balanced_quicksort.h

Go to the documentation of this file.
00001 // -*- C++ -*-
00002 
00003 // Copyright (C) 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
00004 //
00005 // This file is part of the GNU ISO C++ Library.  This library is free
00006 // software; you can redistribute it and/or modify it under the terms
00007 // of the GNU General Public License as published by the Free Software
00008 // Foundation; either version 3, or (at your option) any later
00009 // version.
00010 
00011 // This library is distributed in the hope that it will be useful, but
00012 // WITHOUT ANY WARRANTY; without even the implied warranty of
00013 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00014 // General Public License for more details.
00015 
00016 // Under Section 7 of GPL version 3, you are granted additional
00017 // permissions described in the GCC Runtime Library Exception, version
00018 // 3.1, as published by the Free Software Foundation.
00019 
00020 // You should have received a copy of the GNU General Public License and
00021 // a copy of the GCC Runtime Library Exception along with this program;
00022 // see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
00023 // <http://www.gnu.org/licenses/>.
00024 
00025 /** @file parallel/balanced_quicksort.h
00026  *  @brief Implementation of a dynamically load-balanced parallel quicksort.
00027  *
00028  *  It works in-place and needs only logarithmic extra memory.
00029  *  The algorithm is similar to the one proposed in
00030  *
00031  *  P. Tsigas and Y. Zhang.
00032  *  A simple, fast parallel implementation of quicksort and
00033  *  its performance evaluation on SUN enterprise 10000.
00034  *  In 11th Euromicro Conference on Parallel, Distributed and
00035  *  Network-Based Processing, page 372, 2003.
00036  *
00037  *  This file is a GNU parallel extension to the Standard C++ Library.
00038  */
00039 
00040 // Written by Johannes Singler.
00041 
00042 #ifndef _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H
00043 #define _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H 1
00044 
00045 #include <parallel/basic_iterator.h>
00046 #include <bits/stl_algo.h>
00047 #include <bits/stl_function.h>
00048 
00049 #include <parallel/settings.h>
00050 #include <parallel/partition.h>
00051 #include <parallel/random_number.h>
00052 #include <parallel/queue.h>
00053 
00054 #if _GLIBCXX_ASSERTIONS
00055 #include <parallel/checkers.h>
00056 #endif
00057 
00058 namespace __gnu_parallel
00059 {
00060   /** @brief Information local to one thread in the parallel quicksort run. */
00061   template<typename _RAIter>
00062     struct _QSBThreadLocal
00063     {
00064       typedef std::iterator_traits<_RAIter> _TraitsType;
00065       typedef typename _TraitsType::difference_type _DifferenceType;
00066 
00067       /** @brief Continuous part of the sequence, described by an
00068       iterator pair. */
00069       typedef std::pair<_RAIter, _RAIter> _Piece;
00070 
00071       /** @brief Initial piece to work on. */
00072       _Piece _M_initial;
00073 
00074       /** @brief Work-stealing queue. */
00075       _RestrictedBoundedConcurrentQueue<_Piece> _M_leftover_parts;
00076 
00077       /** @brief Number of threads involved in this algorithm. */
00078       _ThreadIndex _M_num_threads;
00079 
00080       /** @brief Pointer to a counter of elements left over to sort. */
00081       volatile _DifferenceType* _M_elements_leftover;
00082 
00083       /** @brief The complete sequence to sort. */
00084       _Piece _M_global;
00085 
00086       /** @brief Constructor.
00087        *  @param __queue_size size of the work-stealing queue. */
00088       _QSBThreadLocal(int __queue_size) : _M_leftover_parts(__queue_size) { }
00089     };
00090 
00091   /** @brief Balanced quicksort divide step.
00092     *  @param __begin Begin iterator of subsequence.
00093     *  @param __end End iterator of subsequence.
00094     *  @param __comp Comparator.
00095     *  @param __num_threads Number of threads that are allowed to work on
00096     *  this part.
00097     *  @pre @c (__end-__begin)>=1 */
00098   template<typename _RAIter, typename _Compare>
00099     typename std::iterator_traits<_RAIter>::difference_type
00100     __qsb_divide(_RAIter __begin, _RAIter __end,
00101          _Compare __comp, _ThreadIndex __num_threads)
00102     {
00103       _GLIBCXX_PARALLEL_ASSERT(__num_threads > 0);
00104 
00105       typedef std::iterator_traits<_RAIter> _TraitsType;
00106       typedef typename _TraitsType::value_type _ValueType;
00107       typedef typename _TraitsType::difference_type _DifferenceType;
00108 
00109       _RAIter __pivot_pos =
00110     __median_of_three_iterators(__begin, __begin + (__end - __begin) / 2,
00111                     __end  - 1, __comp);
00112 
00113 #if defined(_GLIBCXX_ASSERTIONS)
00114       // Must be in between somewhere.
00115       _DifferenceType __n = __end - __begin;
00116 
00117       _GLIBCXX_PARALLEL_ASSERT((!__comp(*__pivot_pos, *__begin)
00118                 && !__comp(*(__begin + __n / 2),
00119                        *__pivot_pos))
00120                    || (!__comp(*__pivot_pos, *__begin)
00121                    && !__comp(*(__end - 1), *__pivot_pos))
00122                    || (!__comp(*__pivot_pos, *(__begin + __n / 2))
00123                    && !__comp(*__begin, *__pivot_pos))
00124                    || (!__comp(*__pivot_pos, *(__begin + __n / 2))
00125                    && !__comp(*(__end - 1), *__pivot_pos))
00126                    || (!__comp(*__pivot_pos, *(__end - 1))
00127                    && !__comp(*__begin, *__pivot_pos))
00128                    || (!__comp(*__pivot_pos, *(__end - 1))
00129                    && !__comp(*(__begin + __n / 2),
00130                           *__pivot_pos)));
00131 #endif
00132 
00133       // Swap pivot value to end.
00134       if (__pivot_pos != (__end - 1))
00135     std::iter_swap(__pivot_pos, __end - 1);
00136       __pivot_pos = __end - 1;
00137 
00138       __gnu_parallel::__binder2nd<_Compare, _ValueType, _ValueType, bool>
00139     __pred(__comp, *__pivot_pos);
00140 
00141       // Divide, returning __end - __begin - 1 in the worst case.
00142       _DifferenceType __split_pos = __parallel_partition(__begin, __end - 1,
00143                              __pred,
00144                              __num_threads);
00145 
00146       // Swap back pivot to middle.
00147       std::iter_swap(__begin + __split_pos, __pivot_pos);
00148       __pivot_pos = __begin + __split_pos;
00149 
00150 #if _GLIBCXX_ASSERTIONS
00151       _RAIter __r;
00152       for (__r = __begin; __r != __pivot_pos; ++__r)
00153     _GLIBCXX_PARALLEL_ASSERT(__comp(*__r, *__pivot_pos));
00154       for (; __r != __end; ++__r)
00155     _GLIBCXX_PARALLEL_ASSERT(!__comp(*__r, *__pivot_pos));
00156 #endif
00157 
00158       return __split_pos;
00159     }
00160 
00161   /** @brief Quicksort conquer step.
00162     *  @param __tls Array of thread-local storages.
00163     *  @param __begin Begin iterator of subsequence.
00164     *  @param __end End iterator of subsequence.
00165     *  @param __comp Comparator.
00166     *  @param __iam Number of the thread processing this function.
00167     *  @param __num_threads
00168     *          Number of threads that are allowed to work on this part. */
00169   template<typename _RAIter, typename _Compare>
00170     void
00171     __qsb_conquer(_QSBThreadLocal<_RAIter>** __tls,
00172           _RAIter __begin, _RAIter __end,
00173           _Compare __comp,
00174           _ThreadIndex __iam, _ThreadIndex __num_threads,
00175           bool __parent_wait)
00176     {
00177       typedef std::iterator_traits<_RAIter> _TraitsType;
00178       typedef typename _TraitsType::value_type _ValueType;
00179       typedef typename _TraitsType::difference_type _DifferenceType;
00180 
00181       _DifferenceType __n = __end - __begin;
00182 
00183       if (__num_threads <= 1 || __n <= 1)
00184     {
00185           __tls[__iam]->_M_initial.first  = __begin;
00186           __tls[__iam]->_M_initial.second = __end;
00187 
00188           __qsb_local_sort_with_helping(__tls, __comp, __iam, __parent_wait);
00189 
00190           return;
00191     }
00192 
00193       // Divide step.
00194       _DifferenceType __split_pos =
00195     __qsb_divide(__begin, __end, __comp, __num_threads);
00196 
00197 #if _GLIBCXX_ASSERTIONS
00198       _GLIBCXX_PARALLEL_ASSERT(0 <= __split_pos &&
00199                                __split_pos < (__end - __begin));
00200 #endif
00201 
00202       _ThreadIndex
00203     __num_threads_leftside = std::max<_ThreadIndex>
00204     (1, std::min<_ThreadIndex>(__num_threads - 1, __split_pos
00205                    * __num_threads / __n));
00206 
00207 #     pragma omp atomic
00208       *__tls[__iam]->_M_elements_leftover -= (_DifferenceType)1;
00209 
00210       // Conquer step.
00211 #     pragma omp parallel num_threads(2)
00212       {
00213     bool __wait;
00214     if(omp_get_num_threads() < 2)
00215           __wait = false;
00216     else
00217           __wait = __parent_wait;
00218 
00219 #       pragma omp sections
00220     {
00221 #         pragma omp section
00222       {
00223         __qsb_conquer(__tls, __begin, __begin + __split_pos, __comp,
00224               __iam, __num_threads_leftside, __wait);
00225         __wait = __parent_wait;
00226       }
00227       // The pivot_pos is left in place, to ensure termination.
00228 #         pragma omp section
00229       {
00230         __qsb_conquer(__tls, __begin + __split_pos + 1, __end, __comp,
00231               __iam + __num_threads_leftside,
00232               __num_threads - __num_threads_leftside, __wait);
00233         __wait = __parent_wait;
00234       }
00235     }
00236       }
00237     }
00238 
00239   /**
00240     *  @brief Quicksort step doing load-balanced local sort.
00241     *  @param __tls Array of thread-local storages.
00242     *  @param __comp Comparator.
00243     *  @param __iam Number of the thread processing this function.
00244     */
00245   template<typename _RAIter, typename _Compare>
00246     void
00247     __qsb_local_sort_with_helping(_QSBThreadLocal<_RAIter>** __tls,
00248                   _Compare& __comp, _ThreadIndex __iam,
00249                   bool __wait)
00250     {
00251       typedef std::iterator_traits<_RAIter> _TraitsType;
00252       typedef typename _TraitsType::value_type _ValueType;
00253       typedef typename _TraitsType::difference_type _DifferenceType;
00254       typedef std::pair<_RAIter, _RAIter> _Piece;
00255 
00256       _QSBThreadLocal<_RAIter>& __tl = *__tls[__iam];
00257 
00258       _DifferenceType
00259     __base_case_n = _Settings::get().sort_qsb_base_case_maximal_n;
00260       if (__base_case_n < 2)
00261     __base_case_n = 2;
00262       _ThreadIndex __num_threads = __tl._M_num_threads;
00263 
00264       // Every thread has its own random number generator.
00265       _RandomNumber __rng(__iam + 1);
00266 
00267       _Piece __current = __tl._M_initial;
00268 
00269       _DifferenceType __elements_done = 0;
00270 #if _GLIBCXX_ASSERTIONS
00271       _DifferenceType __total_elements_done = 0;
00272 #endif
00273 
00274       for (;;)
00275     {
00276           // Invariant: __current must be a valid (maybe empty) range.
00277           _RAIter __begin = __current.first, __end = __current.second;
00278           _DifferenceType __n = __end - __begin;
00279 
00280           if (__n > __base_case_n)
00281             {
00282               // Divide.
00283               _RAIter __pivot_pos = __begin +  __rng(__n);
00284 
00285               // Swap __pivot_pos value to end.
00286               if (__pivot_pos != (__end - 1))
00287             std::iter_swap(__pivot_pos, __end - 1);
00288               __pivot_pos = __end - 1;
00289 
00290               __gnu_parallel::__binder2nd
00291         <_Compare, _ValueType, _ValueType, bool>
00292         __pred(__comp, *__pivot_pos);
00293 
00294               // Divide, leave pivot unchanged in last place.
00295               _RAIter __split_pos1, __split_pos2;
00296               __split_pos1 = __gnu_sequential::partition(__begin, __end - 1,
00297                              __pred);
00298 
00299               // Left side: < __pivot_pos; __right side: >= __pivot_pos.
00300 #if _GLIBCXX_ASSERTIONS
00301               _GLIBCXX_PARALLEL_ASSERT(__begin <= __split_pos1
00302                                        && __split_pos1 < __end);
00303 #endif
00304               // Swap pivot back to middle.
00305               if (__split_pos1 != __pivot_pos)
00306             std::iter_swap(__split_pos1, __pivot_pos);
00307               __pivot_pos = __split_pos1;
00308 
00309               // In case all elements are equal, __split_pos1 == 0.
00310               if ((__split_pos1 + 1 - __begin) < (__n >> 7)
00311           || (__end - __split_pos1) < (__n >> 7))
00312             {
00313                   // Very unequal split, one part smaller than one 128th
00314                   // elements not strictly larger than the pivot.
00315                   __gnu_parallel::__unary_negate<__gnu_parallel::__binder1st
00316                     <_Compare, _ValueType, _ValueType, bool>, _ValueType>
00317                     __pred(__gnu_parallel::__binder1st
00318                      <_Compare, _ValueType, _ValueType, bool>
00319                (__comp, *__pivot_pos));
00320 
00321                   // Find other end of pivot-equal range.
00322                   __split_pos2 = __gnu_sequential::partition(__split_pos1 + 1,
00323                                  __end, __pred);
00324             }
00325               else
00326             // Only skip the pivot.
00327             __split_pos2 = __split_pos1 + 1;
00328 
00329               // Elements equal to pivot are done.
00330               __elements_done += (__split_pos2 - __split_pos1);
00331 #if _GLIBCXX_ASSERTIONS
00332               __total_elements_done += (__split_pos2 - __split_pos1);
00333 #endif
00334               // Always push larger part onto stack.
00335               if (((__split_pos1 + 1) - __begin) < (__end - (__split_pos2)))
00336             {
00337                   // Right side larger.
00338                   if ((__split_pos2) != __end)
00339                     __tl._M_leftover_parts.push_front
00340               (std::make_pair(__split_pos2, __end));
00341 
00342                   //__current.first = __begin;    //already set anyway
00343                   __current.second = __split_pos1;
00344                   continue;
00345             }
00346               else
00347             {
00348                   // Left side larger.
00349                   if (__begin != __split_pos1)
00350                     __tl._M_leftover_parts.push_front(std::make_pair
00351                               (__begin, __split_pos1));
00352 
00353                   __current.first = __split_pos2;
00354                   //__current.second = __end;     //already set anyway
00355                   continue;
00356             }
00357             }
00358           else
00359             {
00360               __gnu_sequential::sort(__begin, __end, __comp);
00361               __elements_done += __n;
00362 #if _GLIBCXX_ASSERTIONS
00363               __total_elements_done += __n;
00364 #endif
00365 
00366               // Prefer own stack, small pieces.
00367               if (__tl._M_leftover_parts.pop_front(__current))
00368             continue;
00369 
00370 #             pragma omp atomic
00371               *__tl._M_elements_leftover -= __elements_done;
00372 
00373               __elements_done = 0;
00374 
00375 #if _GLIBCXX_ASSERTIONS
00376               double __search_start = omp_get_wtime();
00377 #endif
00378 
00379               // Look for new work.
00380               bool __successfully_stolen = false;
00381               while (__wait && *__tl._M_elements_leftover > 0
00382                      && !__successfully_stolen
00383 #if _GLIBCXX_ASSERTIONS
00384                       // Possible dead-lock.
00385                      && (omp_get_wtime() < (__search_start + 1.0))
00386 #endif
00387              )
00388             {
00389                   _ThreadIndex __victim;
00390                   __victim = __rng(__num_threads);
00391 
00392                   // Large pieces.
00393                   __successfully_stolen = (__victim != __iam)
00394             && __tls[__victim]->_M_leftover_parts.pop_back(__current);
00395                   if (!__successfully_stolen)
00396                     __yield();
00397 #if !defined(__ICC) && !defined(__ECC)
00398 #                 pragma omp flush
00399 #endif
00400             }
00401 
00402 #if _GLIBCXX_ASSERTIONS
00403               if (omp_get_wtime() >= (__search_start + 1.0))
00404             {
00405                   sleep(1);
00406                   _GLIBCXX_PARALLEL_ASSERT(omp_get_wtime()
00407                                            < (__search_start + 1.0));
00408             }
00409 #endif
00410               if (!__successfully_stolen)
00411             {
00412 #if _GLIBCXX_ASSERTIONS
00413                   _GLIBCXX_PARALLEL_ASSERT(*__tl._M_elements_leftover == 0);
00414 #endif
00415                   return;
00416             }
00417             }
00418     }
00419     }
00420 
00421   /** @brief Top-level quicksort routine.
00422     *  @param __begin Begin iterator of sequence.
00423     *  @param __end End iterator of sequence.
00424     *  @param __comp Comparator.
00425     *  @param __num_threads Number of threads that are allowed to work on
00426     *  this part.
00427     */
00428   template<typename _RAIter, typename _Compare>
00429     void
00430     __parallel_sort_qsb(_RAIter __begin, _RAIter __end,
00431             _Compare __comp, _ThreadIndex __num_threads)
00432     {
00433       _GLIBCXX_CALL(__end - __begin)
00434 
00435       typedef std::iterator_traits<_RAIter> _TraitsType;
00436       typedef typename _TraitsType::value_type _ValueType;
00437       typedef typename _TraitsType::difference_type _DifferenceType;
00438       typedef std::pair<_RAIter, _RAIter> _Piece;
00439 
00440       typedef _QSBThreadLocal<_RAIter> _TLSType;
00441 
00442       _DifferenceType __n = __end - __begin;
00443 
00444       if (__n <= 1)
00445     return;
00446 
00447       // At least one element per processor.
00448       if (__num_threads > __n)
00449     __num_threads = static_cast<_ThreadIndex>(__n);
00450 
00451       // Initialize thread local storage
00452       _TLSType** __tls = new _TLSType*[__num_threads];
00453       _DifferenceType __queue_size = (__num_threads
00454                       * (_ThreadIndex)(__rd_log2(__n) + 1));
00455       for (_ThreadIndex __t = 0; __t < __num_threads; ++__t)
00456     __tls[__t] = new _QSBThreadLocal<_RAIter>(__queue_size);
00457 
00458       // There can never be more than ceil(__rd_log2(__n)) ranges on the
00459       // stack, because
00460       // 1. Only one processor pushes onto the stack
00461       // 2. The largest range has at most length __n
00462       // 3. Each range is larger than half of the range remaining
00463       volatile _DifferenceType __elements_leftover = __n;
00464       for (_ThreadIndex __i = 0; __i < __num_threads; ++__i)
00465     {
00466           __tls[__i]->_M_elements_leftover = &__elements_leftover;
00467           __tls[__i]->_M_num_threads = __num_threads;
00468           __tls[__i]->_M_global = std::make_pair(__begin, __end);
00469 
00470           // Just in case nothing is left to assign.
00471           __tls[__i]->_M_initial = std::make_pair(__end, __end);
00472     }
00473 
00474       // Main recursion call.
00475       __qsb_conquer(__tls, __begin, __begin + __n, __comp, 0,
00476             __num_threads, true);
00477 
00478 #if _GLIBCXX_ASSERTIONS
00479       // All stack must be empty.
00480       _Piece __dummy;
00481       for (_ThreadIndex __i = 1; __i < __num_threads; ++__i)
00482     _GLIBCXX_PARALLEL_ASSERT(
00483           !__tls[__i]->_M_leftover_parts.pop_back(__dummy));
00484 #endif
00485 
00486       for (_ThreadIndex __i = 0; __i < __num_threads; ++__i)
00487     delete __tls[__i];
00488       delete[] __tls;
00489     }
00490 } // namespace __gnu_parallel
00491 
00492 #endif /* _GLIBCXX_PARALLEL_BALANCED_QUICKSORT_H */