| 19 | #include "perf.hpp" |
| 20 | |
| 21 | int main(int argc, char *argv[]) |
| 22 | { |
| 23 | perf_parse_args(argc, argv); |
| 24 | |
| 25 | std::cout << "size: " << PERF_N << std::endl; |
| 26 | |
| 27 | ::cl::Device device = bolt::cl::control::getDefault().getDevice(); |
| 28 | std::cout << "device: " << device.getInfo<CL_DEVICE_NAME>() << std::endl; |
| 29 | |
| 30 | // create host vector |
| 31 | std::vector<int> h_vec = generate_random_vector<int>(PERF_N); |
| 32 | // create device vector |
| 33 | bolt::cl::device_vector<int> d_vec(PERF_N); |
| 34 | |
| 35 | perf_timer t; |
| 36 | for(size_t trial = 0; trial < PERF_TRIALS; trial++){ |
| 37 | // transfer data to the device |
| 38 | bolt::cl::copy(h_vec.begin(), h_vec.end(), d_vec.begin()); |
| 39 | |
| 40 | t.start(); |
| 41 | bolt::cl::sort(d_vec.begin(), d_vec.end()); |
| 42 | t.stop(); |
| 43 | } |
| 44 | std::cout << "time: " << t.min_time() / 1e6 << " ms" << std::endl; |
| 45 | |
| 46 | // transfer data back to host |
| 47 | bolt::cl::copy(d_vec.begin(), d_vec.end(), h_vec.begin()); |
| 48 | |
| 49 | return 0; |
| 50 | } |