1 //---------------------------------------------------------------------------//
2 // Copyright (c) 2015 Jakub Szuppe <j.szuppe@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 <iostream>
12 #include <algorithm>
13 #include <vector>
14
15 #include <bolt/cl/copy.h>
16 #include <bolt/cl/device_vector.h>
17 #include <bolt/cl/reduce_by_key.h>
18
19 #include "perf.hpp"
20
rand_int()21 int rand_int()
22 {
23 return static_cast<int>((rand() / double(RAND_MAX)) * 25.0);
24 }
25
26 struct unique_key {
27 int current;
28 int avgValuesNoPerKey;
29
unique_keyunique_key30 unique_key()
31 {
32 current = 0;
33 avgValuesNoPerKey = 512;
34 }
35
operator ()unique_key36 int operator()()
37 {
38 double p = double(1.0) / static_cast<double>(avgValuesNoPerKey);
39 if((rand() / double(RAND_MAX)) <= p)
40 return ++current;
41 return current;
42 }
43 } UniqueKey;
44
main(int argc,char * argv[])45 int main(int argc, char *argv[])
46 {
47 perf_parse_args(argc, argv);
48
49 std::cout << "size: " << PERF_N << std::endl;
50
51 bolt::cl::control ctrl = bolt::cl::control::getDefault();
52 ::cl::Device device = ctrl.getDevice();
53 std::cout << "device: " << device.getInfo<CL_DEVICE_NAME>() << std::endl;
54
55 // create vector of keys and random values
56 std::vector<int> host_keys(PERF_N);
57 std::vector<int> host_values(PERF_N);
58 std::generate(host_keys.begin(), host_keys.end(), UniqueKey);
59 std::generate(host_values.begin(), host_values.end(), rand_int);
60
61 // create device vectors for data
62 bolt::cl::device_vector<int> device_keys(PERF_N);
63 bolt::cl::device_vector<int> device_values(PERF_N);
64
65 // transfer data to the device
66 bolt::cl::copy(host_keys.begin(), host_keys.end(), device_keys.begin());
67 bolt::cl::copy(host_values.begin(), host_values.end(), device_values.begin());
68
69 // create device vectors for the results
70 bolt::cl::device_vector<int> device_keys_results(PERF_N);
71 bolt::cl::device_vector<int> device_values_results(PERF_N);
72
73 typedef bolt::cl::device_vector<int>::iterator iterType;
74 bolt::cl::pair<iterType, iterType> result = {
75 device_keys_results.begin(),
76 device_values_results.begin()
77 };
78
79 perf_timer t;
80 for(size_t trial = 0; trial < PERF_TRIALS; trial++){
81 t.start();
82 result = bolt::cl::reduce_by_key(device_keys.begin(),
83 device_keys.end(),
84 device_values.begin(),
85 device_keys_results.begin(),
86 device_values_results.begin());
87 t.stop();
88 }
89 std::cout << "time: " << t.min_time() / 1e6 << " ms" << std::endl;
90
91 size_t result_size = bolt::cl::distance(device_keys_results.begin(), result.first);
92 if(result_size != static_cast<size_t>(host_keys[PERF_N-1] + 1)){
93 std::cout << "ERROR: "
94 << "wrong number of keys"
95 << std::endl;
96 return -1;
97 }
98
99 return 0;
100 }
101