workstealing.h

Go to the documentation of this file.
00001 // -*- C++ -*-
00002 
00003 // Copyright (C) 2007, 2008, 2009 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/workstealing.h
00026  *  @brief Parallelization of embarrassingly parallel execution by
00027  *  means of work-stealing.
00028  *
00029  *  Work stealing is described in
00030  *
00031  *  R. D. Blumofe and C. E. Leiserson.
00032  *  Scheduling multithreaded computations by work stealing.
00033  *  Journal of the ACM, 46(5):720–748, 1999.
00034  *
00035  *  This file is a GNU parallel extension to the Standard C++ Library.
00036  */
00037 
00038 // Written by Felix Putze.
00039 
00040 #ifndef _GLIBCXX_PARALLEL_WORKSTEALING_H
00041 #define _GLIBCXX_PARALLEL_WORKSTEALING_H 1
00042 
00043 #include <parallel/parallel.h>
00044 #include <parallel/random_number.h>
00045 #include <parallel/compatibility.h>
00046 
00047 namespace __gnu_parallel
00048 {
00049 
00050 #define _GLIBCXX_JOB_VOLATILE volatile
00051 
00052 /** @brief One job for a certain thread. */
00053 template<typename _DifferenceTp>
00054   struct Job
00055   {
00056     typedef _DifferenceTp difference_type;
00057 
00058     /** @brief First element.
00059      *
00060      *  Changed by owning and stealing thread. By stealing thread,
00061      *  always incremented. */
00062     _GLIBCXX_JOB_VOLATILE difference_type first;
00063 
00064     /** @brief Last element.
00065      *
00066      *  Changed by owning thread only. */
00067     _GLIBCXX_JOB_VOLATILE difference_type last;
00068 
00069     /** @brief Number of elements, i. e. @c last-first+1.
00070      *
00071      *  Changed by owning thread only. */
00072     _GLIBCXX_JOB_VOLATILE difference_type load;
00073   };
00074 
00075 /** @brief Work stealing algorithm for random access iterators.
00076   *
00077   *  Uses O(1) additional memory. Synchronization at job lists is
00078   *  done with atomic operations.
00079   *  @param begin Begin iterator of element sequence.
00080   *  @param end End iterator of element sequence.
00081   *  @param op User-supplied functor (comparator, predicate, adding
00082   *  functor, ...).
00083   *  @param f Functor to "process" an element with op (depends on
00084   *  desired functionality, e. g. for std::for_each(), ...).
00085   *  @param r Functor to "add" a single result to the already
00086   *  processed elements (depends on functionality).
00087   *  @param base Base value for reduction.
00088   *  @param output Pointer to position where final result is written to
00089   *  @param bound Maximum number of elements processed (e. g. for
00090   *  std::count_n()).
00091   *  @return User-supplied functor (that may contain a part of the result).
00092   */
00093 template<typename RandomAccessIterator,
00094      typename Op,
00095      typename Fu,
00096      typename Red,
00097      typename Result>
00098   Op
00099   for_each_template_random_access_workstealing(RandomAccessIterator begin,
00100                            RandomAccessIterator end,
00101                            Op op, Fu& f, Red r,
00102                            Result base, Result& output,
00103                            typename std::iterator_traits
00104                            <RandomAccessIterator>::
00105                            difference_type bound)
00106   {
00107     _GLIBCXX_CALL(end - begin)
00108 
00109     typedef std::iterator_traits<RandomAccessIterator> traits_type;
00110     typedef typename traits_type::difference_type difference_type;
00111     
00112     const _Settings& __s = _Settings::get();
00113 
00114     difference_type chunk_size = static_cast<difference_type>(__s.workstealing_chunk_size);
00115 
00116     // How many jobs?
00117     difference_type length = (bound < 0) ? (end - begin) : bound;
00118 
00119     // To avoid false sharing in a cache line.
00120     const int stride = __s.cache_line_size * 10 / sizeof(Job<difference_type>) + 1;
00121 
00122     // Total number of threads currently working.
00123     thread_index_t busy = 0;
00124 
00125     Job<difference_type> *job;
00126 
00127     omp_lock_t output_lock;
00128     omp_init_lock(&output_lock);
00129 
00130     // Write base value to output.
00131     output = base;
00132 
00133     // No more threads than jobs, at least one thread.
00134     thread_index_t num_threads =
00135         __gnu_parallel::max<thread_index_t>(1,
00136             __gnu_parallel::min<difference_type>(length, get_max_threads()));
00137 
00138 #   pragma omp parallel shared(busy) num_threads(num_threads)
00139       {
00140 
00141 #       pragma omp single
00142           {
00143             num_threads = omp_get_num_threads();
00144 
00145             // Create job description array.
00146             job = new Job<difference_type>[num_threads * stride];
00147           }
00148 
00149         // Initialization phase.
00150 
00151         // Flags for every thread if it is doing productive work.
00152         bool iam_working = false;
00153 
00154         // Thread id.
00155         thread_index_t iam = omp_get_thread_num();
00156 
00157         // This job.
00158         Job<difference_type>& my_job = job[iam * stride];
00159 
00160         // Random number (for work stealing).
00161         thread_index_t victim;
00162 
00163         // Local value for reduction.
00164         Result result = Result();
00165 
00166         // Number of elements to steal in one attempt.
00167         difference_type steal;
00168 
00169         // Every thread has its own random number generator
00170         // (modulo num_threads).
00171         random_number rand_gen(iam, num_threads);
00172 
00173         // This thread is currently working.
00174 #       pragma omp atomic
00175           ++busy;
00176 
00177         iam_working = true;
00178 
00179         // How many jobs per thread? last thread gets the rest.
00180         my_job.first =
00181             static_cast<difference_type>(iam * (length / num_threads));
00182 
00183         my_job.last = (iam == (num_threads - 1)) ?
00184             (length - 1) : ((iam + 1) * (length / num_threads) - 1);
00185         my_job.load = my_job.last - my_job.first + 1;
00186 
00187         // Init result with first value (to have a base value for reduction).
00188         if (my_job.first <= my_job.last)
00189           {
00190             // Cannot use volatile variable directly.
00191             difference_type my_first = my_job.first;
00192             result = f(op, begin + my_first);
00193             ++my_job.first;
00194             --my_job.load;
00195           }
00196 
00197         RandomAccessIterator current;
00198 
00199 #       pragma omp barrier
00200 
00201         // Actual work phase
00202         // Work on own or stolen start
00203         while (busy > 0)
00204           {
00205             // Work until no productive thread left.
00206 #           pragma omp flush(busy)
00207 
00208             // Thread has own work to do
00209             while (my_job.first <= my_job.last)
00210               {
00211                 // fetch-and-add call
00212                 // Reserve current job block (size chunk_size) in my queue.
00213                 difference_type current_job =
00214                   fetch_and_add<difference_type>(&(my_job.first), chunk_size);
00215 
00216                 // Update load, to make the three values consistent,
00217                 // first might have been changed in the meantime
00218                 my_job.load = my_job.last - my_job.first + 1;
00219                 for (difference_type job_counter = 0;
00220                      job_counter < chunk_size && current_job <= my_job.last;
00221                      ++job_counter)
00222                   {
00223                     // Yes: process it!
00224                     current = begin + current_job;
00225                     ++current_job;
00226 
00227                     // Do actual work.
00228                     result = r(result, f(op, current));
00229                   }
00230 
00231 #               pragma omp flush(busy)
00232               }
00233 
00234             // After reaching this point, a thread's job list is empty.
00235             if (iam_working)
00236               {
00237                 // This thread no longer has work.
00238 #               pragma omp atomic
00239                 --busy;
00240 
00241                 iam_working = false;
00242               }
00243 
00244             difference_type supposed_first, supposed_last, supposed_load;
00245             do
00246               {
00247                 // Find random nonempty deque (not own), do consistency check.
00248                 yield();
00249 #               pragma omp flush(busy)
00250                 victim = rand_gen();
00251                 supposed_first = job[victim * stride].first;
00252                 supposed_last = job[victim * stride].last;
00253                 supposed_load = job[victim * stride].load;
00254               }
00255             while (busy > 0
00256               && ((supposed_load <= 0)
00257                 || ((supposed_first + supposed_load - 1) != supposed_last)));
00258 
00259             if (busy == 0)
00260               break;
00261 
00262             if (supposed_load > 0)
00263               {
00264                 // Has work and work to do.
00265                 // Number of elements to steal (at least one).
00266                 steal = (supposed_load < 2) ? 1 : supposed_load / 2;
00267 
00268                 // Push victim's start forward.
00269                 difference_type stolen_first =
00270                     fetch_and_add<difference_type>(
00271                         &(job[victim * stride].first), steal);
00272                 difference_type stolen_try =
00273                     stolen_first + steal - difference_type(1);
00274 
00275                 my_job.first = stolen_first;
00276                 my_job.last = __gnu_parallel::min(stolen_try, supposed_last);
00277                 my_job.load = my_job.last - my_job.first + 1;
00278 
00279                 // Has potential work again.
00280 #               pragma omp atomic
00281                   ++busy;
00282                 iam_working = true;
00283 
00284 #               pragma omp flush(busy)
00285               }
00286 #           pragma omp flush(busy)
00287           } // end while busy > 0
00288             // Add accumulated result to output.
00289         omp_set_lock(&output_lock);
00290         output = r(output, result);
00291         omp_unset_lock(&output_lock);
00292       }
00293 
00294     delete[] job;
00295 
00296     // Points to last element processed (needed as return value for
00297     // some algorithms like transform)
00298     f.finish_iterator = begin + length;
00299 
00300     omp_destroy_lock(&output_lock);
00301 
00302     return op;
00303   }
00304 } // end namespace
00305 
00306 #endif /* _GLIBCXX_PARALLEL_WORKSTEALING_H */

Generated on 19 Jun 2018 for libstdc++ by  doxygen 1.6.1