Home | History | Annotate | Download | only in parallel
      1 // -*- C++ -*-
      2 
      3 // Copyright (C) 2007, 2008, 2009 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 3, 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 // Under Section 7 of GPL version 3, you are granted additional
     17 // permissions described in the GCC Runtime Library Exception, version
     18 // 3.1, as published by the Free Software Foundation.
     19 
     20 // You should have received a copy of the GNU General Public License and
     21 // a copy of the GCC Runtime Library Exception along with this program;
     22 // see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
     23 // <http://www.gnu.org/licenses/>.
     24 
     25 /** @file parallel/list_partition.h
     26  *  @brief Functionality to split sequence referenced by only input
     27  *  iterators.
     28  *  This file is a GNU parallel extension to the Standard C++ Library.
     29  */
     30 
     31 // Written by Leonor Frias Moya and Johannes Singler.
     32 
     33 #ifndef _GLIBCXX_PARALLEL_LIST_PARTITION_H
     34 #define _GLIBCXX_PARALLEL_LIST_PARTITION_H 1
     35 
     36 #include <parallel/parallel.h>
     37 #include <vector>
     38 
     39 namespace __gnu_parallel
     40 {
     41   /** @brief Shrinks and doubles the ranges.
     42    *  @param os_starts Start positions worked on (oversampled).
     43    *  @param count_to_two Counts up to 2.
     44    *  @param range_length Current length of a chunk.
     45    *  @param make_twice Whether the @c os_starts is allowed to be
     46    *  grown or not
     47    */
     48   template<typename InputIterator>
     49     void
     50     shrink_and_double(std::vector<InputIterator>& os_starts,
     51 		      size_t& count_to_two, size_t& range_length,
     52 		      const bool make_twice)
     53     {
     54       ++count_to_two;
     55       if (not make_twice or count_to_two < 2)
     56 	shrink(os_starts, count_to_two, range_length);
     57       else
     58 	{
     59 	  os_starts.resize((os_starts.size() - 1) * 2 + 1);
     60 	  count_to_two = 0;
     61 	}
     62     }
     63 
     64   /** @brief Combines two ranges into one and thus halves the number of ranges.
     65    *  @param os_starts Start positions worked on (oversampled).
     66    *  @param count_to_two Counts up to 2.
     67    *  @param range_length Current length of a chunk. */
     68   template<typename InputIterator>
     69     void
     70     shrink(std::vector<InputIterator>& os_starts, size_t& count_to_two,
     71 	   size_t& range_length)
     72     {
     73       for (typename std::vector<InputIterator>::size_type i = 0;
     74 	   i <= (os_starts.size() / 2); ++i)
     75 	os_starts[i] = os_starts[i * 2];
     76       range_length *= 2;
     77     }
     78 
     79   /** @brief Splits a sequence given by input iterators into parts of
     80    * almost equal size
     81    *
     82    *  The function needs only one pass over the sequence.
     83    *  @param begin Begin iterator of input sequence.
     84    *  @param end End iterator of input sequence.
     85    *  @param starts Start iterators for the resulting parts, dimension
     86    *  @c num_parts+1. For convenience, @c starts @c [num_parts]
     87    *  contains the end iterator of the sequence.
     88    *  @param lengths Length of the resulting parts.
     89    *  @param num_parts Number of parts to split the sequence into.
     90    *  @param f Functor to be applied to each element by traversing it
     91    *  @param oversampling Oversampling factor. If 0, then the
     92    *  partitions will differ in at most @f$ \sqrt{\mathrm{end} -
     93    *  \mathrm{begin}} @f$ elements. Otherwise, the ratio between the
     94    *  longest and the shortest part is bounded by @f$
     95    *  1/(\mathrm{oversampling} \cdot \mathrm{num\_parts}) @f$.
     96    *  @return Length of the whole sequence.
     97    */
     98   template<typename InputIterator, typename FunctorType>
     99     size_t
    100     list_partition(const InputIterator begin, const InputIterator end,
    101 		   InputIterator* starts, size_t* lengths, const int num_parts,
    102 		   FunctorType& f, int oversampling = 0)
    103     {
    104       bool make_twice = false;
    105 
    106       // The resizing algorithm is chosen according to the oversampling factor.
    107       if (oversampling == 0)
    108 	{
    109 	  make_twice = true;
    110 	  oversampling = 1;
    111 	}
    112 
    113       std::vector<InputIterator> os_starts(2 * oversampling * num_parts + 1);
    114 
    115       os_starts[0]= begin;
    116       InputIterator prev = begin, it = begin;
    117       size_t dist_limit = 0, dist = 0;
    118       size_t cur = 1, next = 1;
    119       size_t range_length = 1;
    120       size_t count_to_two = 0;
    121       while (it != end)
    122 	{
    123 	  cur = next;
    124 	  for (; cur < os_starts.size() and it != end; ++cur)
    125 	    {
    126 	      for (dist_limit += range_length;
    127 		   dist < dist_limit and it != end; ++dist)
    128 		{
    129 		  f(it);
    130 		  ++it;
    131 		}
    132 	      os_starts[cur] = it;
    133 	    }
    134 
    135 	  // Must compare for end and not cur < os_starts.size() , because
    136 	  // cur could be == os_starts.size() as well
    137 	  if (it == end)
    138 	    break;
    139 
    140 	  shrink_and_double(os_starts, count_to_two, range_length, make_twice);
    141 	  next = os_starts.size() / 2 + 1;
    142 	}
    143 
    144       // Calculation of the parts (one must be extracted from current
    145       // because the partition beginning at end, consists only of
    146       // itself).
    147       size_t size_part = (cur - 1) / num_parts;
    148       int size_greater = static_cast<int>((cur - 1) % num_parts);
    149       starts[0] = os_starts[0];
    150 
    151       size_t index = 0;
    152 
    153       // Smallest partitions.
    154       for (int i = 1; i < (num_parts + 1 - size_greater); ++i)
    155 	{
    156 	  lengths[i - 1] =  size_part * range_length;
    157 	  index += size_part;
    158 	  starts[i] = os_starts[index];
    159 	}
    160 
    161       // Biggest partitions.
    162       for (int i = num_parts + 1 - size_greater; i <= num_parts; ++i)
    163 	{
    164 	  lengths[i - 1] =  (size_part+1) * range_length;
    165 	  index += (size_part+1);
    166 	  starts[i] = os_starts[index];
    167 	}
    168 
    169       // Correction of the end size (the end iteration has not finished).
    170       lengths[num_parts - 1] -= (dist_limit - dist);
    171 
    172       return dist;
    173     }
    174 }
    175 
    176 #endif /* _GLIBCXX_PARALLEL_LIST_PARTITION_H */
    177