1 //---------------------------------------------------------------------------//
2 // Copyright (c) 2013-2014 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 #include <algorithm>
12 #include <iostream>
13 #include <numeric>
14 #include <vector>
15
16 #include <boost/program_options.hpp>
17
18 #include <boost/compute/system.hpp>
19 #include <boost/compute/algorithm/accumulate.hpp>
20 #include <boost/compute/container/vector.hpp>
21
22 #include "perf.hpp"
23
24 namespace po = boost::program_options;
25 namespace compute = boost::compute;
26
rand_int()27 int rand_int()
28 {
29 return static_cast<int>((rand() / double(RAND_MAX)) * 25.0);
30 }
31
32 template<class T>
perf_accumulate(const compute::vector<T> & data,const size_t trials,compute::command_queue & queue)33 double perf_accumulate(const compute::vector<T>& data,
34 const size_t trials,
35 compute::command_queue& queue)
36 {
37 perf_timer t;
38 for(size_t trial = 0; trial < trials; trial++){
39 t.start();
40 compute::accumulate(data.begin(), data.end(), T(0), queue);
41 queue.finish();
42 t.stop();
43 }
44 return t.min_time();
45 }
46
47 template<class T>
tune_accumulate(const compute::vector<T> & data,const size_t trials,compute::command_queue & queue)48 void tune_accumulate(const compute::vector<T>& data,
49 const size_t trials,
50 compute::command_queue& queue)
51 {
52 boost::shared_ptr<compute::detail::parameter_cache>
53 params = compute::detail::parameter_cache::get_global_cache(queue.get_device());
54
55 const std::string cache_key =
56 std::string("__boost_reduce_on_gpu_") + compute::type_name<T>();
57
58 const compute::uint_ tpbs[] = { 4, 8, 16, 32, 64, 128, 256, 512, 1024 };
59 const compute::uint_ vpts[] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16 };
60
61 double min_time = (std::numeric_limits<double>::max)();
62 compute::uint_ best_tpb = 0;
63 compute::uint_ best_vpt = 0;
64
65 for(size_t i = 0; i < sizeof(tpbs) / sizeof(*tpbs); i++){
66 params->set(cache_key, "tpb", tpbs[i]);
67 for(size_t j = 0; j < sizeof(vpts) / sizeof(*vpts); j++){
68 params->set(cache_key, "vpt", vpts[j]);
69
70 try {
71 const double t = perf_accumulate(data, trials, queue);
72 if(t < min_time){
73 best_tpb = tpbs[i];
74 best_vpt = vpts[j];
75 min_time = t;
76 }
77 }
78 catch(compute::opencl_error&){
79 // invalid parameters for this device, skip
80 }
81 }
82 }
83
84 // store optimal parameters
85 params->set(cache_key, "tpb", best_tpb);
86 params->set(cache_key, "vpt", best_vpt);
87 }
88
main(int argc,char * argv[])89 int main(int argc, char *argv[])
90 {
91 // setup command line arguments
92 po::options_description options("options");
93 options.add_options()
94 ("help", "show usage instructions")
95 ("size", po::value<size_t>()->default_value(8192), "input size")
96 ("trials", po::value<size_t>()->default_value(3), "number of trials to run")
97 ("tune", "run tuning procedure")
98 ;
99 po::positional_options_description positional_options;
100 positional_options.add("size", 1);
101
102 // parse command line
103 po::variables_map vm;
104 po::store(
105 po::command_line_parser(argc, argv)
106 .options(options).positional(positional_options).run(),
107 vm
108 );
109 po::notify(vm);
110
111 const size_t size = vm["size"].as<size_t>();
112 const size_t trials = vm["trials"].as<size_t>();
113 std::cout << "size: " << size << std::endl;
114
115 // setup context and queue for the default device
116 compute::device device = compute::system::default_device();
117 compute::context context(device);
118 compute::command_queue queue(context, device);
119 std::cout << "device: " << device.name() << std::endl;
120
121 // create vector of random numbers on the host
122 std::vector<int> host_data(size);
123 std::generate(host_data.begin(), host_data.end(), rand_int);
124
125 // create vector on the device and copy the data
126 compute::vector<int> device_data(
127 host_data.begin(), host_data.end(), queue
128 );
129
130 // run tuning proceure (if requested)
131 if(vm.count("tune")){
132 tune_accumulate(device_data, trials, queue);
133 }
134
135 // run benchmark
136 double t = perf_accumulate(device_data, trials, queue);
137 std::cout << "time: " << t / 1e6 << " ms" << std::endl;
138
139 return 0;
140 }
141