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