runWavePerfRepeat()

WavePerfRepeatResult fl::wave_perf::runWavePerfRepeat ( u32 W, u32 H, u32 iterations, u32 repeats, LaplacianStencil stencil )

inline

Definition at line 146 of file wave_perf_bench.h.

                                           {
    WavePerfRepeatResult r;
    if (!isValidGrid(W, H) || !isValidIterations(iterations)
        || !isValidRepeats(repeats)) {
        return r;
    }
    // Single sim reused across repeats — the warm-up cost is paid once
    // and each repeat measures steady-state cost.
    auto sim = makeBenchSim(W, H, stencil);
    double sum = 0.0;
    double sum_sq = 0.0;
    for (u32 j = 0; j < repeats; ++j) {
        const u32 t0 = fl::micros();
        for (u32 i = 0; i < iterations; ++i) {
            sim->update();
        }
        const u32 t1 = fl::micros();
        const double us_per_update =
            static_cast<double>(t1 - t0) / static_cast<double>(iterations);
        sum += us_per_update;
        sum_sq += us_per_update * us_per_update;
    }
    const double mean = sum / static_cast<double>(repeats);
    const double variance = (sum_sq / static_cast<double>(repeats))
                          - (mean * mean);
    double std_dev = 0.0;
    if (variance > 0.0) {
        // Newton-Raphson sqrt — avoids pulling in <cmath> on platforms
        // where it's costly. 6 iterations gives ~14 significant digits.
        double s = variance;
        for (int k = 0; k < 6; ++k) {
            s = 0.5 * (s + variance / s);
        }
        std_dev = s;
    }
    r.success = true;
    r.repeats = repeats;
    r.mean_us_per_update = mean;
    r.std_dev_us_per_update = std_dev;
    r.std_dev_pct = (mean > 0.0) ? (100.0 * std_dev / mean) : 0.0;
    return r;
}

References fl::FL_NO_EXCEPT, H, isValidGrid(), isValidIterations(), isValidRepeats(), makeBenchSim(), fl::wave_perf::WavePerfRepeatResult::mean_us_per_update, fl::micros(), fl::wave_perf::WavePerfRepeatResult::repeats, fl::wave_perf::WavePerfRepeatResult::std_dev_pct, fl::wave_perf::WavePerfRepeatResult::std_dev_us_per_update, fl::wave_perf::WavePerfRepeatResult::success, and W.

Here is the call graph for this function: