1 // -*- C++ -*- 2 3 // Copyright (C) 2007-2018 Free Software Foundation, Inc. 4 // 5 // This file is part of the GNU ISO C++ Library. This library is free 6 // software; you can redistribute it and/or modify it under the terms 7 // of the GNU General Public License as published by the Free Software 8 // Foundation; either version 3, or (at your option) any later 9 // version. 10 11 // This library is distributed in the hope that it will be useful, but 12 // WITHOUT ANY WARRANTY; without even the implied warranty of 13 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 // General Public License for more details. 15 16 // Under Section 7 of GPL version 3, you are granted additional 17 // permissions described in the GCC Runtime Library Exception, version 18 // 3.1, as published by the Free Software Foundation. 19 20 // You should have received a copy of the GNU General Public License and 21 // a copy of the GCC Runtime Library Exception along with this program; 22 // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 23 // <http://www.gnu.org/licenses/>. 24 25 /** @file parallel/balanced_quicksort.h 26 * @brief Implementation of a dynamically load-balanced parallel quicksort. 27 * 28 * It works in-place and needs only logarithmic extra memory. 29 * The algorithm is similar to the one proposed in 30 * 31 * P. Tsigas and Y. Zhang. 32 * A simple, fast parallel implementation of quicksort and 33 * its performance evaluation on SUN enterprise 10000. 34 * In 11th Euromicro Conference on Parallel, Distributed and 35 * Network-Based Processing, page 372, 2003. 36 * 37 * This file is a GNU parallel extension to the Standard C++ Library. 38 */ 39 40 // Written by Johannes Singler. 41 42 #ifndef _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H 43 #define _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H 1 44 45 #include <parallel/basic_iterator.h> 46 #include <bits/stl_algo.h> 47 #include <bits/stl_function.h> 48 49 #include <parallel/settings.h> 50 #include <parallel/partition.h> 51 #include <parallel/random_number.h> 52 #include <parallel/queue.h> 53 54 #if _GLIBCXX_PARALLEL_ASSERTIONS 55 #include <parallel/checkers.h> 56 #ifdef _GLIBCXX_HAVE_UNISTD_H 57 #include <unistd.h> 58 #endif 59 #endif 60 61 namespace __gnu_parallel 62 { 63 /** @brief Information local to one thread in the parallel quicksort run. */ 64 template<typename _RAIter> 65 struct _QSBThreadLocal 66 { 67 typedef std::iterator_traits<_RAIter> _TraitsType; 68 typedef typename _TraitsType::difference_type _DifferenceType; 69 70 /** @brief Continuous part of the sequence, described by an 71 iterator pair. */ 72 typedef std::pair<_RAIter, _RAIter> _Piece; 73 74 /** @brief Initial piece to work on. */ 75 _Piece _M_initial; 76 77 /** @brief Work-stealing queue. */ 78 _RestrictedBoundedConcurrentQueue<_Piece> _M_leftover_parts; 79 80 /** @brief Number of threads involved in this algorithm. */ 81 _ThreadIndex _M_num_threads; 82 83 /** @brief Pointer to a counter of elements left over to sort. */ 84 volatile _DifferenceType* _M_elements_leftover; 85 86 /** @brief The complete sequence to sort. */ 87 _Piece _M_global; 88 89 /** @brief Constructor. 90 * @param __queue_size size of the work-stealing queue. */ 91 _QSBThreadLocal(int __queue_size) : _M_leftover_parts(__queue_size) { } 92 }; 93 94 /** @brief Balanced quicksort divide step. 95 * @param __begin Begin iterator of subsequence. 96 * @param __end End iterator of subsequence. 97 * @param __comp Comparator. 98 * @param __num_threads Number of threads that are allowed to work on 99 * this part. 100 * @pre @c (__end-__begin)>=1 */ 101 template<typename _RAIter, typename _Compare> 102 typename std::iterator_traits<_RAIter>::difference_type 103 __qsb_divide(_RAIter __begin, _RAIter __end, 104 _Compare __comp, _ThreadIndex __num_threads) 105 { 106 _GLIBCXX_PARALLEL_ASSERT(__num_threads > 0); 107 108 typedef std::iterator_traits<_RAIter> _TraitsType; 109 typedef typename _TraitsType::value_type _ValueType; 110 typedef typename _TraitsType::difference_type _DifferenceType; 111 112 _RAIter __pivot_pos = 113 __median_of_three_iterators(__begin, __begin + (__end - __begin) / 2, 114 __end - 1, __comp); 115 116 #if defined(_GLIBCXX_PARALLEL_ASSERTIONS) 117 // Must be in between somewhere. 118 _DifferenceType __n = __end - __begin; 119 120 _GLIBCXX_PARALLEL_ASSERT((!__comp(*__pivot_pos, *__begin) 121 && !__comp(*(__begin + __n / 2), 122 *__pivot_pos)) 123 || (!__comp(*__pivot_pos, *__begin) 124 && !__comp(*(__end - 1), *__pivot_pos)) 125 || (!__comp(*__pivot_pos, *(__begin + __n / 2)) 126 && !__comp(*__begin, *__pivot_pos)) 127 || (!__comp(*__pivot_pos, *(__begin + __n / 2)) 128 && !__comp(*(__end - 1), *__pivot_pos)) 129 || (!__comp(*__pivot_pos, *(__end - 1)) 130 && !__comp(*__begin, *__pivot_pos)) 131 || (!__comp(*__pivot_pos, *(__end - 1)) 132 && !__comp(*(__begin + __n / 2), 133 *__pivot_pos))); 134 #endif 135 136 // Swap pivot value to end. 137 if (__pivot_pos != (__end - 1)) 138 std::iter_swap(__pivot_pos, __end - 1); 139 __pivot_pos = __end - 1; 140 141 __gnu_parallel::__binder2nd<_Compare, _ValueType, _ValueType, bool> 142 __pred(__comp, *__pivot_pos); 143 144 // Divide, returning __end - __begin - 1 in the worst case. 145 _DifferenceType __split_pos = __parallel_partition(__begin, __end - 1, 146 __pred, 147 __num_threads); 148 149 // Swap back pivot to middle. 150 std::iter_swap(__begin + __split_pos, __pivot_pos); 151 __pivot_pos = __begin + __split_pos; 152 153 #if _GLIBCXX_PARALLEL_ASSERTIONS 154 _RAIter __r; 155 for (__r = __begin; __r != __pivot_pos; ++__r) 156 _GLIBCXX_PARALLEL_ASSERT(__comp(*__r, *__pivot_pos)); 157 for (; __r != __end; ++__r) 158 _GLIBCXX_PARALLEL_ASSERT(!__comp(*__r, *__pivot_pos)); 159 #endif 160 161 return __split_pos; 162 } 163 164 /** @brief Quicksort conquer step. 165 * @param __tls Array of thread-local storages. 166 * @param __begin Begin iterator of subsequence. 167 * @param __end End iterator of subsequence. 168 * @param __comp Comparator. 169 * @param __iam Number of the thread processing this function. 170 * @param __num_threads 171 * Number of threads that are allowed to work on this part. */ 172 template<typename _RAIter, typename _Compare> 173 void 174 __qsb_conquer(_QSBThreadLocal<_RAIter>** __tls, 175 _RAIter __begin, _RAIter __end, 176 _Compare __comp, 177 _ThreadIndex __iam, _ThreadIndex __num_threads, 178 bool __parent_wait) 179 { 180 typedef std::iterator_traits<_RAIter> _TraitsType; 181 typedef typename _TraitsType::value_type _ValueType; 182 typedef typename _TraitsType::difference_type _DifferenceType; 183 184 _DifferenceType __n = __end - __begin; 185 186 if (__num_threads <= 1 || __n <= 1) 187 { 188 __tls[__iam]->_M_initial.first = __begin; 189 __tls[__iam]->_M_initial.second = __end; 190 191 __qsb_local_sort_with_helping(__tls, __comp, __iam, __parent_wait); 192 193 return; 194 } 195 196 // Divide step. 197 _DifferenceType __split_pos = 198 __qsb_divide(__begin, __end, __comp, __num_threads); 199 200 #if _GLIBCXX_PARALLEL_ASSERTIONS 201 _GLIBCXX_PARALLEL_ASSERT(0 <= __split_pos && 202 __split_pos < (__end - __begin)); 203 #endif 204 205 _ThreadIndex 206 __num_threads_leftside = std::max<_ThreadIndex> 207 (1, std::min<_ThreadIndex>(__num_threads - 1, __split_pos 208 * __num_threads / __n)); 209 210 # pragma omp atomic 211 *__tls[__iam]->_M_elements_leftover -= (_DifferenceType)1; 212 213 // Conquer step. 214 # pragma omp parallel num_threads(2) 215 { 216 bool __wait; 217 if(omp_get_num_threads() < 2) 218 __wait = false; 219 else 220 __wait = __parent_wait; 221 222 # pragma omp sections 223 { 224 # pragma omp section 225 { 226 __qsb_conquer(__tls, __begin, __begin + __split_pos, __comp, 227 __iam, __num_threads_leftside, __wait); 228 __wait = __parent_wait; 229 } 230 // The pivot_pos is left in place, to ensure termination. 231 # pragma omp section 232 { 233 __qsb_conquer(__tls, __begin + __split_pos + 1, __end, __comp, 234 __iam + __num_threads_leftside, 235 __num_threads - __num_threads_leftside, __wait); 236 __wait = __parent_wait; 237 } 238 } 239 } 240 } 241 242 /** 243 * @brief Quicksort step doing load-balanced local sort. 244 * @param __tls Array of thread-local storages. 245 * @param __comp Comparator. 246 * @param __iam Number of the thread processing this function. 247 */ 248 template<typename _RAIter, typename _Compare> 249 void 250 __qsb_local_sort_with_helping(_QSBThreadLocal<_RAIter>** __tls, 251 _Compare& __comp, _ThreadIndex __iam, 252 bool __wait) 253 { 254 typedef std::iterator_traits<_RAIter> _TraitsType; 255 typedef typename _TraitsType::value_type _ValueType; 256 typedef typename _TraitsType::difference_type _DifferenceType; 257 typedef std::pair<_RAIter, _RAIter> _Piece; 258 259 _QSBThreadLocal<_RAIter>& __tl = *__tls[__iam]; 260 261 _DifferenceType 262 __base_case_n = _Settings::get().sort_qsb_base_case_maximal_n; 263 if (__base_case_n < 2) 264 __base_case_n = 2; 265 _ThreadIndex __num_threads = __tl._M_num_threads; 266 267 // Every thread has its own random number generator. 268 _RandomNumber __rng(__iam + 1); 269 270 _Piece __current = __tl._M_initial; 271 272 _DifferenceType __elements_done = 0; 273 #if _GLIBCXX_PARALLEL_ASSERTIONS 274 _DifferenceType __total_elements_done = 0; 275 #endif 276 277 for (;;) 278 { 279 // Invariant: __current must be a valid (maybe empty) range. 280 _RAIter __begin = __current.first, __end = __current.second; 281 _DifferenceType __n = __end - __begin; 282 283 if (__n > __base_case_n) 284 { 285 // Divide. 286 _RAIter __pivot_pos = __begin + __rng(__n); 287 288 // Swap __pivot_pos value to end. 289 if (__pivot_pos != (__end - 1)) 290 std::iter_swap(__pivot_pos, __end - 1); 291 __pivot_pos = __end - 1; 292 293 __gnu_parallel::__binder2nd 294 <_Compare, _ValueType, _ValueType, bool> 295 __pred(__comp, *__pivot_pos); 296 297 // Divide, leave pivot unchanged in last place. 298 _RAIter __split_pos1, __split_pos2; 299 __split_pos1 = __gnu_sequential::partition(__begin, __end - 1, 300 __pred); 301 302 // Left side: < __pivot_pos; __right side: >= __pivot_pos. 303 #if _GLIBCXX_PARALLEL_ASSERTIONS 304 _GLIBCXX_PARALLEL_ASSERT(__begin <= __split_pos1 305 && __split_pos1 < __end); 306 #endif 307 // Swap pivot back to middle. 308 if (__split_pos1 != __pivot_pos) 309 std::iter_swap(__split_pos1, __pivot_pos); 310 __pivot_pos = __split_pos1; 311 312 // In case all elements are equal, __split_pos1 == 0. 313 if ((__split_pos1 + 1 - __begin) < (__n >> 7) 314 || (__end - __split_pos1) < (__n >> 7)) 315 { 316 // Very unequal split, one part smaller than one 128th 317 // elements not strictly larger than the pivot. 318 __gnu_parallel::__unary_negate<__gnu_parallel::__binder1st 319 <_Compare, _ValueType, _ValueType, bool>, _ValueType> 320 __pred(__gnu_parallel::__binder1st 321 <_Compare, _ValueType, _ValueType, bool> 322 (__comp, *__pivot_pos)); 323 324 // Find other end of pivot-equal range. 325 __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1, 326 __end, __pred); 327 } 328 else 329 // Only skip the pivot. 330 __split_pos2 = __split_pos1 + 1; 331 332 // Elements equal to pivot are done. 333 __elements_done += (__split_pos2 - __split_pos1); 334 #if _GLIBCXX_PARALLEL_ASSERTIONS 335 __total_elements_done += (__split_pos2 - __split_pos1); 336 #endif 337 // Always push larger part onto stack. 338 if (((__split_pos1 + 1) - __begin) < (__end - (__split_pos2))) 339 { 340 // Right side larger. 341 if ((__split_pos2) != __end) 342 __tl._M_leftover_parts.push_front 343 (std::make_pair(__split_pos2, __end)); 344 345 //__current.first = __begin; //already set anyway 346 __current.second = __split_pos1; 347 continue; 348 } 349 else 350 { 351 // Left side larger. 352 if (__begin != __split_pos1) 353 __tl._M_leftover_parts.push_front(std::make_pair 354 (__begin, __split_pos1)); 355 356 __current.first = __split_pos2; 357 //__current.second = __end; //already set anyway 358 continue; 359 } 360 } 361 else 362 { 363 __gnu_sequential::sort(__begin, __end, __comp); 364 __elements_done += __n; 365 #if _GLIBCXX_PARALLEL_ASSERTIONS 366 __total_elements_done += __n; 367 #endif 368 369 // Prefer own stack, small pieces. 370 if (__tl._M_leftover_parts.pop_front(__current)) 371 continue; 372 373 # pragma omp atomic 374 *__tl._M_elements_leftover -= __elements_done; 375 376 __elements_done = 0; 377 378 #if _GLIBCXX_PARALLEL_ASSERTIONS 379 double __search_start = omp_get_wtime(); 380 #endif 381 382 // Look for new work. 383 bool __successfully_stolen = false; 384 while (__wait && *__tl._M_elements_leftover > 0 385 && !__successfully_stolen 386 #if _GLIBCXX_PARALLEL_ASSERTIONS 387 // Possible dead-lock. 388 && (omp_get_wtime() < (__search_start + 1.0)) 389 #endif 390 ) 391 { 392 _ThreadIndex __victim; 393 __victim = __rng(__num_threads); 394 395 // Large pieces. 396 __successfully_stolen = (__victim != __iam) 397 && __tls[__victim]->_M_leftover_parts.pop_back(__current); 398 if (!__successfully_stolen) 399 __yield(); 400 #if !defined(__ICC) && !defined(__ECC) 401 # pragma omp flush 402 #endif 403 } 404 405 #if _GLIBCXX_PARALLEL_ASSERTIONS 406 if (omp_get_wtime() >= (__search_start + 1.0)) 407 { 408 sleep(1); 409 _GLIBCXX_PARALLEL_ASSERT(omp_get_wtime() 410 < (__search_start + 1.0)); 411 } 412 #endif 413 if (!__successfully_stolen) 414 { 415 #if _GLIBCXX_PARALLEL_ASSERTIONS 416 _GLIBCXX_PARALLEL_ASSERT(*__tl._M_elements_leftover == 0); 417 #endif 418 return; 419 } 420 } 421 } 422 } 423 424 /** @brief Top-level quicksort routine. 425 * @param __begin Begin iterator of sequence. 426 * @param __end End iterator of sequence. 427 * @param __comp Comparator. 428 * @param __num_threads Number of threads that are allowed to work on 429 * this part. 430 */ 431 template<typename _RAIter, typename _Compare> 432 void 433 __parallel_sort_qsb(_RAIter __begin, _RAIter __end, 434 _Compare __comp, _ThreadIndex __num_threads) 435 { 436 _GLIBCXX_CALL(__end - __begin) 437 438 typedef std::iterator_traits<_RAIter> _TraitsType; 439 typedef typename _TraitsType::value_type _ValueType; 440 typedef typename _TraitsType::difference_type _DifferenceType; 441 typedef std::pair<_RAIter, _RAIter> _Piece; 442 443 typedef _QSBThreadLocal<_RAIter> _TLSType; 444 445 _DifferenceType __n = __end - __begin; 446 447 if (__n <= 1) 448 return; 449 450 // At least one element per processor. 451 if (__num_threads > __n) 452 __num_threads = static_cast<_ThreadIndex>(__n); 453 454 // Initialize thread local storage 455 _TLSType** __tls = new _TLSType*[__num_threads]; 456 _DifferenceType __queue_size = (__num_threads 457 * (_ThreadIndex)(__rd_log2(__n) + 1)); 458 for (_ThreadIndex __t = 0; __t < __num_threads; ++__t) 459 __tls[__t] = new _QSBThreadLocal<_RAIter>(__queue_size); 460 461 // There can never be more than ceil(__rd_log2(__n)) ranges on the 462 // stack, because 463 // 1. Only one processor pushes onto the stack 464 // 2. The largest range has at most length __n 465 // 3. Each range is larger than half of the range remaining 466 volatile _DifferenceType __elements_leftover = __n; 467 for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) 468 { 469 __tls[__i]->_M_elements_leftover = &__elements_leftover; 470 __tls[__i]->_M_num_threads = __num_threads; 471 __tls[__i]->_M_global = std::make_pair(__begin, __end); 472 473 // Just in case nothing is left to assign. 474 __tls[__i]->_M_initial = std::make_pair(__end, __end); 475 } 476 477 // Main recursion call. 478 __qsb_conquer(__tls, __begin, __begin + __n, __comp, 0, 479 __num_threads, true); 480 481 #if _GLIBCXX_PARALLEL_ASSERTIONS 482 // All stack must be empty. 483 _Piece __dummy; 484 for (_ThreadIndex __i = 1; __i < __num_threads; ++__i) 485 _GLIBCXX_PARALLEL_ASSERT( 486 !__tls[__i]->_M_leftover_parts.pop_back(__dummy)); 487 #endif 488 489 for (_ThreadIndex __i = 0; __i < __num_threads; ++__i) 490 delete __tls[__i]; 491 delete[] __tls; 492 } 493 } // namespace __gnu_parallel 494 495 #endif /* _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H */ 496