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

   1 // -*- C++ -*-
   2
   3 // Copyright (C) 2007 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 2, 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 // You should have received a copy of the GNU General Public License
  17 // along with this library; see the file COPYING.  If not, write to
  18 // the Free Software Foundation, 59 Temple Place - Suite 330, Boston,
  19 // MA 02111-1307, USA.
  20
  21 // As a special exception, you may use this file as part of a free
  22 // software library without restriction.  Specifically, if other files
  23 // instantiate templates or use macros or inline functions from this
  24 // file, or you compile this file and link it with other files to
  25 // produce an executable, this file does not by itself cause the
  26 // resulting executable to be covered by the GNU General Public
  27 // License.  This exception does not however invalidate any other
  28 // reasons why the executable file might be covered by the GNU General
  29 // Public License.
  30
  31 /** @file parallel/multiway_mergesort.h
  32  *  @brief Parallel multiway merge sort.
  33  *  This file is a GNU parallel extension to the Standard C++ Library.
  34  */
  35
  36 // Written by Johannes Singler.
  37
  38 #ifndef _GLIBCXX_PARALLEL_MERGESORT_H
  39 #define _GLIBCXX_PARALLEL_MERGESORT_H 1
  40
  41 #include <vector>
  42
  43 #include <parallel/basic_iterator.h>
  44 #include <bits/stl_algo.h>
  45 #include <parallel/parallel.h>
  46 #include <parallel/multiway_merge.h>
  47
  48 namespace __gnu_parallel
  49 {
  50
  51 /** @brief Subsequence description. */
  52 template<typename _DifferenceTp>
  53   struct Piece
  54   {
  55     typedef _DifferenceTp difference_type;
  56
  57     /** @brief Begin of subsequence. */
  58     difference_type begin;
  59
  60     /** @brief End of subsequence. */
  61     difference_type end;
  62   };
  63
  64 /** @brief Data accessed by all threads.
  65   *
  66   *  PMWMS = parallel multiway mergesort */
  67 template<typename RandomAccessIterator>
  68   struct PMWMSSortingData
  69   {
  70     typedef std::iterator_traits<RandomAccessIterator> traits_type;
  71     typedef typename traits_type::value_type value_type;
  72     typedef typename traits_type::difference_type difference_type;
  73
  74     /** @brief Number of threads involved. */
  75     thread_index_t num_threads;
  76
  77     /** @brief Input begin. */
  78     RandomAccessIterator source;
  79
  80     /** @brief Start indices, per thread. */
  81     difference_type* starts;
  82
  83     /** @brief Temporary arrays for each thread.
  84      *
  85      *  Indirection Allows using the temporary storage in different
  86      *  ways, without code duplication.
  87      *  @see _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST */
  88     value_type** temporaries;
  89
  90 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
  91     /** @brief Storage in which to sort. */
  92     RandomAccessIterator* sorting_places;
  93
  94     /** @brief Storage into which to merge. */
  95     value_type** merging_places;
  96 #else
  97     /** @brief Storage in which to sort. */
  98     value_type** sorting_places;
  99
 100     /** @brief Storage into which to merge. */
 101     RandomAccessIterator* merging_places;
 102 #endif
 103     /** @brief Samples. */
 104     value_type* samples;
 105
 106     /** @brief Offsets to add to the found positions. */
 107     difference_type* offsets;
 108
 109     /** @brief Pieces of data to merge @c [thread][sequence] */
 110     std::vector<Piece<difference_type> >* pieces;
 111
 112     /** @brief Stable sorting desired. */
 113     bool stable;
 114 };
 115
 116 /**
 117   *  @brief Select samples from a sequence.
 118   *  @param sd Pointer to algorithm data. Result will be placed in
 119   *  @c sd->samples.
 120   *  @param num_samples Number of samples to select.
 121   */
 122 template<typename RandomAccessIterator, typename _DifferenceTp>
 123   inline void
 124   determine_samples(PMWMSSortingData<RandomAccessIterator>* sd,
 125                     _DifferenceTp& num_samples)
 126   {
 127     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 128     typedef typename traits_type::value_type value_type;
 129     typedef _DifferenceTp difference_type;
 130
 131     thread_index_t iam = omp_get_thread_num();
 132
 133     num_samples =
 134         Settings::sort_mwms_oversampling * sd->num_threads - 1;
 135
 136     difference_type* es = new difference_type[num_samples + 2];
 137
 138     equally_split(sd->starts[iam + 1] - sd->starts[iam],
 139                   num_samples + 1, es);
 140
 141     for (difference_type i = 0; i < num_samples; i++)
 142       new(&(sd->samples[iam * num_samples + i])) value_type(
 143           sd->source[sd->starts[iam] + es[i + 1]]);
 144
 145     delete[] es;
 146   }
 147
 148 /** @brief PMWMS code executed by each thread.
 149   *  @param sd Pointer to algorithm data.
 150   *  @param comp Comparator.
 151   */
 152 template<typename RandomAccessIterator, typename Comparator>
 153   inline void
 154   parallel_sort_mwms_pu(PMWMSSortingData<RandomAccessIterator>* sd,
 155                         Comparator& comp)
 156   {
 157     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 158     typedef typename traits_type::value_type value_type;
 159     typedef typename traits_type::difference_type difference_type;
 160
 161     thread_index_t iam = omp_get_thread_num();
 162
 163     // Length of this thread's chunk, before merging.
 164     difference_type length_local = sd->starts[iam + 1] - sd->starts[iam];
 165
 166 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 167     typedef RandomAccessIterator SortingPlacesIterator;
 168
 169     // Sort in input storage.
 170     sd->sorting_places[iam] = sd->source + sd->starts[iam];
 171 #else
 172     typedef value_type* SortingPlacesIterator;
 173
 174     // Sort in temporary storage, leave space for sentinel.
 175     sd->sorting_places[iam] = sd->temporaries[iam] =
 176         static_cast<value_type*>(
 177         ::operator new(sizeof(value_type) * (length_local + 1)));
 178
 179     // Copy there.
 180     std::uninitialized_copy(sd->source + sd->starts[iam],
 181                             sd->source + sd->starts[iam] + length_local,
 182                             sd->sorting_places[iam]);
 183 #endif
 184
 185     // Sort locally.
 186     if (sd->stable)
 187       __gnu_sequential::stable_sort(sd->sorting_places[iam],
 188                                     sd->sorting_places[iam] + length_local,
 189                                     comp);
 190     else
 191       __gnu_sequential::sort(sd->sorting_places[iam],
 192                              sd->sorting_places[iam] + length_local,
 193                              comp);
 194
 195     // Invariant: locally sorted subsequence in sd->sorting_places[iam],
 196     // sd->sorting_places[iam] + length_local.
 197
 198     if (Settings::sort_splitting == Settings::SAMPLING)
 199       {
 200         difference_type num_samples;
 201         determine_samples(sd, num_samples);
 202
 203 #       pragma omp barrier
 204
 205 #       pragma omp single
 206         __gnu_sequential::sort(sd->samples,
 207                                sd->samples + (num_samples * sd->num_threads),
 208                                comp);
 209
 210 #       pragma omp barrier
 211
 212         for (int s = 0; s < sd->num_threads; s++)
 213           {
 214             // For each sequence.
 215               if (num_samples * iam > 0)
 216                 sd->pieces[iam][s].begin =
 217                     std::lower_bound(sd->sorting_places[s],
 218                         sd->sorting_places[s]
 219                             + (sd->starts[s + 1] - sd->starts[s]),
 220                         sd->samples[num_samples * iam],
 221                         comp)
 222                     - sd->sorting_places[s];
 223             else
 224               // Absolute beginning.
 225               sd->pieces[iam][s].begin = 0;
 226
 227             if ((num_samples * (iam + 1)) < (num_samples * sd->num_threads))
 228               sd->pieces[iam][s].end =
 229                   std::lower_bound(sd->sorting_places[s],
 230                           sd->sorting_places[s]
 231                               + (sd->starts[s + 1] - sd->starts[s]),
 232                           sd->samples[num_samples * (iam + 1)],
 233                           comp)
 234                   - sd->sorting_places[s];
 235             else
 236               // Absolute end.
 237               sd->pieces[iam][s].end = sd->starts[s + 1] - sd->starts[s];
 238             }
 239       }
 240     else if (Settings::sort_splitting == Settings::EXACT)
 241       {
 242 #       pragma omp barrier
 243
 244         std::vector<std::pair<SortingPlacesIterator, SortingPlacesIterator> >
 245             seqs(sd->num_threads);
 246         for (int s = 0; s < sd->num_threads; s++)
 247           seqs[s] = std::make_pair(sd->sorting_places[s],
 248                                    sd->sorting_places[s]
 249                                        + (sd->starts[s + 1] - sd->starts[s]));
 250
 251         std::vector<SortingPlacesIterator> offsets(sd->num_threads);
 252
 253         // if not last thread
 254         if (iam < sd->num_threads - 1)
 255           multiseq_partition(seqs.begin(), seqs.end(),
 256                              sd->starts[iam + 1], offsets.begin(), comp);
 257
 258         for (int seq = 0; seq < sd->num_threads; seq++)
 259           {
 260             // for each sequence
 261             if (iam < (sd->num_threads - 1))
 262               sd->pieces[iam][seq].end = offsets[seq] - seqs[seq].first;
 263             else
 264               // very end of this sequence
 265               sd->pieces[iam][seq].end =
 266                   sd->starts[seq + 1] - sd->starts[seq];
 267           }
 268
 269 #       pragma omp barrier
 270
 271         for (int seq = 0; seq < sd->num_threads; seq++)
 272           {
 273             // For each sequence.
 274             if (iam > 0)
 275               sd->pieces[iam][seq].begin = sd->pieces[iam - 1][seq].end;
 276             else
 277               // Absolute beginning.
 278               sd->pieces[iam][seq].begin = 0;
 279           }
 280       }
 281
 282     // Offset from target begin, length after merging.
 283     difference_type offset = 0, length_am = 0;
 284     for (int s = 0; s < sd->num_threads; s++)
 285       {
 286         length_am += sd->pieces[iam][s].end - sd->pieces[iam][s].begin;
 287         offset += sd->pieces[iam][s].begin;
 288       }
 289
 290 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 291     // Merge to temporary storage, uninitialized creation not possible
 292     // since there is no multiway_merge calling the placement new
 293     // instead of the assignment operator.
 294     // XXX incorrect (de)construction
 295     sd->merging_places[iam] = sd->temporaries[iam] =
 296         static_cast<value_type*>(
 297         ::operator new(sizeof(value_type) * length_am));
 298 #else
 299     // Merge directly to target.
 300     sd->merging_places[iam] = sd->source + offset;
 301 #endif
 302     std::vector<std::pair<SortingPlacesIterator, SortingPlacesIterator> >
 303         seqs(sd->num_threads);
 304
 305     for (int s = 0; s < sd->num_threads; s++)
 306       {
 307         seqs[s] =
 308             std::make_pair(sd->sorting_places[s] + sd->pieces[iam][s].begin,
 309                            sd->sorting_places[s] + sd->pieces[iam][s].end);
 310       }
 311
 312     multiway_merge(seqs.begin(), seqs.end(), sd->merging_places[iam], comp,
 313                    length_am, sd->stable, false, sequential_tag());
 314
 315 #   pragma omp barrier
 316
 317 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 318     // Write back.
 319     std::copy(sd->merging_places[iam],
 320               sd->merging_places[iam] + length_am,
 321               sd->source + offset);
 322 #endif
 323
 324     delete[] sd->temporaries[iam];
 325   }
 326
 327 /** @brief PMWMS main call.
 328   *  @param begin Begin iterator of sequence.
 329   *  @param end End iterator of sequence.
 330   *  @param comp Comparator.
 331   *  @param n Length of sequence.
 332   *  @param num_threads Number of threads to use.
 333   *  @param stable Stable sorting.
 334   */
 335 template<typename RandomAccessIterator, typename Comparator>
 336   inline void
 337   parallel_sort_mwms(RandomAccessIterator begin, RandomAccessIterator end,
 338                      Comparator comp,
 339                      typename std::iterator_traits<RandomAccessIterator>
 340                         ::difference_type n,
 341                      int num_threads,
 342                      bool stable)
 343   {
 344     _GLIBCXX_CALL(n)
 345
 346     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 347     typedef typename traits_type::value_type value_type;
 348     typedef typename traits_type::difference_type difference_type;
 349
 350     if (n <= 1)
 351       return;
 352
 353     // at least one element per thread
 354     if (num_threads > n)
 355       num_threads = static_cast<thread_index_t>(n);
 356
 357     // shared variables
 358     PMWMSSortingData<RandomAccessIterator> sd;
 359     difference_type* starts;
 360
 361 #   pragma omp parallel num_threads(num_threads)
 362       {
 363         num_threads = omp_get_num_threads();  //no more threads than requested
 364
 365 #       pragma omp single
 366           {
 367             sd.num_threads = num_threads;
 368             sd.source = begin;
 369             sd.temporaries = new value_type*[num_threads];
 370
 371 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 372             sd.sorting_places = new RandomAccessIterator[num_threads];
 373             sd.merging_places = new value_type*[num_threads];
 374 #else
 375             sd.sorting_places = new value_type*[num_threads];
 376             sd.merging_places = new RandomAccessIterator[num_threads];
 377 #endif
 378
 379             if (Settings::sort_splitting == Settings::SAMPLING)
 380               {
 381                 unsigned int size =
 382                     (Settings::sort_mwms_oversampling * num_threads - 1)
 383                         * num_threads;
 384                 sd.samples = static_cast<value_type*>(
 385                     ::operator new(size * sizeof(value_type)));
 386               }
 387             else
 388               sd.samples = NULL;
 389
 390             sd.offsets = new difference_type[num_threads - 1];
 391             sd.pieces = new std::vector<Piece<difference_type> >[num_threads];
 392             for (int s = 0; s < num_threads; s++)
 393               sd.pieces[s].resize(num_threads);
 394             starts = sd.starts = new difference_type[num_threads + 1];
 395             sd.stable = stable;
 396
 397             difference_type chunk_length = n / num_threads;
 398             difference_type split = n % num_threads;
 399             difference_type pos = 0;
 400             for (int i = 0; i < num_threads; i++)
 401               {
 402                 starts[i] = pos;
 403                 pos += (i < split) ? (chunk_length + 1) : chunk_length;
 404               }
 405             starts[num_threads] = pos;
 406           }
 407
 408         // Now sort in parallel.
 409         parallel_sort_mwms_pu(&sd, comp);
 410       } //parallel
 411
 412     delete[] starts;
 413     delete[] sd.temporaries;
 414     delete[] sd.sorting_places;
 415     delete[] sd.merging_places;
 416
 417     if (Settings::sort_splitting == Settings::SAMPLING)
 418         delete[] sd.samples;
 419
 420     delete[] sd.offsets;
 421     delete[] sd.pieces;
 422   }
 423 } //namespace __gnu_parallel
 424
 425 #endif