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 Input begin. */
  75     RandomAccessIterator source;
  76
  77     /** @brief Start indices, per thread. */
  78     difference_type* starts;
  79
  80     /** @brief Temporary arrays for each thread.
  81      *
  82      *  Indirection Allows using the temporary storage in different
  83      *  ways, without code duplication.
  84      *  @see _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST */
  85     value_type** temporaries;
  86
  87 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
  88     /** @brief Storage in which to sort. */
  89     RandomAccessIterator* sorting_places;
  90
  91     /** @brief Storage into which to merge. */
  92     value_type** merging_places;
  93 #else
  94     /** @brief Storage in which to sort. */
  95     value_type** sorting_places;
  96
  97     /** @brief Storage into which to merge. */
  98     RandomAccessIterator* merging_places;
  99 #endif
 100     /** @brief Samples. */
 101     value_type* samples;
 102
 103     /** @brief Offsets to add to the found positions. */
 104     difference_type* offsets;
 105
 106     /** @brief Pieces of data to merge @c [thread][sequence] */
 107     std::vector<Piece<difference_type> >* pieces;
 108   };
 109
 110   /** @brief Thread local data for PMWMS. */
 111   template<typename RandomAccessIterator>
 112   struct PMWMSSorterPU
 113   {
 114     /** @brief Total number of thread involved. */
 115     thread_index_t num_threads;
 116     /** @brief Number of owning thread. */
 117     thread_index_t iam;
 118     /** @brief Stable sorting desired. */
 119     bool stable;
 120     /** @brief Pointer to global data. */
 121     PMWMSSortingData<RandomAccessIterator>* sd;
 122   };
 123
 124   /**
 125    *  @brief Select samples from a sequence.
 126    *  @param d Pointer to thread-local data. Result will be placed in
 127    *  @c d->ds->samples.
 128    *  @param num_samples Number of samples to select.
 129    */
 130   template<typename RandomAccessIterator, typename _DifferenceTp>
 131   inline void
 132   determine_samples(PMWMSSorterPU<RandomAccessIterator>* d,
 133                     _DifferenceTp& num_samples)
 134   {
 135     typedef _DifferenceTp difference_type;
 136
 137     PMWMSSortingData<RandomAccessIterator>* sd = d->sd;
 138
 139     num_samples = Settings::sort_mwms_oversampling * d->num_threads - 1;
 140
 141     difference_type* es = static_cast<difference_type*>(__builtin_alloca(sizeof(difference_type) * (num_samples + 2)));
 142
 143     equally_split(sd->starts[d->iam + 1] - sd->starts[d->iam], num_samples + 1, es);
 144
 145     for (difference_type i = 0; i < num_samples; i++)
 146       sd->samples[d->iam * num_samples + i] = sd->source[sd->starts[d->iam] + es[i + 1]];
 147   }
 148
 149   /** @brief PMWMS code executed by each thread.
 150    *  @param d Pointer to thread-local data.
 151    *  @param comp Comparator.
 152    */
 153   template<typename RandomAccessIterator, typename Comparator>
 154   inline void
 155   parallel_sort_mwms_pu(PMWMSSorterPU<RandomAccessIterator>* d,
 156                         Comparator& comp)
 157   {
 158     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 159     typedef typename traits_type::value_type value_type;
 160     typedef typename traits_type::difference_type difference_type;
 161
 162     PMWMSSortingData<RandomAccessIterator>* sd = d->sd;
 163     thread_index_t iam = d->iam;
 164
 165     // Length of this thread's chunk, before merging.
 166     difference_type length_local = sd->starts[iam + 1] - sd->starts[iam];
 167
 168 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 169     typedef RandomAccessIterator SortingPlacesIterator;
 170
 171     // Sort in input storage.
 172     sd->sorting_places[iam] = sd->source + sd->starts[iam];
 173 #else
 174     typedef value_type* SortingPlacesIterator;
 175
 176     // Sort in temporary storage, leave space for sentinel.
 177     sd->sorting_places[iam] = sd->temporaries[iam] = static_cast<value_type*>(::operator new(sizeof(value_type) * (length_local + 1)));
 178
 179     // Copy there.
 180     std::uninitialized_copy(sd->source + sd->starts[iam], sd->source + sd->starts[iam] + length_local, sd->sorting_places[iam]);
 181 #endif
 182
 183     // Sort locally.
 184     if (d->stable)
 185       __gnu_sequential::stable_sort(sd->sorting_places[iam], sd->sorting_places[iam] + length_local, comp);
 186     else
 187       __gnu_sequential::sort(sd->sorting_places[iam], sd->sorting_places[iam] + length_local, comp);
 188
 189 #if _GLIBCXX_ASSERTIONS
 190     _GLIBCXX_PARALLEL_ASSERT(is_sorted(sd->sorting_places[iam], sd->sorting_places[iam] + length_local, comp));
 191 #endif
 192
 193     // Invariant: locally sorted subsequence in sd->sorting_places[iam],
 194     // sd->sorting_places[iam] + length_local.
 195
 196     if (Settings::sort_splitting == Settings::SAMPLING)
 197       {
 198         difference_type num_samples;
 199         determine_samples(d, num_samples);
 200
 201 #pragma omp barrier
 202
 203 #pragma omp single
 204         __gnu_sequential::sort(sd->samples,
 205                                sd->samples + (num_samples * d->num_threads),
 206                                comp);
 207
 208 #pragma omp barrier
 209
 210         for (int s = 0; s < d->num_threads; s++)
 211           {
 212             // For each sequence.
 213             if (num_samples * iam > 0)
 214               sd->pieces[iam][s].begin = std::lower_bound(sd->sorting_places[s],
 215                                  sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s],
 216                                  sd->samples[num_samples * iam],
 217                                  comp)
 218                 - sd->sorting_places[s];
 219             else
 220               // Absolute beginning.
 221               sd->pieces[iam][s].begin = 0;
 222
 223             if ((num_samples * (iam + 1)) < (num_samples * d->num_threads))
 224               sd->pieces[iam][s].end = std::lower_bound(sd->sorting_places[s],
 225                                                         sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s], sd->samples[num_samples * (iam + 1)], comp)
 226                 - sd->sorting_places[s];
 227             else
 228               // Absolute end.
 229               sd->pieces[iam][s].end = sd->starts[s + 1] - sd->starts[s];
 230           }
 231
 232       }
 233     else if (Settings::sort_splitting == Settings::EXACT)
 234       {
 235 #pragma omp barrier
 236
 237         std::vector<std::pair<SortingPlacesIterator, SortingPlacesIterator> > seqs(d->num_threads);
 238         for (int s = 0; s < d->num_threads; s++)
 239           seqs[s] = std::make_pair(sd->sorting_places[s], sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s]);
 240
 241         std::vector<SortingPlacesIterator> offsets(d->num_threads);
 242
 243         // If not last thread.
 244         if (iam < d->num_threads - 1)
 245           multiseq_partition(seqs.begin(), seqs.end(), sd->starts[iam + 1], offsets.begin(), comp);
 246
 247         for (int seq = 0; seq < d->num_threads; seq++)
 248           {
 249             // For each sequence.
 250             if (iam < (d->num_threads - 1))
 251               sd->pieces[iam][seq].end = offsets[seq] - seqs[seq].first;
 252             else
 253               // Absolute end of this sequence.
 254               sd->pieces[iam][seq].end = sd->starts[seq + 1] - sd->starts[seq];
 255           }
 256
 257 #pragma omp barrier
 258
 259         for (int seq = 0; seq < d->num_threads; seq++)
 260           {
 261             // For each sequence.
 262             if (iam > 0)
 263               sd->pieces[iam][seq].begin = sd->pieces[iam - 1][seq].end;
 264             else
 265               // Absolute beginning.
 266               sd->pieces[iam][seq].begin = 0;
 267           }
 268       }
 269
 270     // Offset from target begin, length after merging.
 271     difference_type offset = 0, length_am = 0;
 272     for (int s = 0; s < d->num_threads; s++)
 273       {
 274         length_am += sd->pieces[iam][s].end - sd->pieces[iam][s].begin;
 275         offset += sd->pieces[iam][s].begin;
 276       }
 277
 278 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 279     // Merge to temporary storage, uninitialized creation not possible
 280     // since there is no multiway_merge calling the placement new
 281     // instead of the assignment operator.
 282     sd->merging_places[iam] = sd->temporaries[iam] = static_cast<value_type*>(::operator new(sizeof(value_type) * length_am));
 283 #else
 284     // Merge directly to target.
 285     sd->merging_places[iam] = sd->source + offset;
 286 #endif
 287     std::vector<std::pair<SortingPlacesIterator, SortingPlacesIterator> > seqs(d->num_threads);
 288
 289     for (int s = 0; s < d->num_threads; s++)
 290       {
 291         seqs[s] = std::make_pair(sd->sorting_places[s] + sd->pieces[iam][s].begin, sd->sorting_places[s] + sd->pieces[iam][s].end);
 292
 293 #if _GLIBCXX_ASSERTIONS
 294         _GLIBCXX_PARALLEL_ASSERT(is_sorted(seqs[s].first, seqs[s].second, comp));
 295 #endif
 296       }
 297
 298     multiway_merge(seqs.begin(), seqs.end(), sd->merging_places[iam], comp, length_am, d->stable, false, sequential_tag());
 299
 300 #if _GLIBCXX_ASSERTIONS
 301     _GLIBCXX_PARALLEL_ASSERT(is_sorted(sd->merging_places[iam], sd->merging_places[iam] + length_am, comp));
 302 #endif
 303
 304 #       pragma omp barrier
 305
 306 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 307     // Write back.
 308     std::copy(sd->merging_places[iam], sd->merging_places[iam] + length_am,
 309               sd->source + offset);
 310 #endif
 311
 312     delete[] sd->temporaries[iam];
 313   }
 314
 315   /** @brief PMWMS main call.
 316    *  @param begin Begin iterator of sequence.
 317    *  @param end End iterator of sequence.
 318    *  @param comp Comparator.
 319    *  @param n Length of sequence.
 320    *  @param num_threads Number of threads to use.
 321    *  @param stable Stable sorting.
 322    */
 323   template<typename RandomAccessIterator, typename Comparator>
 324   inline void
 325   parallel_sort_mwms(RandomAccessIterator begin, RandomAccessIterator end,
 326                      Comparator comp,
 327        typename std::iterator_traits<RandomAccessIterator>::difference_type n,
 328                      int num_threads, bool stable)
 329   {
 330     _GLIBCXX_CALL(n)
 331
 332     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 333     typedef typename traits_type::value_type value_type;
 334     typedef typename traits_type::difference_type difference_type;
 335
 336     if (n <= 1)
 337       return;
 338
 339     // At least one element per thread.
 340     if (num_threads > n)
 341       num_threads = static_cast<thread_index_t>(n);
 342
 343     PMWMSSortingData<RandomAccessIterator> sd;
 344
 345     sd.source = begin;
 346     sd.temporaries = new value_type*[num_threads];
 347
 348 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 349     sd.sorting_places = new RandomAccessIterator[num_threads];
 350     sd.merging_places = new value_type*[num_threads];
 351 #else
 352     sd.sorting_places = new value_type*[num_threads];
 353     sd.merging_places = new RandomAccessIterator[num_threads];
 354 #endif
 355
 356     if (Settings::sort_splitting == Settings::SAMPLING)
 357       {
 358         unsigned int sz = Settings::sort_mwms_oversampling * num_threads - 1;
 359         sz *= num_threads;
 360
 361         // Equivalent to value_type[sz], without need of default construction.
 362         sz *= sizeof(value_type);
 363         sd.samples = static_cast<value_type*>(::operator new(sz));
 364       }
 365     else
 366       sd.samples = NULL;
 367
 368     sd.offsets = new difference_type[num_threads - 1];
 369     sd.pieces = new std::vector<Piece<difference_type> >[num_threads];
 370     for (int s = 0; s < num_threads; s++)
 371       sd.pieces[s].resize(num_threads);
 372     PMWMSSorterPU<RandomAccessIterator>* pus = new PMWMSSorterPU<RandomAccessIterator>[num_threads];
 373     difference_type* starts = sd.starts = new difference_type[num_threads + 1];
 374
 375     difference_type chunk_length = n / num_threads;
 376     difference_type split = n % num_threads;
 377     difference_type start = 0;
 378     for (int i = 0; i < num_threads; i++)
 379       {
 380         starts[i] = start;
 381         start += (i < split) ? (chunk_length + 1) : chunk_length;
 382         pus[i].num_threads = num_threads;
 383         pus[i].iam = i;
 384         pus[i].sd = &sd;
 385         pus[i].stable = stable;
 386       }
 387     starts[num_threads] = start;
 388
 389     // Now sort in parallel.
 390 #pragma omp parallel num_threads(num_threads)
 391     parallel_sort_mwms_pu(&(pus[omp_get_thread_num()]), comp);
 392
 393     // XXX sd as RAII
 394     delete[] starts;
 395     delete[] sd.temporaries;
 396     delete[] sd.sorting_places;
 397     delete[] sd.merging_places;
 398
 399     if (Settings::sort_splitting == Settings::SAMPLING)
 400       delete[] sd.samples;
 401
 402     delete[] sd.offsets;
 403     delete[] sd.pieces;
 404
 405     delete[] pus;
 406   }
 407
 408 }
 409
 410 #endif