libstdc++-v3/include/parallel/balanced_quicksort.h

   1 // -*- C++ -*-
   2
   3 // Copyright (C) 2007, 2008, 2009 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
  48 #include <parallel/settings.h>
  49 #include <parallel/partition.h>
  50 #include <parallel/random_number.h>
  51 #include <parallel/queue.h>
  52 #include <functional>
  53
  54 #if _GLIBCXX_ASSERTIONS
  55 #include <parallel/checkers.h>
  56 #endif
  57
  58 namespace __gnu_parallel
  59 {
  60 /** @brief Information local to one thread in the parallel quicksort run. */
  61 template<typename _RAIter>
  62   struct _QSBThreadLocal
  63   {
  64     typedef std::iterator_traits<_RAIter> _TraitsType;
  65     typedef typename _TraitsType::difference_type _DifferenceType;
  66
  67     /** @brief Continuous part of the sequence, described by an
  68     iterator pair. */
  69     typedef std::pair<_RAIter, _RAIter> _Piece;
  70
  71     /** @brief Initial piece to work on. */
  72     _Piece _M_initial;
  73
  74     /** @brief Work-stealing queue. */
  75     _RestrictedBoundedConcurrentQueue<_Piece> _M_leftover_parts;
  76
  77     /** @brief Number of threads involved in this algorithm. */
  78     _ThreadIndex _M_num_threads;
  79
  80     /** @brief Pointer to a counter of elements left over to sort. */
  81     volatile _DifferenceType* _M_elements_leftover;
  82
  83     /** @brief The complete sequence to sort. */
  84     _Piece _M_global;
  85
  86     /** @brief Constructor.
  87      *  @param __queue_size size of the work-stealing queue. */
  88     _QSBThreadLocal(int __queue_size) : _M_leftover_parts(__queue_size) { }
  89   };
  90
  91 /** @brief Balanced quicksort divide step.
  92   *  @param __begin Begin iterator of subsequence.
  93   *  @param __end End iterator of subsequence.
  94   *  @param __comp Comparator.
  95   *  @param __num_threads Number of threads that are allowed to work on
  96   *  this part.
  97   *  @pre @__c (__end-__begin)>=1 */
  98 template<typename _RAIter, typename _Compare>
  99   typename std::iterator_traits<_RAIter>::difference_type
 100   __qsb_divide(_RAIter __begin, _RAIter __end,
 101              _Compare __comp, _ThreadIndex __num_threads)
 102   {
 103     _GLIBCXX_PARALLEL_ASSERT(__num_threads > 0);
 104
 105     typedef std::iterator_traits<_RAIter> _TraitsType;
 106     typedef typename _TraitsType::value_type _ValueType;
 107     typedef typename _TraitsType::difference_type _DifferenceType;
 108
 109     _RAIter __pivot_pos =
 110       __median_of_three_iterators(__begin, __begin + (__end - __begin) / 2,
 111                                 __end  - 1, __comp);
 112
 113 #if defined(_GLIBCXX_ASSERTIONS)
 114     // Must be in between somewhere.
 115     _DifferenceType __n = __end - __begin;
 116
 117     _GLIBCXX_PARALLEL_ASSERT(
 118            (!__comp(*__pivot_pos, *__begin) &&
 119               !__comp(*(__begin + __n / 2), *__pivot_pos))
 120         || (!__comp(*__pivot_pos, *__begin) &&
 121               !__comp(*(__end - 1), *__pivot_pos))
 122         || (!__comp(*__pivot_pos, *(__begin + __n / 2)) &&
 123               !__comp(*__begin, *__pivot_pos))
 124         || (!__comp(*__pivot_pos, *(__begin + __n / 2)) &&
 125               !__comp(*(__end - 1), *__pivot_pos))
 126         || (!__comp(*__pivot_pos, *(__end - 1)) &&
 127               !__comp(*__begin, *__pivot_pos))
 128         || (!__comp(*__pivot_pos, *(__end - 1)) &&
 129               !__comp(*(__begin + __n / 2), *__pivot_pos)));
 130 #endif
 131
 132     // Swap pivot value to end.
 133     if (__pivot_pos != (__end - 1))
 134       std::swap(*__pivot_pos, *(__end - 1));
 135     __pivot_pos = __end - 1;
 136
 137     __gnu_parallel::binder2nd<_Compare, _ValueType, _ValueType, bool>
 138         __pred(__comp, *__pivot_pos);
 139
 140     // Divide, returning __end - __begin - 1 in the worst case.
 141     _DifferenceType __split_pos = __parallel_partition(
 142         __begin, __end - 1, __pred, __num_threads);
 143
 144     // Swap back pivot to middle.
 145     std::swap(*(__begin + __split_pos), *__pivot_pos);
 146     __pivot_pos = __begin + __split_pos;
 147
 148 #if _GLIBCXX_ASSERTIONS
 149     _RAIter __r;
 150     for (__r = __begin; __r != __pivot_pos; ++__r)
 151       _GLIBCXX_PARALLEL_ASSERT(__comp(*__r, *__pivot_pos));
 152     for (; __r != __end; ++__r)
 153       _GLIBCXX_PARALLEL_ASSERT(!__comp(*__r, *__pivot_pos));
 154 #endif
 155
 156     return __split_pos;
 157   }
 158
 159 /** @brief Quicksort conquer step.
 160   *  @param __tls Array of thread-local storages.
 161   *  @param __begin Begin iterator of subsequence.
 162   *  @param __end End iterator of subsequence.
 163   *  @param __comp Comparator.
 164   *  @param __iam Number of the thread processing this function.
 165   *  @param __num_threads
 166   *          Number of threads that are allowed to work on this part. */
 167 template<typename _RAIter, typename _Compare>
 168   void
 169   __qsb_conquer(_QSBThreadLocal<_RAIter>** __tls,
 170               _RAIter __begin, _RAIter __end,
 171               _Compare __comp,
 172               _ThreadIndex __iam, _ThreadIndex __num_threads,
 173               bool __parent_wait)
 174   {
 175     typedef std::iterator_traits<_RAIter> _TraitsType;
 176     typedef typename _TraitsType::value_type _ValueType;
 177     typedef typename _TraitsType::difference_type _DifferenceType;
 178
 179     _DifferenceType __n = __end - __begin;
 180
 181     if (__num_threads <= 1 || __n <= 1)
 182       {
 183         __tls[__iam]->_M_initial.first  = __begin;
 184         __tls[__iam]->_M_initial.second = __end;
 185
 186         __qsb_local_sort_with_helping(__tls, __comp, __iam, __parent_wait);
 187
 188         return;
 189       }
 190
 191     // Divide step.
 192     _DifferenceType __split_pos =
 193             __qsb_divide(__begin, __end, __comp, __num_threads);
 194
 195 #if _GLIBCXX_ASSERTIONS
 196     _GLIBCXX_PARALLEL_ASSERT(0 <= __split_pos &&
 197                              __split_pos < (__end - __begin));
 198 #endif
 199
 200     _ThreadIndex __num_threads_leftside =
 201         std::max<_ThreadIndex>(1, std::min<_ThreadIndex>(
 202             __num_threads - 1, __split_pos * __num_threads / __n));
 203
 204 #   pragma omp atomic
 205     *__tls[__iam]->_M_elements_leftover -= (_DifferenceType)1;
 206
 207     // Conquer step.
 208 #   pragma omp parallel num_threads(2)
 209     {
 210       bool __wait;
 211       if(omp_get_num_threads() < 2)
 212         __wait = false;
 213       else
 214         __wait = __parent_wait;
 215
 216 #     pragma omp sections
 217         {
 218 #         pragma omp section
 219             {
 220               __qsb_conquer(__tls, __begin, __begin + __split_pos, __comp,
 221                           __iam,
 222                           __num_threads_leftside,
 223                           __wait);
 224               __wait = __parent_wait;
 225             }
 226           // The pivot_pos is left in place, to ensure termination.
 227 #         pragma omp section
 228             {
 229               __qsb_conquer(__tls, __begin + __split_pos + 1, __end, __comp,
 230                           __iam + __num_threads_leftside,
 231                           __num_threads - __num_threads_leftside,
 232                           __wait);
 233               __wait = __parent_wait;
 234             }
 235         }
 236     }
 237   }
 238
 239 /**
 240   *  @brief Quicksort step doing load-balanced local sort.
 241   *  @param __tls Array of thread-local storages.
 242   *  @param __comp Comparator.
 243   *  @param __iam Number of the thread processing this function.
 244   */
 245 template<typename _RAIter, typename _Compare>
 246   void
 247   __qsb_local_sort_with_helping(_QSBThreadLocal<_RAIter>** __tls,
 248                               _Compare& __comp, int __iam, bool __wait)
 249   {
 250     typedef std::iterator_traits<_RAIter> _TraitsType;
 251     typedef typename _TraitsType::value_type _ValueType;
 252     typedef typename _TraitsType::difference_type _DifferenceType;
 253     typedef std::pair<_RAIter, _RAIter> _Piece;
 254
 255     _QSBThreadLocal<_RAIter>& __tl = *__tls[__iam];
 256
 257     _DifferenceType __base_case_n =
 258         _Settings::get().sort_qsb_base_case_maximal_n;
 259     if (__base_case_n < 2)
 260       __base_case_n = 2;
 261     _ThreadIndex __num_threads = __tl._M_num_threads;
 262
 263     // Every thread has its own random number generator.
 264     _RandomNumber __rng(__iam + 1);
 265
 266     _Piece __current = __tl._M_initial;
 267
 268     _DifferenceType __elements_done = 0;
 269 #if _GLIBCXX_ASSERTIONS
 270     _DifferenceType __total_elements_done = 0;
 271 #endif
 272
 273     for (;;)
 274       {
 275         // Invariant: __current must be a valid (maybe empty) range.
 276         _RAIter __begin = __current.first, __end = __current.second;
 277         _DifferenceType __n = __end - __begin;
 278
 279         if (__n > __base_case_n)
 280           {
 281             // Divide.
 282             _RAIter __pivot_pos = __begin +  __rng(__n);
 283
 284             // Swap __pivot_pos value to end.
 285             if (__pivot_pos != (__end - 1))
 286               std::swap(*__pivot_pos, *(__end - 1));
 287             __pivot_pos = __end - 1;
 288
 289             __gnu_parallel::binder2nd
 290                 <_Compare, _ValueType, _ValueType, bool>
 291                 __pred(__comp, *__pivot_pos);
 292
 293             // Divide, leave pivot unchanged in last place.
 294             _RAIter __split_pos1, __split_pos2;
 295             __split_pos1 =
 296                     __gnu_sequential::partition(__begin, __end - 1, __pred);
 297
 298             // Left side: < __pivot_pos; __right side: >= __pivot_pos.
 299 #if _GLIBCXX_ASSERTIONS
 300             _GLIBCXX_PARALLEL_ASSERT(__begin <= __split_pos1
 301                                      && __split_pos1 < __end);
 302 #endif
 303             // Swap pivot back to middle.
 304             if (__split_pos1 != __pivot_pos)
 305               std::swap(*__split_pos1, *__pivot_pos);
 306             __pivot_pos = __split_pos1;
 307
 308             // In case all elements are equal, __split_pos1 == 0.
 309             if ((__split_pos1 + 1 - __begin) < (__n >> 7)
 310             || (__end - __split_pos1) < (__n >> 7))
 311               {
 312                 // Very unequal split, one part smaller than one 128th
 313                 // elements not strictly larger than the pivot.
 314                 __gnu_parallel::__unary_negate<__gnu_parallel::__binder1st
 315                   <_Compare, _ValueType, _ValueType, bool>, _ValueType>
 316                   __pred(__gnu_parallel::__binder1st
 317                        <_Compare, _ValueType, _ValueType, bool>(
 318                           __comp, *__pivot_pos));
 319
 320                 // Find other end of pivot-equal range.
 321                 __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1,
 322                                                          __end, __pred);
 323               }
 324             else
 325               // Only skip the pivot.
 326               __split_pos2 = __split_pos1 + 1;
 327
 328             // Elements equal to pivot are done.
 329             __elements_done += (__split_pos2 - __split_pos1);
 330 #if _GLIBCXX_ASSERTIONS
 331             __total_elements_done += (__split_pos2 - __split_pos1);
 332 #endif
 333             // Always push larger part onto stack.
 334             if (((__split_pos1 + 1) - __begin) < (__end - (__split_pos2)))
 335               {
 336                 // Right side larger.
 337                 if ((__split_pos2) != __end)
 338                   __tl._M_leftover_parts.push_front(
 339                     std::make_pair(__split_pos2, __end));
 340
 341                 //__current.first = __begin;    //already set anyway
 342                 __current.second = __split_pos1;
 343                 continue;
 344               }
 345             else
 346               {
 347                 // Left side larger.
 348                 if (__begin != __split_pos1)
 349                   __tl._M_leftover_parts.push_front(std::make_pair(__begin,
 350                                                               __split_pos1));
 351
 352                 __current.first = __split_pos2;
 353                 //__current.second = __end;     //already set anyway
 354                 continue;
 355               }
 356           }
 357         else
 358           {
 359             __gnu_sequential::sort(__begin, __end, __comp);
 360             __elements_done += __n;
 361 #if _GLIBCXX_ASSERTIONS
 362             __total_elements_done += __n;
 363 #endif
 364
 365             // Prefer own stack, small pieces.
 366             if (__tl._M_leftover_parts.pop_front(__current))
 367               continue;
 368
 369 #           pragma omp atomic
 370             *__tl._M_elements_leftover -= __elements_done;
 371
 372             __elements_done = 0;
 373
 374 #if _GLIBCXX_ASSERTIONS
 375             double __search_start = omp_get_wtime();
 376 #endif
 377
 378             // Look for new work.
 379             bool __successfully_stolen = false;
 380             while (__wait && *__tl._M_elements_leftover > 0
 381                    && !__successfully_stolen
 382 #if _GLIBCXX_ASSERTIONS
 383                     // Possible dead-lock.
 384                    && (omp_get_wtime() < (__search_start + 1.0))
 385 #endif
 386               )
 387               {
 388                 _ThreadIndex __victim;
 389                 __victim = __rng(__num_threads);
 390
 391                 // Large pieces.
 392                 __successfully_stolen = (__victim != __iam)
 393                     && __tls[__victim]->_M_leftover_parts.pop_back(__current);
 394                 if (!__successfully_stolen)
 395                   __yield();
 396 #if !defined(__ICC) && !defined(__ECC)
 397 #               pragma omp flush
 398 #endif
 399               }
 400
 401 #if _GLIBCXX_ASSERTIONS
 402             if (omp_get_wtime() >= (__search_start + 1.0))
 403               {
 404                 sleep(1);
 405                 _GLIBCXX_PARALLEL_ASSERT(omp_get_wtime()
 406                                          < (__search_start + 1.0));
 407               }
 408 #endif
 409             if (!__successfully_stolen)
 410               {
 411 #if _GLIBCXX_ASSERTIONS
 412                 _GLIBCXX_PARALLEL_ASSERT(*__tl._M_elements_leftover == 0);
 413 #endif
 414                 return;
 415               }
 416           }
 417       }
 418   }
 419
 420 /** @brief Top-level quicksort routine.
 421   *  @param __begin Begin iterator of sequence.
 422   *  @param __end End iterator of sequence.
 423   *  @param __comp Comparator.
 424   *  @param __num_threads Number of threads that are allowed to work on
 425   *  this part.
 426   */
 427 template<typename _RAIter, typename _Compare>
 428   void
 429   __parallel_sort_qsb(_RAIter __begin, _RAIter __end,
 430                     _Compare __comp,
 431                     _ThreadIndex __num_threads)
 432   {
 433     _GLIBCXX_CALL(__end - __begin)
 434
 435     typedef std::iterator_traits<_RAIter> _TraitsType;
 436     typedef typename _TraitsType::value_type _ValueType;
 437     typedef typename _TraitsType::difference_type _DifferenceType;
 438     typedef std::pair<_RAIter, _RAIter> _Piece;
 439
 440     typedef _QSBThreadLocal<_RAIter> _TLSType;
 441
 442     _DifferenceType __n = __end - __begin;
 443
 444     if (__n <= 1)
 445       return;
 446
 447     // At least one element per processor.
 448     if (__num_threads > __n)
 449       __num_threads = static_cast<_ThreadIndex>(__n);
 450
 451     // Initialize thread local storage
 452     _TLSType** __tls = new _TLSType*[__num_threads];
 453     _DifferenceType __queue_size =
 454                            __num_threads * (_ThreadIndex)(log2(__n) + 1);
 455     for (_ThreadIndex __t = 0; __t < __num_threads; ++__t)
 456       __tls[__t] = new _QSBThreadLocal<_RAIter>(__queue_size);
 457
 458     // There can never be more than ceil(log2(__n)) ranges on the stack,
 459     // because
 460     // 1. Only one processor pushes onto the stack
 461     // 2. The largest range has at most length __n
 462     // 3. Each range is larger than half of the range remaining
 463     volatile _DifferenceType _M_elements_leftover = __n;
 464     for (int __i = 0; __i < __num_threads; ++__i)
 465       {
 466         __tls[__i]->_M_elements_leftover = &_M_elements_leftover;
 467         __tls[__i]->_M_num_threads = __num_threads;
 468         __tls[__i]->_M_global = std::make_pair(__begin, __end);
 469
 470         // Just in case nothing is left to assign.
 471         __tls[__i]->_M_initial = std::make_pair(__end, __end);
 472       }
 473
 474     // Main recursion call.
 475     __qsb_conquer(
 476       __tls, __begin, __begin + __n, __comp, 0, __num_threads, true);
 477
 478 #if _GLIBCXX_ASSERTIONS
 479     // All stack must be empty.
 480     _Piece __dummy;
 481     for (int __i = 1; __i < __num_threads; ++__i)
 482       _GLIBCXX_PARALLEL_ASSERT(
 483         !__tls[__i]->_M_leftover_parts.pop_back(__dummy));
 484 #endif
 485
 486     for (int __i = 0; __i < __num_threads; ++__i)
 487       delete __tls[__i];
 488     delete[] __tls;
 489   }
 490 } // namespace __gnu_parallel
 491
 492 #endif /* _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H */