Function Profiling

tools/pplnn.cc:1082–1131

}

static bool Profiling(const vector<string>& input_data, Runtime* runtime) {
    if (g_flag_warmup_iterations > 0) {
        LOG(INFO) << "Warm up start for " << g_flag_warmup_iterations << " times.";
        for (uint32_t i = 0; i < g_flag_warmup_iterations; ++i) {
            runtime->Run();
        }
        LOG(INFO) << "Warm up end.";
    }

#ifdef PPLNN_ENABLE_KERNEL_PROFILING
    auto status = runtime->Configure(RUNTIME_CONF_SET_KERNEL_PROFILING_FLAG, true);
    if (status != RC_SUCCESS) {
        LOG(WARNING) << "enable profiling failed: " << GetRetCodeStr(status);
    }
#endif
    LOG(INFO) << "Profiling start";

    double run_dur = 0;
    uint32_t run_count = 0;
    while (run_dur < g_flag_min_profiling_seconds * 1000 || run_count < g_flag_min_profiling_iterations) {
        auto run_begin_ts = std::chrono::system_clock::now();
        if (g_flag_perf_with_io) {
            SetInputs(input_data, runtime);
        }
        runtime->Run();
        if (g_flag_perf_with_io) {
            GetOutputs(runtime);
        }
        auto run_end_ts = std::chrono::system_clock::now();
        auto diff = std::chrono::duration_cast<std::chrono::microseconds>(run_end_ts - run_begin_ts);
        run_dur += (double)diff.count() / 1000;
        run_count += 1;
    }

    LOG(INFO) << "Total duration: " << run_dur << " ms";

#ifdef PPLNN_ENABLE_KERNEL_PROFILING
    ProfilingStatistics stat;
    status = runtime->GetProfilingStatistics(&stat);
    if (status != RC_SUCCESS) {
        LOG(WARNING) << "Get profiling statistics failed: " << GetRetCodeStr(status);
    }
    PrintProfilingStatistics(stat, run_dur, run_count);
#else
    LOG(INFO) << "Average run costs: " << (run_dur / run_count) << " ms.";
#endif

    LOG(INFO) << "Profiling End";
    return true;
}

static uint32_t CalcModelNum() {
    uint32_t counter = 0;