1 //---------------------------------------------------------------------------//
2 // Copyright (c) 2013 Kyle Lutz <kyle.r.lutz@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_DETAIL_SERIAL_SCAN_HPP
12 #define BOOST_COMPUTE_ALGORITHM_DETAIL_SERIAL_SCAN_HPP
13
14 #include <iterator>
15
16 #include <boost/compute/device.hpp>
17 #include <boost/compute/kernel.hpp>
18 #include <boost/compute/command_queue.hpp>
19 #include <boost/compute/detail/meta_kernel.hpp>
20 #include <boost/compute/detail/iterator_range_size.hpp>
21
22 namespace boost {
23 namespace compute {
24 namespace detail {
25
26 template<class InputIterator, class OutputIterator, class T, class BinaryOperator>
serial_scan(InputIterator first,InputIterator last,OutputIterator result,bool exclusive,T init,BinaryOperator op,command_queue & queue)27 inline OutputIterator serial_scan(InputIterator first,
28 InputIterator last,
29 OutputIterator result,
30 bool exclusive,
31 T init,
32 BinaryOperator op,
33 command_queue &queue)
34 {
35 if(first == last){
36 return result;
37 }
38
39 typedef typename
40 std::iterator_traits<InputIterator>::value_type input_type;
41 typedef typename
42 std::iterator_traits<OutputIterator>::value_type output_type;
43
44 const context &context = queue.get_context();
45
46 // create scan kernel
47 meta_kernel k("serial_scan");
48
49 // Arguments
50 size_t n_arg = k.add_arg<ulong_>("n");
51 size_t init_arg = k.add_arg<output_type>("initial_value");
52
53 if(!exclusive){
54 k <<
55 k.decl<const ulong_>("start_idx") << " = 1;\n" <<
56 k.decl<output_type>("sum") << " = " << first[0] << ";\n" <<
57 result[0] << " = sum;\n";
58 }
59 else {
60 k <<
61 k.decl<const ulong_>("start_idx") << " = 0;\n" <<
62 k.decl<output_type>("sum") << " = initial_value;\n";
63 }
64
65 k <<
66 "for(ulong i = start_idx; i < n; i++){\n" <<
67 k.decl<const input_type>("x") << " = "
68 << first[k.var<ulong_>("i")] << ";\n";
69
70 if(exclusive){
71 k << result[k.var<ulong_>("i")] << " = sum;\n";
72 }
73
74 k << " sum = "
75 << op(k.var<output_type>("sum"), k.var<output_type>("x"))
76 << ";\n";
77
78 if(!exclusive){
79 k << result[k.var<ulong_>("i")] << " = sum;\n";
80 }
81
82 k << "}\n";
83
84 // compile scan kernel
85 kernel scan_kernel = k.compile(context);
86
87 // setup kernel arguments
88 size_t n = detail::iterator_range_size(first, last);
89 scan_kernel.set_arg<ulong_>(n_arg, n);
90 scan_kernel.set_arg<output_type>(init_arg, static_cast<output_type>(init));
91
92 // execute the kernel
93 queue.enqueue_1d_range_kernel(scan_kernel, 0, 1, 1);
94
95 // return iterator pointing to the end of the result range
96 return result + n;
97 }
98
99 } // end detail namespace
100 } // end compute namespace
101 } // end boost namespace
102
103 #endif // BOOST_COMPUTE_ALGORITHM_DETAIL_SERIAL_SCAN_HPP
104