algorithm/detail/merge_path.hpp

0001 //---------------------------------------------------------------------------//
0002 // Copyright (c) 2014 Roshan <thisisroshansmail@gmail.com>
0003 //
0004 // Distributed under the Boost Software License, Version 1.0
0005 // See accompanying file LICENSE_1_0.txt or copy at
0006 // http://www.boost.org/LICENSE_1_0.txt
0007 //
0008 // See http://boostorg.github.com/compute for more information.
0009 //---------------------------------------------------------------------------//
0010
0011 #ifndef BOOST_COMPUTE_ALGORITHM_DETAIL_MERGE_PATH_HPP
0012 #define BOOST_COMPUTE_ALGORITHM_DETAIL_MERGE_PATH_HPP
0013
0014 #include <iterator>
0015
0016 #include <boost/compute/algorithm/find_if.hpp>
0017 #include <boost/compute/container/vector.hpp>
0018 #include <boost/compute/detail/iterator_range_size.hpp>
0019 #include <boost/compute/detail/meta_kernel.hpp>
0020 #include <boost/compute/lambda.hpp>
0021 #include <boost/compute/system.hpp>
0022
0023 namespace boost {
0024 namespace compute {
0025 namespace detail {
0026
0027 ///
0028 /// \brief Merge Path kernel class
0029 ///
0030 /// Subclass of meta_kernel to break two sets into tiles according
0031 /// to their merge path
0032 ///
0033 class merge_path_kernel : public meta_kernel
0034 {
0035 public:
0036     unsigned int tile_size;
0037
0038     merge_path_kernel() : meta_kernel("merge_path")
0039     {
0040         tile_size = 4;
0041     }
0042
0043     template<class InputIterator1, class InputIterator2,
0044              class OutputIterator1, class OutputIterator2,
0045              class Compare>
0046     void set_range(InputIterator1 first1,
0047                    InputIterator1 last1,
0048                    InputIterator2 first2,
0049                    InputIterator2 last2,
0050                    OutputIterator1 result_a,
0051                    OutputIterator2 result_b,
0052                    Compare comp)
0053     {
0054         m_a_count = iterator_range_size(first1, last1);
0055         m_a_count_arg = add_arg<uint_>("a_count");
0056
0057         m_b_count = iterator_range_size(first2, last2);
0058         m_b_count_arg = add_arg<uint_>("b_count");
0059
0060         *this <<
0061             "uint i = get_global_id(0);\n" <<
0062             "uint target = (i+1)*" << tile_size << ";\n" <<
0063             "uint start = max(convert_int(0),convert_int(target)-convert_int(b_count));\n" <<
0064             "uint end = min(target,a_count);\n" <<
0065             "uint a_index, b_index;\n" <<
0066             "while(start<end)\n" <<
0067             "{\n" <<
0068             "   a_index = (start + end)/2;\n" <<
0069             "   b_index = target - a_index - 1;\n" <<
0070             "   if(!(" << comp(first2[expr<uint_>("b_index")],
0071                               first1[expr<uint_>("a_index")]) << "))\n" <<
0072             "       start = a_index + 1;\n" <<
0073             "   else end = a_index;\n" <<
0074             "}\n" <<
0075             result_a[expr<uint_>("i")] << " = start;\n" <<
0076             result_b[expr<uint_>("i")] << " = target - start;\n";
0077     }
0078
0079     template<class InputIterator1, class InputIterator2,
0080              class OutputIterator1, class OutputIterator2>
0081     void set_range(InputIterator1 first1,
0082                    InputIterator1 last1,
0083                    InputIterator2 first2,
0084                    InputIterator2 last2,
0085                    OutputIterator1 result_a,
0086                    OutputIterator2 result_b)
0087     {
0088         typedef typename std::iterator_traits<InputIterator1>::value_type value_type;
0089         ::boost::compute::less<value_type> less_than;
0090         set_range(first1, last1, first2, last2, result_a, result_b, less_than);
0091     }
0092
0093     event exec(command_queue &queue)
0094     {
0095         if((m_a_count + m_b_count)/tile_size == 0) {
0096             return event();
0097         }
0098
0099         set_arg(m_a_count_arg, uint_(m_a_count));
0100         set_arg(m_b_count_arg, uint_(m_b_count));
0101
0102         return exec_1d(queue, 0, (m_a_count + m_b_count)/tile_size);
0103     }
0104
0105 private:
0106     size_t m_a_count;
0107     size_t m_a_count_arg;
0108     size_t m_b_count;
0109     size_t m_b_count_arg;
0110 };
0111
0112 } //end detail namespace
0113 } //end compute namespace
0114 } //end boost namespace
0115
0116 #endif // BOOST_COMPUTE_ALGORITHM_DETAIL_MERGE_PATH_HPP