Name	Size	Mode	Actions
balanced_path.hpp	6267	0644	editdlrm
binary_find.hpp	4626	0644	editdlrm
compact.hpp	2188	0644	editdlrm
copy_on_device.hpp	7382	0644	editdlrm
copy_to_device.hpp	6731	0644	editdlrm
copy_to_host.hpp	6658	0644	editdlrm
count_if_with_ballot.hpp	2605	0644	editdlrm
count_if_with_reduce.hpp	2648	0644	editdlrm
count_if_with_threads.hpp	4241	0644	editdlrm
find_extrema.hpp	2441	0644	editdlrm
find_extrema_on_cpu.hpp	5333	0644	editdlrm
find_extrema_with_atomics.hpp	4251	0644	editdlrm
find_extrema_with_reduce.hpp	18672	0644	editdlrm
find_if_with_atomics.hpp	8509	0644	editdlrm
inplace_reduce.hpp	4839	0644	editdlrm
insertion_sort.hpp	6030	0644	editdlrm
merge_path.hpp	3874	0644	editdlrm
merge_sort_on_cpu.hpp	14654	0644	editdlrm
merge_sort_on_gpu.hpp	22093	0644	editdlrm
merge_with_merge_path.hpp	7339	0644	editdlrm
radix_sort.hpp	15502	0644	editdlrm
random_fill.hpp	1880	0644	editdlrm
reduce_by_key.hpp	5243	0644	editdlrm
reduce_by_key_with_scan.hpp	23735	0644	editdlrm
reduce_on_cpu.hpp	3949	0644	editdlrm
reduce_on_gpu.hpp	10391	0644	editdlrm
scan.hpp	1545	0644	editdlrm
scan_on_cpu.hpp	7056	0644	editdlrm
scan_on_gpu.hpp	11320	0644	editdlrm
search_all.hpp	2579	0644	editdlrm
serial_accumulate.hpp	1990	0644	editdlrm
serial_count_if.hpp	2243	0644	editdlrm
serial_find_extrema.hpp	3151	0644	editdlrm
serial_merge.hpp	3504	0644	editdlrm
serial_reduce.hpp	2145	0644	editdlrm
serial_reduce_by_key.hpp	4240	0644	editdlrm
serial_scan.hpp	3184	0644	editdlrm

//---------------------------------------------------------------------------// // Copyright (c) 2014 Roshan // // Distributed under the Boost Software License, Version 1.0 // See accompanying file LICENSE_1_0.txt or copy at // http://www.boost.org/LICENSE_1_0.txt // // See http://boostorg.github.com/compute for more information. //---------------------------------------------------------------------------// #ifndef BOOST_COMPUTE_ALGORITHM_DETAIL_BALANCED_PATH_HPP #define BOOST_COMPUTE_ALGORITHM_DETAIL_BALANCED_PATH_HPP #include #include #include #include #include #include #include namespace boost { namespace compute { namespace detail { /// /// \brief Balanced Path kernel class /// /// Subclass of meta_kernel to break two sets into tiles according /// to their balanced path. /// class balanced_path_kernel : public meta_kernel { public: unsigned int tile_size; balanced_path_kernel() : meta_kernel("balanced_path") { tile_size = 4; } template void set_range(InputIterator1 first1, InputIterator1 last1, InputIterator2 first2, InputIterator2 last2, OutputIterator1 result_a, OutputIterator2 result_b, Compare comp) { typedef typename std::iterator_traits::value_type value_type; m_a_count = iterator_range_size(first1, last1); m_a_count_arg = add_arg("a_count"); m_b_count = iterator_range_size(first2, last2); m_b_count_arg = add_arg("b_count"); *this << "uint i = get_global_id(0);\n" << "uint target = (i+1)*" << tile_size << ";\n" << "uint start = max(convert_int(0),convert_int(target)-convert_int(b_count));\n" << "uint end = min(target,a_count);\n" << "uint a_index, b_index;\n" << "while(start("b_index")], first1[expr("a_index")]) << "))\n" << " start = a_index + 1;\n" << " else end = a_index;\n" << "}\n" << "a_index = start;\n" << "b_index = target - start;\n" << "if(b_index < b_count)\n" << "{\n" << " " << decl("x") << " = " << first2[expr("b_index")] << ";\n" << " uint a_start = 0, a_end = a_index, a_mid;\n" << " uint b_start = 0, b_end = b_index, b_mid;\n" << " while(a_start("a_mid")], expr("x")) << ")\n" << " a_start = a_mid+1;\n" << " else a_end = a_mid;\n" << " }\n" << " while(b_start("b_mid")], expr("x")) << ")\n" << " b_start = b_mid+1;\n" << " else b_end = b_mid;\n" << " }\n" << " uint a_run = a_index - a_start;\n" << " uint b_run = b_index - b_start;\n" << " uint x_count = a_run + b_run;\n" << " uint b_advance = max(x_count / 2, x_count - a_run);\n" << " b_end = min(b_count, b_start + b_advance + 1);\n" << " uint temp_start = b_index, temp_end = b_end, temp_mid;" << " while(temp_start < temp_end)\n" << " {\n" << " temp_mid = (temp_start + temp_end + 1)/2;\n" << " if(" << comp(expr("x"), first2[expr("temp_mid")]) << ")\n" << " temp_end = temp_mid-1;\n" << " else temp_start = temp_mid;\n" << " }\n" << " b_run = temp_start - b_start + 1;\n" << " b_advance = min(b_advance, b_run);\n" << " uint a_advance = x_count - b_advance;\n" << " uint star = convert_uint((a_advance == b_advance + 1) " << "&& (b_advance < b_run));\n" << " a_index = a_start + a_advance;\n" << " b_index = target - a_index + star;\n" << "}\n" << result_a[expr("i")] << " = a_index;\n" << result_b[expr("i")] << " = b_index;\n"; } template void set_range(InputIterator1 first1, InputIterator1 last1, InputIterator2 first2, InputIterator2 last2, OutputIterator1 result_a, OutputIterator2 result_b) { typedef typename std::iterator_traits::value_type value_type; ::boost::compute::less less_than; set_range(first1, last1, first2, last2, result_a, result_b, less_than); } event exec(command_queue &queue) { if((m_a_count + m_b_count)/tile_size == 0) { return event(); } set_arg(m_a_count_arg, uint_(m_a_count)); set_arg(m_b_count_arg, uint_(m_b_count)); return exec_1d(queue, 0, (m_a_count + m_b_count)/tile_size); } private: size_t m_a_count; size_t m_a_count_arg; size_t m_b_count; size_t m_b_count_arg; }; } //end detail namespace } //end compute namespace } //end boost namespace #endif // BOOST_COMPUTE_ALGORITHM_DETAIL_BALANCED_PATH_HPP