| 31 | |
| 32 | template<class T> |
| 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> |
| 48 | void tune_accumulate(const compute::vector<T>& data, |