libstdc++-v3/include/parallel/partial_sum.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/partial_sum.h
  32  *  @brief Parallel implementation of std::partial_sum(), i. e. prefix
  33  *  sums.
  34  *  This file is a GNU parallel extension to the Standard C++ Library.
  35  */
  36
  37 // Written by Johannes Singler.
  38
  39 #ifndef _GLIBCXX_PARALLEL_PARTIAL_SUM_H
  40 #define _GLIBCXX_PARALLEL_PARTIAL_SUM_H 1
  41
  42 #include <omp.h>
  43 #include <bits/stl_algobase.h>
  44 #include <parallel/parallel.h>
  45 #include <parallel/numericfwd.h>
  46
  47 namespace __gnu_parallel
  48 {
  49   // Problem: there is no 0-element given.
  50
  51 /** @brief Base case prefix sum routine.
  52   *  @param begin Begin iterator of input sequence.
  53   *  @param end End iterator of input sequence.
  54   *  @param result Begin iterator of output sequence.
  55   *  @param bin_op Associative binary function.
  56   *  @param value Start value. Must be passed since the neutral
  57   *  element is unknown in general.
  58   *  @return End iterator of output sequence. */
  59 template<
  60     typename InputIterator,
  61     typename OutputIterator,
  62     typename BinaryOperation>
  63   inline OutputIterator
  64   parallel_partial_sum_basecase(
  65             InputIterator begin, InputIterator end,
  66             OutputIterator result, BinaryOperation bin_op,
  67             typename std::iterator_traits<InputIterator>::value_type value)
  68   {
  69     if (begin == end)
  70       return result;
  71
  72     while (begin != end)
  73       {
  74         value = bin_op(value, *begin);
  75         *result = value;
  76         result++;
  77         begin++;
  78       }
  79     return result;
  80   }
  81
  82 /** @brief Parallel partial sum implementation, two-phase approach,
  83     no recursion.
  84     *  @param begin Begin iterator of input sequence.
  85     *  @param end End iterator of input sequence.
  86     *  @param result Begin iterator of output sequence.
  87     *  @param bin_op Associative binary function.
  88     *  @param n Length of sequence.
  89     *  @param num_threads Number of threads to use.
  90     *  @return End iterator of output sequence.
  91     */
  92 template<
  93     typename InputIterator,
  94     typename OutputIterator,
  95     typename BinaryOperation>
  96   OutputIterator
  97   parallel_partial_sum_linear(
  98             InputIterator begin, InputIterator end,
  99             OutputIterator result, BinaryOperation bin_op,
 100             typename std::iterator_traits<InputIterator>::difference_type n)
 101   {
 102     typedef std::iterator_traits<InputIterator> traits_type;
 103     typedef typename traits_type::value_type value_type;
 104     typedef typename traits_type::difference_type difference_type;
 105
 106     thread_index_t num_threads =
 107         std::min<difference_type>(get_max_threads(), n - 1);
 108
 109     if (num_threads < 2)
 110       {
 111         *result = *begin;
 112         return parallel_partial_sum_basecase(
 113             begin + 1, end, result + 1, bin_op, *begin);
 114       }
 115
 116     difference_type* borders;
 117     value_type* sums;
 118
 119 #   pragma omp parallel num_threads(num_threads)
 120       {
 121 #       pragma omp single
 122           {
 123             num_threads = omp_get_num_threads();
 124
 125             borders = new difference_type[num_threads + 2];
 126
 127             if (Settings::partial_sum_dilatation == 1.0f)
 128               equally_split(n, num_threads + 1, borders);
 129             else
 130               {
 131                 difference_type chunk_length =
 132                     ((double)n /
 133                     ((double)num_threads + Settings::partial_sum_dilatation)),
 134                     borderstart = n - num_threads * chunk_length;
 135                 borders[0] = 0;
 136                 for (int i = 1; i < (num_threads + 1); i++)
 137                   {
 138                     borders[i] = borderstart;
 139                     borderstart += chunk_length;
 140                   }
 141                 borders[num_threads + 1] = n;
 142               }
 143
 144             sums = static_cast<value_type*>(
 145                 ::operator new(sizeof(value_type) * num_threads));
 146             OutputIterator target_end;
 147           } //single
 148
 149         int iam = omp_get_thread_num();
 150         if (iam == 0)
 151           {
 152             *result = *begin;
 153             parallel_partial_sum_basecase(begin + 1, begin + borders[1],
 154                           result + 1, bin_op, *begin);
 155             sums[0] = *(result + borders[1] - 1);
 156           }
 157         else
 158           {
 159             sums[iam] = std::accumulate(begin + borders[iam] + 1,
 160                             begin + borders[iam + 1],
 161                             *(begin + borders[iam]),
 162                             bin_op, __gnu_parallel::sequential_tag());
 163           }
 164
 165 #       pragma omp barrier
 166
 167 #       pragma omp single
 168           parallel_partial_sum_basecase(
 169               sums + 1, sums + num_threads, sums + 1, bin_op, sums[0]);
 170
 171 #       pragma omp barrier
 172
 173         // Still same team.
 174         parallel_partial_sum_basecase(begin + borders[iam + 1],
 175                       begin + borders[iam + 2],
 176                       result + borders[iam + 1], bin_op,
 177                       sums[iam]);
 178       } //parallel
 179
 180     delete[] sums;
 181     delete[] borders;
 182
 183     return result + n;
 184   }
 185
 186 /** @brief Parallel partial sum front-end.
 187   *  @param begin Begin iterator of input sequence.
 188   *  @param end End iterator of input sequence.
 189   *  @param result Begin iterator of output sequence.
 190   *  @param bin_op Associative binary function.
 191   *  @return End iterator of output sequence. */
 192 template<
 193     typename InputIterator,
 194     typename OutputIterator,
 195     typename BinaryOperation>
 196   OutputIterator
 197   parallel_partial_sum(InputIterator begin, InputIterator end,
 198                        OutputIterator result, BinaryOperation bin_op)
 199   {
 200     _GLIBCXX_CALL(begin - end)
 201
 202     typedef std::iterator_traits<InputIterator> traits_type;
 203     typedef typename traits_type::value_type value_type;
 204     typedef typename traits_type::difference_type difference_type;
 205
 206     difference_type n = end - begin;
 207
 208     switch (Settings::partial_sum_algorithm)
 209       {
 210       case Settings::LINEAR:
 211         // Need an initial offset.
 212         return parallel_partial_sum_linear(begin, end, result, bin_op, n);
 213       default:
 214     // Partial_sum algorithm not implemented.
 215         _GLIBCXX_PARALLEL_ASSERT(0);
 216         return result + n;
 217       }
 218   }
 219 }
 220
 221 #endif