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 #include <parallel/timing.h>
  48
  49 namespace __gnu_parallel
  50 {
  51
  52   /** @brief Subsequence description. */
  53   template<typename _DifferenceTp>
  54   struct Piece
  55   {
  56     typedef _DifferenceTp difference_type;
  57
  58     /** @brief Begin of subsequence. */
  59     difference_type begin;
  60
  61     /** @brief End of subsequence. */
  62     difference_type end;
  63   };
  64
  65   /** @brief Data accessed by all threads.
  66    *
  67    *  PMWMS = parallel multiway mergesort */
  68   template<typename RandomAccessIterator>
  69   struct PMWMSSortingData
  70   {
  71     typedef std::iterator_traits<RandomAccessIterator> traits_type;
  72     typedef typename traits_type::value_type value_type;
  73     typedef typename traits_type::difference_type difference_type;
  74
  75     /** @brief Input begin. */
  76     RandomAccessIterator source;
  77
  78     /** @brief Start indices, per thread. */
  79     difference_type* starts;
  80
  81     /** @brief Temporary arrays for each thread.
  82      *
  83      *  Indirection Allows using the temporary storage in different
  84      *  ways, without code duplication.
  85      *  @see _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST */
  86     value_type** temporaries;
  87
  88 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
  89     /** @brief Storage in which to sort. */
  90     RandomAccessIterator* sorting_places;
  91
  92     /** @brief Storage into which to merge. */
  93     value_type** merging_places;
  94 #else
  95     /** @brief Storage in which to sort. */
  96     value_type** sorting_places;
  97
  98     /** @brief Storage into which to merge. */
  99     RandomAccessIterator* merging_places;
 100 #endif
 101     /** @brief Samples. */
 102     value_type* samples;
 103
 104     /** @brief Offsets to add to the found positions. */
 105     difference_type* offsets;
 106
 107     /** @brief Pieces of data to merge @c [thread][sequence] */
 108     std::vector<Piece<difference_type> >* pieces;
 109   };
 110
 111   /** @brief Thread local data for PMWMS. */
 112   template<typename RandomAccessIterator>
 113   struct PMWMSSorterPU
 114   {
 115     /** @brief Total number of thread involved. */
 116     thread_index_t num_threads;
 117     /** @brief Number of owning thread. */
 118     thread_index_t iam;
 119     /** @brief Stable sorting desired. */
 120     bool stable;
 121     /** @brief Pointer to global data. */
 122     PMWMSSortingData<RandomAccessIterator>* sd;
 123   };
 124
 125   /**
 126    *  @brief Select samples from a sequence.
 127    *  @param d Pointer to thread-local data. Result will be placed in
 128    *  @c d->ds->samples.
 129    *  @param num_samples Number of samples to select.
 130    */
 131   template<typename RandomAccessIterator, typename _DifferenceTp>
 132   inline void
 133   determine_samples(PMWMSSorterPU<RandomAccessIterator>* d,
 134                     _DifferenceTp& num_samples)
 135   {
 136     typedef _DifferenceTp difference_type;
 137
 138     PMWMSSortingData<RandomAccessIterator>* sd = d->sd;
 139
 140     num_samples = Settings::sort_mwms_oversampling * d->num_threads - 1;
 141
 142     difference_type* es = static_cast<difference_type*>(__builtin_alloca(sizeof(difference_type) * (num_samples + 2)));
 143
 144     equally_split(sd->starts[d->iam + 1] - sd->starts[d->iam], num_samples + 1, es);
 145
 146     for (difference_type i = 0; i < num_samples; i++)
 147       sd->samples[d->iam * num_samples + i] = sd->source[sd->starts[d->iam] + es[i + 1]];
 148   }
 149
 150   /** @brief PMWMS code executed by each thread.
 151    *  @param d Pointer to thread-local data.
 152    *  @param comp Comparator.
 153    */
 154   template<typename RandomAccessIterator, typename Comparator>
 155   inline void
 156   parallel_sort_mwms_pu(PMWMSSorterPU<RandomAccessIterator>* d,
 157                         Comparator& comp)
 158   {
 159     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 160     typedef typename traits_type::value_type value_type;
 161     typedef typename traits_type::difference_type difference_type;
 162
 163     Timing<sequential_tag> t;
 164     t.tic();
 165
 166     PMWMSSortingData<RandomAccessIterator>* sd = d->sd;
 167     thread_index_t iam = d->iam;
 168
 169     // Length of this thread's chunk, before merging.
 170     difference_type length_local = sd->starts[iam + 1] - sd->starts[iam];
 171
 172 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 173     typedef RandomAccessIterator SortingPlacesIterator;
 174
 175     // Sort in input storage.
 176     sd->sorting_places[iam] = sd->source + sd->starts[iam];
 177 #else
 178     typedef value_type* SortingPlacesIterator;
 179
 180     // Sort in temporary storage, leave space for sentinel.
 181     sd->sorting_places[iam] = sd->temporaries[iam] = static_cast<value_type*>(::operator new(sizeof(value_type) * (length_local + 1)));
 182
 183     // Copy there.
 184     std::uninitialized_copy(sd->source + sd->starts[iam], sd->source + sd->starts[iam] + length_local, sd->sorting_places[iam]);
 185 #endif
 186
 187     // Sort locally.
 188     if (d->stable)
 189       __gnu_sequential::stable_sort(sd->sorting_places[iam], sd->sorting_places[iam] + length_local, comp);
 190     else
 191       __gnu_sequential::sort(sd->sorting_places[iam], sd->sorting_places[iam] + length_local, comp);
 192
 193 #if _GLIBCXX_ASSERTIONS
 194     _GLIBCXX_PARALLEL_ASSERT(is_sorted(sd->sorting_places[iam], sd->sorting_places[iam] + length_local, comp));
 195 #endif
 196
 197     // Invariant: locally sorted subsequence in sd->sorting_places[iam],
 198     // sd->sorting_places[iam] + length_local.
 199     t.tic("local sort");
 200
 201     if (Settings::sort_splitting == Settings::SAMPLING)
 202       {
 203         difference_type num_samples;
 204         determine_samples(d, num_samples);
 205
 206 #pragma omp barrier
 207
 208         t.tic("sample/wait");
 209
 210 #pragma omp single
 211         __gnu_sequential::sort(sd->samples,
 212                                sd->samples + (num_samples * d->num_threads),
 213                                comp);
 214
 215 #pragma omp barrier
 216
 217         for (int s = 0; s < d->num_threads; s++)
 218           {
 219             // For each sequence.
 220             if (num_samples * iam > 0)
 221               sd->pieces[iam][s].begin = std::lower_bound(sd->sorting_places[s],
 222                                  sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s],
 223                                  sd->samples[num_samples * iam],
 224                                  comp)
 225                 - sd->sorting_places[s];
 226             else
 227               // Absolute beginning.
 228               sd->pieces[iam][s].begin = 0;
 229
 230             if ((num_samples * (iam + 1)) < (num_samples * d->num_threads))
 231               sd->pieces[iam][s].end = std::lower_bound(sd->sorting_places[s],
 232                                                         sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s], sd->samples[num_samples * (iam + 1)], comp)
 233                 - sd->sorting_places[s];
 234             else
 235               // Absolute end.
 236               sd->pieces[iam][s].end = sd->starts[s + 1] - sd->starts[s];
 237           }
 238
 239       }
 240     else if (Settings::sort_splitting == Settings::EXACT)
 241       {
 242 #pragma omp barrier
 243
 244         t.tic("wait");
 245
 246         std::vector<std::pair<SortingPlacesIterator, SortingPlacesIterator> > seqs(d->num_threads);
 247         for (int s = 0; s < d->num_threads; s++)
 248           seqs[s] = std::make_pair(sd->sorting_places[s], sd->sorting_places[s] + sd->starts[s + 1] - sd->starts[s]);
 249
 250         std::vector<SortingPlacesIterator> offsets(d->num_threads);
 251
 252         // If not last thread.
 253         if (iam < d->num_threads - 1)
 254           multiseq_partition(seqs.begin(), seqs.end(), sd->starts[iam + 1], offsets.begin(), comp);
 255
 256         for (int seq = 0; seq < d->num_threads; seq++)
 257           {
 258             // For each sequence.
 259             if (iam < (d->num_threads - 1))
 260               sd->pieces[iam][seq].end = offsets[seq] - seqs[seq].first;
 261             else
 262               // Absolute end of this sequence.
 263               sd->pieces[iam][seq].end = sd->starts[seq + 1] - sd->starts[seq];
 264           }
 265
 266 #pragma omp barrier
 267
 268         for (int seq = 0; seq < d->num_threads; seq++)
 269           {
 270             // For each sequence.
 271             if (iam > 0)
 272               sd->pieces[iam][seq].begin = sd->pieces[iam - 1][seq].end;
 273             else
 274               // Absolute beginning.
 275               sd->pieces[iam][seq].begin = 0;
 276           }
 277       }
 278
 279     t.tic("split");
 280
 281     // Offset from target begin, length after merging.
 282     difference_type offset = 0, length_am = 0;
 283     for (int s = 0; s < d->num_threads; s++)
 284       {
 285         length_am += sd->pieces[iam][s].end - sd->pieces[iam][s].begin;
 286         offset += sd->pieces[iam][s].begin;
 287       }
 288
 289 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 290     // Merge to temporary storage, uninitialized creation not possible
 291     // since there is no multiway_merge calling the placement new
 292     // instead of the assignment operator.
 293     sd->merging_places[iam] = sd->temporaries[iam] = static_cast<value_type*>(::operator new(sizeof(value_type) * length_am));
 294 #else
 295     // Merge directly to target.
 296     sd->merging_places[iam] = sd->source + offset;
 297 #endif
 298     std::vector<std::pair<SortingPlacesIterator, SortingPlacesIterator> > seqs(d->num_threads);
 299
 300     for (int s = 0; s < d->num_threads; s++)
 301       {
 302         seqs[s] = std::make_pair(sd->sorting_places[s] + sd->pieces[iam][s].begin, sd->sorting_places[s] + sd->pieces[iam][s].end);
 303
 304 #if _GLIBCXX_ASSERTIONS
 305         _GLIBCXX_PARALLEL_ASSERT(is_sorted(seqs[s].first, seqs[s].second, comp));
 306 #endif
 307       }
 308
 309     multiway_merge(seqs.begin(), seqs.end(), sd->merging_places[iam], comp, length_am, d->stable, false, sequential_tag());
 310
 311     t.tic("merge");
 312
 313 #if _GLIBCXX_ASSERTIONS
 314     _GLIBCXX_PARALLEL_ASSERT(is_sorted(sd->merging_places[iam], sd->merging_places[iam] + length_am, comp));
 315 #endif
 316
 317 #       pragma omp barrier
 318
 319 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 320     // Write back.
 321     std::copy(sd->merging_places[iam], sd->merging_places[iam] + length_am,
 322               sd->source + offset);
 323 #endif
 324
 325     delete[] sd->temporaries[iam];
 326
 327     t.tic("copy back");
 328
 329     t.print();
 330   }
 331
 332   /** @brief PMWMS main call.
 333    *  @param begin Begin iterator of sequence.
 334    *  @param end End iterator of sequence.
 335    *  @param comp Comparator.
 336    *  @param n Length of sequence.
 337    *  @param num_threads Number of threads to use.
 338    *  @param stable Stable sorting.
 339    */
 340   template<typename RandomAccessIterator, typename Comparator>
 341   inline void
 342   parallel_sort_mwms(RandomAccessIterator begin, RandomAccessIterator end,
 343                      Comparator comp,
 344        typename std::iterator_traits<RandomAccessIterator>::difference_type n,
 345                      int num_threads, bool stable)
 346   {
 347     _GLIBCXX_CALL(n)
 348
 349     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 350     typedef typename traits_type::value_type value_type;
 351     typedef typename traits_type::difference_type difference_type;
 352
 353     if (n <= 1)
 354       return;
 355
 356     // At least one element per thread.
 357     if (num_threads > n)
 358       num_threads = static_cast<thread_index_t>(n);
 359
 360     PMWMSSortingData<RandomAccessIterator> sd;
 361
 362     sd.source = begin;
 363     sd.temporaries = new value_type*[num_threads];
 364
 365 #if _GLIBCXX_MULTIWAY_MERGESORT_COPY_LAST
 366     sd.sorting_places = new RandomAccessIterator[num_threads];
 367     sd.merging_places = new value_type*[num_threads];
 368 #else
 369     sd.sorting_places = new value_type*[num_threads];
 370     sd.merging_places = new RandomAccessIterator[num_threads];
 371 #endif
 372
 373     if (Settings::sort_splitting == Settings::SAMPLING)
 374       {
 375         unsigned int sz = Settings::sort_mwms_oversampling * num_threads - 1;
 376         sz *= num_threads;
 377
 378         // Equivalent to value_type[sz], without need of default construction.
 379         sz *= sizeof(value_type);
 380         sd.samples = static_cast<value_type*>(::operator new(sz));
 381       }
 382     else
 383       sd.samples = NULL;
 384
 385     sd.offsets = new difference_type[num_threads - 1];
 386     sd.pieces = new std::vector<Piece<difference_type> >[num_threads];
 387     for (int s = 0; s < num_threads; s++)
 388       sd.pieces[s].resize(num_threads);
 389     PMWMSSorterPU<RandomAccessIterator>* pus = new PMWMSSorterPU<RandomAccessIterator>[num_threads];
 390     difference_type* starts = sd.starts = new difference_type[num_threads + 1];
 391
 392     difference_type chunk_length = n / num_threads;
 393     difference_type split = n % num_threads;
 394     difference_type start = 0;
 395     for (int i = 0; i < num_threads; i++)
 396       {
 397         starts[i] = start;
 398         start += (i < split) ? (chunk_length + 1) : chunk_length;
 399         pus[i].num_threads = num_threads;
 400         pus[i].iam = i;
 401         pus[i].sd = &sd;
 402         pus[i].stable = stable;
 403       }
 404     starts[num_threads] = start;
 405
 406     // Now sort in parallel.
 407 #pragma omp parallel num_threads(num_threads)
 408     parallel_sort_mwms_pu(&(pus[omp_get_thread_num()]), comp);
 409
 410     // XXX sd as RAII
 411     delete[] starts;
 412     delete[] sd.temporaries;
 413     delete[] sd.sorting_places;
 414     delete[] sd.merging_places;
 415
 416     if (Settings::sort_splitting == Settings::SAMPLING)
 417       delete[] sd.samples;
 418
 419     delete[] sd.offsets;
 420     delete[] sd.pieces;
 421
 422     delete[] pus;
 423   }
 424
 425 }
 426
 427 #endif