lck(m_lock); 169 if (!m_init_done) 170 { 171 m_task_count = init_func(); 172 m_init_done = true; 173 } 174 } 175 176 /** 177 * @brief Trigger the pipeline stage init step. 178 * 179 * This can be called from multi-threaded code. The first thread to hit this will process the 180 * initialization. Other threads will block and wait for it to complete. 181 * 182 * @param task_count Total number of tasks needing processing. 183 * @param callback Function pointer for progress status callbacks. 184 */ init(unsigned int task_count,astcenc_progress_callback callback)185 void init(unsigned int task_count, astcenc_progress_callback callback) 186 { 187 std::lock_guard<std::mutex> lck(m_lock); 188 if (!m_init_done) 189 { 190 m_callback = callback; 191 m_task_count = task_count; 192 m_init_done = true; 193 194 // Report every 1% or 4096 blocks, whichever is larger, to avoid callback overhead 195 float min_diff = (4096.0f / static_cast<float>(task_count)) * 100.0f; 196 m_callback_min_diff = astc::max(min_diff, 1.0f); 197 } 198 } 199 200 /** 201 * @brief Request a task assignment. 202 * 203 * Assign up to @c granule tasks to the caller for processing. 204 * 205 * @param granule Maximum number of tasks that can be assigned. 206 * @param[out] count Actual number of tasks assigned, or zero if no tasks were assigned. 207 * 208 * @return Task index of the first assigned task; assigned tasks increment from this. 209 */ get_task_assignment(unsigned int granule,unsigned int & count)210 unsigned int get_task_assignment(unsigned int granule, unsigned int& count) 211 { 212 unsigned int base = m_start_count.fetch_add(granule, std::memory_order_relaxed); 213 if (base >= m_task_count) 214 { 215 count = 0; 216 return 0; 217 } 218 219 count = astc::min(m_task_count - base, granule); 220 return base; 221 } 222 223 /** 224 * @brief Complete a task assignment. 225 * 226 * Mark @c count tasks as complete. This will notify all threads blocked on @c wait() if this 227 * completes the processing of the stage. 228 * 229 * @param count The number of completed tasks. 230 */ complete_task_assignment(unsigned int count)231 void complete_task_assignment(unsigned int count) 232 { 233 // Note: m_done_count cannot use an atomic without the mutex; this has a race between the 234 // update here and the wait() for other threads 235 unsigned int local_count; 236 float local_last_value; 237 { 238 std::unique_lock<std::mutex> lck(m_lock); 239 m_done_count += count; 240 local_count = m_done_count; 241 local_last_value = m_callback_last_value; 242 243 if (m_done_count == m_task_count) 244 { 245 // Ensure the progress bar hits 100% 246 if (m_callback) 247 { 248 std::unique_lock<std::mutex> cblck(m_callback_lock); 249 m_callback(100.0f); 250 m_callback_last_value = 100.0f; 251 } 252 253 lck.unlock(); 254 m_complete.notify_all(); 255 } 256 } 257 258 // Process progress callback if we have one 259 if (m_callback) 260 { 261 // Initial lockless test - have we progressed enough to emit? 262 float num = static_cast<float>(local_count); 263 float den = static_cast<float>(m_task_count); 264 float this_value = (num / den) * 100.0f; 265 bool report_test = (this_value - local_last_value) > m_callback_min_diff; 266 267 // Recheck under lock, because another thread might report first 268 if (report_test) 269 { 270 std::unique_lock<std::mutex> cblck(m_callback_lock); 271 bool report_retest = (this_value - m_callback_last_value) > m_callback_min_diff; 272 if (report_retest) 273 { 274 m_callback(this_value); 275 m_callback_last_value = this_value; 276 } 277 } 278 } 279 } 280 281 /** 282 * @brief Wait for stage processing to complete. 283 */ wait()284 void wait() 285 { 286 std::unique_lock<std::mutex> lck(m_lock); 287 m_complete.wait(lck, [this]{ return m_done_count == m_task_count; }); 288 } 289 290 /** 291 * @brief Trigger the pipeline stage term step. 292 * 293 * This can be called from multi-threaded code. The first thread to hit this will process the 294 * work pool termination. Caller must have called @c wait() prior to calling this function to 295 * ensure that processing is complete. 296 * 297 * @param term_func Callable which executes the stage termination. 298 */ term(std::function<void (void)> term_func)299 void term(std::function<void(void)> term_func) 300 { 301 std::lock_guard<std::mutex> lck(m_lock); 302 if (!m_term_done) 303 { 304 term_func(); 305 m_term_done = true; 306 } 307 } 308 }; 309 310 /** 311 * @brief The astcenc compression context. 312 */ 313 struct astcenc_context 314 { 315 /** @brief The context internal state. */ 316 astcenc_contexti context; 317 318 #if !defined(ASTCENC_DECOMPRESS_ONLY) 319 /** @brief The parallel manager for averages computation. */ 320 ParallelManager manage_avg; 321 322 /** @brief The parallel manager for compression. */ 323 ParallelManager manage_compress; 324 #endif 325 326 /** @brief The parallel manager for decompression. */ 327 ParallelManager manage_decompress; 328 }; 329 330 #endif 331