mingw/lib/gcc/mingw32/4.3.3/include/c++/parallel/workstealing.h

   1 // -*- C++ -*-
   2
   3 // Copyright (C) 2007, 2008 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/workstealing.h
  32  *  @brief Parallelization of embarrassingly parallel execution by
  33  *  means of work-stealing.
  34  *
  35  *  Work stealing is described in
  36  *
  37  *  R. D. Blumofe and C. E. Leiserson.
  38  *  Scheduling multithreaded computations by work stealing.
  39  *  Journal of the ACM, 46(5):720–748, 1999.
  40  *
  41  *  This file is a GNU parallel extension to the Standard C++ Library.
  42  */
  43
  44 // Written by Felix Putze.
  45
  46 #ifndef _GLIBCXX_PARALLEL_WORKSTEALING_H
  47 #define _GLIBCXX_PARALLEL_WORKSTEALING_H 1
  48
  49 #include <parallel/parallel.h>
  50 #include <parallel/random_number.h>
  51 #include <parallel/compatibility.h>
  52
  53 namespace __gnu_parallel
  54 {
  55
  56 #define _GLIBCXX_JOB_VOLATILE volatile
  57
  58 /** @brief One job for a certain thread. */
  59 template<typename _DifferenceTp>
  60   struct Job
  61   {
  62     typedef _DifferenceTp difference_type;
  63
  64     /** @brief First element.
  65      *
  66      *  Changed by owning and stealing thread. By stealing thread,
  67      *  always incremented. */
  68     _GLIBCXX_JOB_VOLATILE difference_type first;
  69
  70     /** @brief Last element.
  71      *
  72      *  Changed by owning thread only. */
  73     _GLIBCXX_JOB_VOLATILE difference_type last;
  74
  75     /** @brief Number of elements, i. e. @c last-first+1.
  76      *
  77      *  Changed by owning thread only. */
  78     _GLIBCXX_JOB_VOLATILE difference_type load;
  79   };
  80
  81 /** @brief Work stealing algorithm for random access iterators.
  82   *
  83   *  Uses O(1) additional memory. Synchronization at job lists is
  84   *  done with atomic operations.
  85   *  @param begin Begin iterator of element sequence.
  86   *  @param end End iterator of element sequence.
  87   *  @param op User-supplied functor (comparator, predicate, adding
  88   *  functor, ...).
  89   *  @param f Functor to "process" an element with op (depends on
  90   *  desired functionality, e. g. for std::for_each(), ...).
  91   *  @param r Functor to "add" a single result to the already
  92   *  processed elements (depends on functionality).
  93   *  @param base Base value for reduction.
  94   *  @param output Pointer to position where final result is written to
  95   *  @param bound Maximum number of elements processed (e. g. for
  96   *  std::count_n()).
  97   *  @return User-supplied functor (that may contain a part of the result).
  98   */
  99 template<typename RandomAccessIterator,
 100          typename Op,
 101          typename Fu,
 102          typename Red,
 103          typename Result>
 104   Op
 105   for_each_template_random_access_workstealing(RandomAccessIterator begin,
 106                                                RandomAccessIterator end,
 107                                                Op op, Fu& f, Red r,
 108                                                Result base, Result& output,
 109                                                typename std::iterator_traits
 110                                                <RandomAccessIterator>::
 111                                                difference_type bound)
 112   {
 113     _GLIBCXX_CALL(end - begin)
 114
 115     typedef std::iterator_traits<RandomAccessIterator> traits_type;
 116     typedef typename traits_type::difference_type difference_type;
 117
 118     const _Settings& __s = _Settings::get();
 119
 120     difference_type chunk_size = static_cast<difference_type>(__s.workstealing_chunk_size);
 121
 122     // How many jobs?
 123     difference_type length = (bound < 0) ? (end - begin) : bound;
 124
 125     // To avoid false sharing in a cache line.
 126     const int stride = __s.cache_line_size * 10 / sizeof(Job<difference_type>) + 1;
 127
 128     // Total number of threads currently working.
 129     thread_index_t busy = 0;
 130
 131     Job<difference_type> *job;
 132
 133     omp_lock_t output_lock;
 134     omp_init_lock(&output_lock);
 135
 136     // Write base value to output.
 137     output = base;
 138
 139     // No more threads than jobs, at least one thread.
 140     thread_index_t num_threads =
 141         __gnu_parallel::max<thread_index_t>(1,
 142             __gnu_parallel::min<difference_type>(length, get_max_threads()));
 143
 144 #   pragma omp parallel shared(busy) num_threads(num_threads)
 145       {
 146
 147 #       pragma omp single
 148           {
 149             num_threads = omp_get_num_threads();
 150
 151             // Create job description array.
 152             job = new Job<difference_type>[num_threads * stride];
 153           }
 154
 155         // Initialization phase.
 156
 157         // Flags for every thread if it is doing productive work.
 158         bool iam_working = false;
 159
 160         // Thread id.
 161         thread_index_t iam = omp_get_thread_num();
 162
 163         // This job.
 164         Job<difference_type>& my_job = job[iam * stride];
 165
 166         // Random number (for work stealing).
 167         thread_index_t victim;
 168
 169         // Local value for reduction.
 170         Result result = Result();
 171
 172         // Number of elements to steal in one attempt.
 173         difference_type steal;
 174
 175         // Every thread has its own random number generator
 176         // (modulo num_threads).
 177         random_number rand_gen(iam, num_threads);
 178
 179         // This thread is currently working.
 180 #       pragma omp atomic
 181           ++busy;
 182
 183         iam_working = true;
 184
 185         // How many jobs per thread? last thread gets the rest.
 186         my_job.first =
 187             static_cast<difference_type>(iam * (length / num_threads));
 188
 189         my_job.last = (iam == (num_threads - 1)) ?
 190             (length - 1) : ((iam + 1) * (length / num_threads) - 1);
 191         my_job.load = my_job.last - my_job.first + 1;
 192
 193         // Init result with first value (to have a base value for reduction).
 194         if (my_job.first <= my_job.last)
 195           {
 196             // Cannot use volatile variable directly.
 197             difference_type my_first = my_job.first;
 198             result = f(op, begin + my_first);
 199             ++my_job.first;
 200             --my_job.load;
 201           }
 202
 203         RandomAccessIterator current;
 204
 205 #       pragma omp barrier
 206
 207         // Actual work phase
 208         // Work on own or stolen start
 209         while (busy > 0)
 210           {
 211             // Work until no productive thread left.
 212 #           pragma omp flush(busy)
 213
 214             // Thread has own work to do
 215             while (my_job.first <= my_job.last)
 216               {
 217                 // fetch-and-add call
 218                 // Reserve current job block (size chunk_size) in my queue.
 219                 difference_type current_job =
 220                   fetch_and_add<difference_type>(&(my_job.first), chunk_size);
 221
 222                 // Update load, to make the three values consistent,
 223                 // first might have been changed in the meantime
 224                 my_job.load = my_job.last - my_job.first + 1;
 225                 for (difference_type job_counter = 0;
 226                      job_counter < chunk_size && current_job <= my_job.last;
 227                      ++job_counter)
 228                   {
 229                     // Yes: process it!
 230                     current = begin + current_job;
 231                     ++current_job;
 232
 233                     // Do actual work.
 234                     result = r(result, f(op, current));
 235                   }
 236
 237 #               pragma omp flush(busy)
 238               }
 239
 240             // After reaching this point, a thread's job list is empty.
 241             if (iam_working)
 242               {
 243                 // This thread no longer has work.
 244 #               pragma omp atomic
 245                 --busy;
 246
 247                 iam_working = false;
 248               }
 249
 250             difference_type supposed_first, supposed_last, supposed_load;
 251             do
 252               {
 253                 // Find random nonempty deque (not own), do consistency check.
 254                 yield();
 255 #               pragma omp flush(busy)
 256                 victim = rand_gen();
 257                 supposed_first = job[victim * stride].first;
 258                 supposed_last = job[victim * stride].last;
 259                 supposed_load = job[victim * stride].load;
 260               }
 261             while (busy > 0
 262               && ((supposed_load <= 0)
 263                 || ((supposed_first + supposed_load - 1) != supposed_last)));
 264
 265             if (busy == 0)
 266               break;
 267
 268             if (supposed_load > 0)
 269               {
 270                 // Has work and work to do.
 271                 // Number of elements to steal (at least one).
 272                 steal = (supposed_load < 2) ? 1 : supposed_load / 2;
 273
 274                 // Push victim's start forward.
 275                 difference_type stolen_first =
 276                     fetch_and_add<difference_type>(
 277                         &(job[victim * stride].first), steal);
 278                 difference_type stolen_try =
 279                     stolen_first + steal - difference_type(1);
 280
 281                 my_job.first = stolen_first;
 282                 my_job.last = __gnu_parallel::min(stolen_try, supposed_last);
 283                 my_job.load = my_job.last - my_job.first + 1;
 284
 285                 // Has potential work again.
 286 #               pragma omp atomic
 287                   ++busy;
 288                 iam_working = true;
 289
 290 #               pragma omp flush(busy)
 291               }
 292 #           pragma omp flush(busy)
 293           } // end while busy > 0
 294             // Add accumulated result to output.
 295         omp_set_lock(&output_lock);
 296         output = r(output, result);
 297         omp_unset_lock(&output_lock);
 298       }
 299
 300     delete[] job;
 301
 302     // Points to last element processed (needed as return value for
 303     // some algorithms like transform)
 304     f.finish_iterator = begin + length;
 305
 306     omp_destroy_lock(&output_lock);
 307
 308     return op;
 309   }
 310 } // end namespace
 311
 312 #endif