ceph/src/boost/libs/compute/include/boost/compute/algorithm/set_union.hpp

   1 //---------------------------------------------------------------------------//
   2 // Copyright (c) 2014 Roshan <thisisroshansmail@gmail.com>
   3 //
   4 // Distributed under the Boost Software License, Version 1.0
   5 // See accompanying file LICENSE_1_0.txt or copy at
   6 // http://www.boost.org/LICENSE_1_0.txt
   7 //
   8 // See http://boostorg.github.com/compute for more information.
   9 //---------------------------------------------------------------------------//
  10
  11 #ifndef BOOST_COMPUTE_ALGORITHM_SET_UNION_HPP
  12 #define BOOST_COMPUTE_ALGORITHM_SET_UNION_HPP
  13
  14 #include <iterator>
  15
  16 #include <boost/compute/algorithm/detail/balanced_path.hpp>
  17 #include <boost/compute/algorithm/detail/compact.hpp>
  18 #include <boost/compute/algorithm/exclusive_scan.hpp>
  19 #include <boost/compute/algorithm/fill_n.hpp>
  20 #include <boost/compute/container/vector.hpp>
  21 #include <boost/compute/detail/iterator_range_size.hpp>
  22 #include <boost/compute/detail/meta_kernel.hpp>
  23 #include <boost/compute/system.hpp>
  24
  25 namespace boost {
  26 namespace compute {
  27 namespace detail {
  28
  29 ///
  30 /// \brief Serial set union kernel class
  31 ///
  32 /// Subclass of meta_kernel to perform serial set union after tiling
  33 ///
  34 class serial_set_union_kernel : meta_kernel
  35 {
  36 public:
  37     unsigned int tile_size;
  38
  39     serial_set_union_kernel() : meta_kernel("set_union")
  40     {
  41         tile_size = 4;
  42     }
  43
  44     template<class InputIterator1, class InputIterator2,
  45              class InputIterator3, class InputIterator4,
  46              class OutputIterator1, class OutputIterator2>
  47     void set_range(InputIterator1 first1,
  48                     InputIterator2 first2,
  49                     InputIterator3 tile_first1,
  50                     InputIterator3 tile_last1,
  51                     InputIterator4 tile_first2,
  52                     OutputIterator1 result,
  53                     OutputIterator2 counts)
  54     {
  55         m_count = iterator_range_size(tile_first1, tile_last1) - 1;
  56
  57         *this <<
  58         "uint i = get_global_id(0);\n" <<
  59         "uint start1 = " << tile_first1[expr<uint_>("i")] << ";\n" <<
  60         "uint end1 = " << tile_first1[expr<uint_>("i+1")] << ";\n" <<
  61         "uint start2 = " << tile_first2[expr<uint_>("i")] << ";\n" <<
  62         "uint end2 = " << tile_first2[expr<uint_>("i+1")] << ";\n" <<
  63         "uint index = i*" << tile_size << ";\n" <<
  64         "uint count = 0;\n" <<
  65         "while(start1<end1 && start2<end2)\n" <<
  66         "{\n" <<
  67         "   if(" << first1[expr<uint_>("start1")] << " == " <<
  68                     first2[expr<uint_>("start2")] << ")\n" <<
  69         "   {\n" <<
  70                 result[expr<uint_>("index")] <<
  71                     " = " << first1[expr<uint_>("start1")] << ";\n" <<
  72         "       index++; count++;\n" <<
  73         "       start1++; start2++;\n" <<
  74         "   }\n" <<
  75         "   else if(" << first1[expr<uint_>("start1")] << " < " <<
  76                         first2[expr<uint_>("start2")] << ")\n" <<
  77         "   {\n" <<
  78                 result[expr<uint_>("index")] <<
  79                     " = " << first1[expr<uint_>("start1")] << ";\n" <<
  80         "       index++; count++;\n" <<
  81         "       start1++;\n" <<
  82         "   }\n" <<
  83         "   else\n" <<
  84         "   {\n" <<
  85                 result[expr<uint_>("index")] <<
  86                     " = " << first2[expr<uint_>("start2")] << ";\n" <<
  87         "       index++; count++;\n" <<
  88         "       start2++;\n" <<
  89         "   }\n" <<
  90         "}\n" <<
  91         "while(start1<end1)\n" <<
  92         "{\n" <<
  93             result[expr<uint_>("index")] <<
  94                 " = " << first1[expr<uint_>("start1")] << ";\n" <<
  95         "   index++; count++;\n" <<
  96         "   start1++;\n" <<
  97         "}\n" <<
  98         "while(start2<end2)\n" <<
  99         "{\n" <<
 100             result[expr<uint_>("index")] <<
 101                 " = " << first2[expr<uint_>("start2")] << ";\n" <<
 102         "   index++; count++;\n" <<
 103         "   start2++;\n" <<
 104         "}\n" <<
 105         counts[expr<uint_>("i")] << " = count;\n";
 106     }
 107
 108     event exec(command_queue &queue)
 109     {
 110         if(m_count == 0) {
 111             return event();
 112         }
 113
 114         return exec_1d(queue, 0, m_count);
 115     }
 116
 117 private:
 118     size_t m_count;
 119 };
 120
 121 } //end detail namespace
 122
 123 ///
 124 /// \brief Set union algorithm
 125 ///
 126 /// Finds the union of the sorted range [first1, last1) with the sorted
 127 /// range [first2, last2) and stores it in range starting at result
 128 /// \return Iterator pointing to end of union
 129 ///
 130 /// \param first1 Iterator pointing to start of first set
 131 /// \param last1 Iterator pointing to end of first set
 132 /// \param first2 Iterator pointing to start of second set
 133 /// \param last2 Iterator pointing to end of second set
 134 /// \param result Iterator pointing to start of range in which the union
 135 /// will be stored
 136 /// \param queue Queue on which to execute
 137 ///
 138 template<class InputIterator1, class InputIterator2, class OutputIterator>
 139 inline OutputIterator set_union(InputIterator1 first1,
 140                                 InputIterator1 last1,
 141                                 InputIterator2 first2,
 142                                 InputIterator2 last2,
 143                                 OutputIterator result,
 144                                 command_queue &queue = system::default_queue())
 145 {
 146     typedef typename std::iterator_traits<InputIterator1>::value_type value_type;
 147
 148     int tile_size = 1024;
 149
 150     int count1 = detail::iterator_range_size(first1, last1);
 151     int count2 = detail::iterator_range_size(first2, last2);
 152
 153     vector<uint_> tile_a((count1+count2+tile_size-1)/tile_size+1, queue.get_context());
 154     vector<uint_> tile_b((count1+count2+tile_size-1)/tile_size+1, queue.get_context());
 155
 156     // Tile the sets
 157     detail::balanced_path_kernel tiling_kernel;
 158     tiling_kernel.tile_size = tile_size;
 159     tiling_kernel.set_range(first1, last1, first2, last2,
 160                             tile_a.begin()+1, tile_b.begin()+1);
 161     fill_n(tile_a.begin(), 1, 0, queue);
 162     fill_n(tile_b.begin(), 1, 0, queue);
 163     tiling_kernel.exec(queue);
 164
 165     fill_n(tile_a.end()-1, 1, count1, queue);
 166     fill_n(tile_b.end()-1, 1, count2, queue);
 167
 168     vector<value_type> temp_result(count1+count2, queue.get_context());
 169     vector<uint_> counts((count1+count2+tile_size-1)/tile_size + 1, queue.get_context());
 170     fill_n(counts.end()-1, 1, 0, queue);
 171
 172     // Find individual unions
 173     detail::serial_set_union_kernel union_kernel;
 174     union_kernel.tile_size = tile_size;
 175     union_kernel.set_range(first1, first2, tile_a.begin(), tile_a.end(),
 176                                   tile_b.begin(), temp_result.begin(), counts.begin());
 177
 178     union_kernel.exec(queue);
 179
 180     exclusive_scan(counts.begin(), counts.end(), counts.begin(), queue);
 181
 182     // Compact the results
 183     detail::compact_kernel compact_kernel;
 184     compact_kernel.tile_size = tile_size;
 185     compact_kernel.set_range(temp_result.begin(), counts.begin(), counts.end(), result);
 186
 187     compact_kernel.exec(queue);
 188
 189     return result + (counts.end() - 1).read(queue);
 190 }
 191
 192 } //end compute namespace
 193 } //end boost namespace
 194
 195 #endif // BOOST_COMPUTE_ALGORITHM_SET_UNION_HPP