| 48 | } |
| 49 | |
| 50 | double |
| 51 | time_loop(migraphx::gpu::context& gctx, int bundle, int nruns, const std::function<void()>& f) |
| 52 | { |
| 53 | // check for manual overrides |
| 54 | bundle = value_of(MIGRAPHX_BENCHMARKING_BUNDLE{}, bundle); |
| 55 | nruns = value_of(MIGRAPHX_BENCHMARKING_NRUNS{}, nruns); |
| 56 | |
| 57 | std::vector<std::pair<hip_event_ptr, hip_event_ptr>> events(nruns); |
| 58 | std::generate(events.begin(), events.end(), [] { |
| 59 | return std::make_pair(context::create_event_for_timing(), |
| 60 | context::create_event_for_timing()); |
| 61 | }); |
| 62 | std::vector<double> times; |
| 63 | // Warmup |
| 64 | f(); |
| 65 | for(auto i : range(nruns)) |
| 66 | { |
| 67 | gctx.get_stream().record(events[i].first.get()); |
| 68 | for(auto j : range(bundle)) |
| 69 | { |
| 70 | (void)j; |
| 71 | f(); |
| 72 | } |
| 73 | gctx.get_stream().record(events[i].second.get()); |
| 74 | } |
| 75 | gctx.finish(); |
| 76 | std::transform(events.begin(), events.end(), std::back_inserter(times), [&](const auto& p) { |
| 77 | return context::get_elapsed_ms(p.first.get(), p.second.get()) / bundle; |
| 78 | }); |
| 79 | std::sort(times.begin(), times.end()); |
| 80 | |
| 81 | // compute common average by removing top and bottom 25% of values |
| 82 | std::size_t quarters = times.size() / 4; |
| 83 | double total = std::accumulate(times.begin() + quarters, times.end() - quarters, 0.0); |
| 84 | return total / std::distance(times.begin() + quarters, times.end() - quarters); |
| 85 | } |
| 86 | |
| 87 | double |
| 88 | time_op(const context& ictx, operation op, const std::vector<shape>& inputs, int bundle, int nruns) |