CuptiProfilerResult evalMetricValues()

in libkineto/src/CuptiNvPerfMetric.cpp [370:467]
74 lines of code
9 McCabe index (conditional complexity)

CuptiProfilerResult evalMetricValues(
    const std::string& chipName,
    const std::vector<uint8_t>& counterDataImage,
    const std::vector<std::string>& metricNames,
    bool verbose) {

  if (!counterDataImage.size()) {
    LOG(ERROR) << "Counter Data Image is empty!";
    return {};
  }

  NVPW_CUDA_MetricsContext_Create_Params metricsContextCreateParams = {
      NVPW_CUDA_MetricsContext_Create_Params_STRUCT_SIZE, nullptr};
  metricsContextCreateParams.pChipName = chipName.c_str();
  if (!NVPW_CALL(
        NVPW_CUDA_MetricsContext_Create(&metricsContextCreateParams))) {
    return {};
  }

  NVPW_MetricsContext_Destroy_Params metricsContextDestroyParams = {
      NVPW_MetricsContext_Destroy_Params_STRUCT_SIZE, nullptr};
  metricsContextDestroyParams.pMetricsContext =
      metricsContextCreateParams.pMetricsContext;
  SCOPE_EXIT([&]() {
    NVPW_MetricsContext_Destroy(
        (NVPW_MetricsContext_Destroy_Params*)&metricsContextDestroyParams);
  });

  NVPW_CounterData_GetNumRanges_Params getNumRangesParams = {
      NVPW_CounterData_GetNumRanges_Params_STRUCT_SIZE, nullptr};
  getNumRangesParams.pCounterDataImage = counterDataImage.data();
  if (!NVPW_CALL(
      NVPW_CounterData_GetNumRanges(&getNumRangesParams))) {
    return {};
  }

  // TBD in the future support special chars in metric name
  // for now these are default
  const bool isolated = true;

  // API takes a 2D array of chars
  std::vector<const char*> metricNamePtrs;

  for (const auto& metric : metricNames) {
    metricNamePtrs.push_back(metric.c_str());
  }

  CuptiProfilerResult result{
    .metricNames = metricNames};

  for (size_t rangeIndex = 0; rangeIndex < getNumRangesParams.numRanges;
       ++rangeIndex) {

    CuptiRangeMeasurement rangeData {
    .rangeName = getRangeDescription(counterDataImage, rangeIndex)};
    rangeData.values.resize(metricNames.size());

    // First set Counter data image with current range
    NVPW_MetricsContext_SetCounterData_Params setCounterDataParams = {
        NVPW_MetricsContext_SetCounterData_Params_STRUCT_SIZE, nullptr};

    setCounterDataParams.pMetricsContext =
        metricsContextCreateParams.pMetricsContext;
    setCounterDataParams.pCounterDataImage = counterDataImage.data();
    setCounterDataParams.isolated = isolated;
    setCounterDataParams.rangeIndex = rangeIndex;

    NVPW_CALL(NVPW_MetricsContext_SetCounterData(&setCounterDataParams));


    // Now we can evaluate GPU metrics
    NVPW_MetricsContext_EvaluateToGpuValues_Params evalToGpuParams = {
        NVPW_MetricsContext_EvaluateToGpuValues_Params_STRUCT_SIZE, nullptr};
    evalToGpuParams.pMetricsContext =
        metricsContextCreateParams.pMetricsContext;
    evalToGpuParams.numMetrics = metricNamePtrs.size();
    evalToGpuParams.ppMetricNames = metricNamePtrs.data();
    evalToGpuParams.pMetricValues = rangeData.values.data();

    if (!NVPW_CALL(NVPW_MetricsContext_EvaluateToGpuValues(&evalToGpuParams))) {
      LOG(WARNING) << "Failed to evaluate metris for range : "
                   << rangeData.rangeName;
      continue;
    }

    if (verbose) {
      for (size_t i = 0; i < metricNames.size(); i++) {
        LOG(INFO) << "rangeName: " << rangeData.rangeName
                  << "\tmetricName: " << metricNames[i]
                  << "\tgpuValue: " << rangeData.values[i];
      }
    }

    result.rangeVals.emplace_back(std::move(rangeData));
  }

  return result;
}