• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_SYMMETRIC_DIFFERENCE_HPP
12 #define BOOST_COMPUTE_ALGORITHM_SET_SYMMETRIC_DIFFERENCE_HPP
13 
14 #include <iterator>
15 
16 #include <boost/static_assert.hpp>
17 
18 #include <boost/compute/algorithm/detail/compact.hpp>
19 #include <boost/compute/algorithm/detail/balanced_path.hpp>
20 #include <boost/compute/algorithm/exclusive_scan.hpp>
21 #include <boost/compute/algorithm/fill_n.hpp>
22 #include <boost/compute/container/vector.hpp>
23 #include <boost/compute/detail/iterator_range_size.hpp>
24 #include <boost/compute/detail/meta_kernel.hpp>
25 #include <boost/compute/system.hpp>
26 #include <boost/compute/type_traits/is_device_iterator.hpp>
27 
28 namespace boost {
29 namespace compute {
30 namespace detail {
31 
32 ///
33 /// \brief Serial set symmetric difference kernel class
34 ///
35 /// Subclass of meta_kernel to perform serial set symmetric
36 /// difference after tiling
37 ///
38 class serial_set_symmetric_difference_kernel : meta_kernel
39 {
40 public:
41     unsigned int tile_size;
42 
serial_set_symmetric_difference_kernel()43     serial_set_symmetric_difference_kernel() : meta_kernel("set_symmetric_difference")
44     {
45         tile_size = 4;
46     }
47 
48     template<class InputIterator1, class InputIterator2,
49              class InputIterator3, class InputIterator4,
50              class OutputIterator1, class OutputIterator2>
set_range(InputIterator1 first1,InputIterator2 first2,InputIterator3 tile_first1,InputIterator3 tile_last1,InputIterator4 tile_first2,OutputIterator1 result,OutputIterator2 counts)51     void set_range(InputIterator1 first1,
52                     InputIterator2 first2,
53                     InputIterator3 tile_first1,
54                     InputIterator3 tile_last1,
55                     InputIterator4 tile_first2,
56                     OutputIterator1 result,
57                     OutputIterator2 counts)
58     {
59         m_count = iterator_range_size(tile_first1, tile_last1) - 1;
60 
61         *this <<
62         "uint i = get_global_id(0);\n" <<
63         "uint start1 = " << tile_first1[expr<uint_>("i")] << ";\n" <<
64         "uint end1 = " << tile_first1[expr<uint_>("i+1")] << ";\n" <<
65         "uint start2 = " << tile_first2[expr<uint_>("i")] << ";\n" <<
66         "uint end2 = " << tile_first2[expr<uint_>("i+1")] << ";\n" <<
67         "uint index = i*" << tile_size << ";\n" <<
68         "uint count = 0;\n" <<
69         "while(start1<end1 && start2<end2)\n" <<
70         "{\n" <<
71         "   if(" << first1[expr<uint_>("start1")] << " == " <<
72                     first2[expr<uint_>("start2")] << ")\n" <<
73         "   {\n" <<
74         "       start1++; start2++;\n" <<
75         "   }\n" <<
76         "   else if(" << first1[expr<uint_>("start1")] << " < " <<
77                         first2[expr<uint_>("start2")] << ")\n" <<
78         "   {\n" <<
79                 result[expr<uint_>("index")] <<
80                     " = " << first1[expr<uint_>("start1")] << ";\n" <<
81         "       index++; count++;\n" <<
82         "       start1++;\n" <<
83         "   }\n" <<
84         "   else\n" <<
85         "   {\n" <<
86                 result[expr<uint_>("index")] <<
87                     " = " << first2[expr<uint_>("start2")] << ";\n" <<
88         "       index++; count++;\n" <<
89         "       start2++;\n" <<
90         "   }\n" <<
91         "}\n" <<
92         "while(start1<end1)\n" <<
93         "{\n" <<
94             result[expr<uint_>("index")] <<
95                 " = " << first1[expr<uint_>("start1")] << ";\n" <<
96         "   index++; count++;\n" <<
97         "   start1++;\n" <<
98         "}\n" <<
99         "while(start2<end2)\n" <<
100         "{\n" <<
101             result[expr<uint_>("index")] <<
102                 " = " << first2[expr<uint_>("start2")] << ";\n" <<
103         "   index++; count++;\n" <<
104         "   start2++;\n" <<
105         "}\n" <<
106         counts[expr<uint_>("i")] << " = count;\n";
107     }
108 
exec(command_queue & queue)109     event exec(command_queue &queue)
110     {
111         if(m_count == 0) {
112             return event();
113         }
114 
115         return exec_1d(queue, 0, m_count);
116     }
117 
118 private:
119     size_t m_count;
120 };
121 
122 } //end detail namespace
123 
124 ///
125 /// \brief Set symmetric difference algorithm
126 ///
127 /// Finds the symmetric difference of the sorted range [first2, last2) from
128 /// the sorted range [first1, last1) and stores it in range starting at result
129 /// \return Iterator pointing to end of symmetric difference
130 ///
131 /// \param first1 Iterator pointing to start of first set
132 /// \param last1 Iterator pointing to end of first set
133 /// \param first2 Iterator pointing to start of second set
134 /// \param last2 Iterator pointing to end of second set
135 /// \param result Iterator pointing to start of range in which the symmetric
136 /// difference will be stored
137 /// \param queue Queue on which to execute
138 ///
139 /// Space complexity:
140 /// \Omega(2(distance(\p first1, \p last1) + distance(\p first2, \p last2)))
141 template<class InputIterator1, class InputIterator2, class OutputIterator>
142 inline OutputIterator
set_symmetric_difference(InputIterator1 first1,InputIterator1 last1,InputIterator2 first2,InputIterator2 last2,OutputIterator result,command_queue & queue=system::default_queue ())143 set_symmetric_difference(InputIterator1 first1,
144                          InputIterator1 last1,
145                          InputIterator2 first2,
146                          InputIterator2 last2,
147                          OutputIterator result,
148                          command_queue &queue = system::default_queue())
149 {
150     BOOST_STATIC_ASSERT(is_device_iterator<InputIterator1>::value);
151     BOOST_STATIC_ASSERT(is_device_iterator<InputIterator2>::value);
152     BOOST_STATIC_ASSERT(is_device_iterator<OutputIterator>::value);
153 
154     typedef typename std::iterator_traits<InputIterator1>::value_type value_type;
155 
156     int tile_size = 1024;
157 
158     int count1 = detail::iterator_range_size(first1, last1);
159     int count2 = detail::iterator_range_size(first2, last2);
160 
161     vector<uint_> tile_a((count1+count2+tile_size-1)/tile_size+1, queue.get_context());
162     vector<uint_> tile_b((count1+count2+tile_size-1)/tile_size+1, queue.get_context());
163 
164     // Tile the sets
165     detail::balanced_path_kernel tiling_kernel;
166     tiling_kernel.tile_size = tile_size;
167     tiling_kernel.set_range(first1, last1, first2, last2,
168                             tile_a.begin()+1, tile_b.begin()+1);
169     fill_n(tile_a.begin(), 1, 0, queue);
170     fill_n(tile_b.begin(), 1, 0, queue);
171     tiling_kernel.exec(queue);
172 
173     fill_n(tile_a.end()-1, 1, count1, queue);
174     fill_n(tile_b.end()-1, 1, count2, queue);
175 
176     vector<value_type> temp_result(count1+count2, queue.get_context());
177     vector<uint_> counts((count1+count2+tile_size-1)/tile_size + 1, queue.get_context());
178     fill_n(counts.end()-1, 1, 0, queue);
179 
180     // Find individual symmetric differences
181     detail::serial_set_symmetric_difference_kernel symmetric_difference_kernel;
182     symmetric_difference_kernel.tile_size = tile_size;
183     symmetric_difference_kernel.set_range(first1, first2, tile_a.begin(),
184                                             tile_a.end(), tile_b.begin(),
185                                             temp_result.begin(), counts.begin());
186 
187     symmetric_difference_kernel.exec(queue);
188 
189     exclusive_scan(counts.begin(), counts.end(), counts.begin(), queue);
190 
191     // Compact the results
192     detail::compact_kernel compact_kernel;
193     compact_kernel.tile_size = tile_size;
194     compact_kernel.set_range(temp_result.begin(), counts.begin(), counts.end(), result);
195 
196     compact_kernel.exec(queue);
197 
198     return result + (counts.end() - 1).read(queue);
199 }
200 
201 } //end compute namespace
202 } //end boost namespace
203 
204 #endif // BOOST_COMPUTE_ALGORITHM_SET_SYMMETRIC_DIFFERENCE_HPP
205