Function printPerformanceReport

samples/common/sampleReporting.cpp:286–311

}

void printPerformanceReport(std::vector<InferenceTrace> const& trace, ReportingOptions const& reportingOpts,
    InferenceOptions const& infOpts, std::ostream& osInfo, std::ostream& osWarning, std::ostream& osVerbose)
{
    int32_t batchSize = infOpts.batch;
    float const warmupMs = infOpts.warmup;
    auto const isNotWarmup = [&warmupMs](const InferenceTrace& a) { return a.computeStart >= warmupMs; };
    auto const noWarmup = std::find_if(trace.begin(), trace.end(), isNotWarmup);
    int32_t const warmups = noWarmup - trace.begin();
    float const benchTime = trace.back().d2hEnd - noWarmup->h2dStart;
    // when implicit batch used, batchSize = options.inference.batch, which is parsed through --batch
    // when explicit batch used, batchSize = options.inference.batch = 0
    // treat inference with explicit batch as a single query and report the throughput
    batchSize = batchSize ? batchSize : 1;
    printProlog(warmups * batchSize, (trace.size() - warmups) * batchSize, warmupMs, benchTime, osInfo);

    std::vector<InferenceTime> timings(trace.size() - warmups);
    std::transform(noWarmup, trace.end(), timings.begin(), traceToTiming);
    printTiming(timings, reportingOpts.avgs, osInfo);
    printEpilog(
        timings, benchTime, reportingOpts.percentiles, batchSize, infOpts.infStreams, osInfo, osWarning, osVerbose);

    if (!reportingOpts.exportTimes.empty())
    {
        exportJSONTrace(trace, reportingOpts.exportTimes, warmups);
    }
}

//! Printed format:
//! [ value, ...]