// CKECK-SAME: <{calibration_method = 5 : i32, id = "skipping_id", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>) -> (tensor<1x3x4x3xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)
// CHECK: %[[CUSTOM_AGGREGATOR_1:.*]], %[[MIN_1:.*]], %[[MAX_1:.*]], %[[HISTOGRAM_1:.*]] = "tf.CustomAggregator"

	// GaugeMT - represents a gauge metric.
	GaugeMT
	// HistogramMT - represents a histogram metric.
	HistogramMT
	// rangeL - represents a range label.
	rangeL = "range"
)

func (mt MetricType) String() string {
	switch mt {
	case CounterMT:
		return "counter"
	case GaugeMT:
		return "gauge"
	case HistogramMT:
		return "histogram"
	default:
		return "*unknown*"
	}
}

  inputs.emplace_back("min", 0, DT_FLOAT);
  inputs.emplace_back("max", 0, DT_FLOAT);
  inputs.emplace_back("histogram", 0, DT_INT64);
  inputs.emplace_back("min", 0, DT_FLOAT);
  inputs.emplace_back("max", 0, DT_FLOAT);
  inputs.emplace_back("histogram", 0, DT_INT64);

  const std::string dir = testing::TmpDir();
  const std::string output_file_path = io::JoinPath(dir, "statistics.pbtxt");

    CALIBRATION_METHOD_AVERAGE_MIN_MAX = 2;
    // Use the min/max percentile value of histogram.
    CALIBRATION_METHOD_HISTOGRAM_PERCENTILE = 3;
    // Use the histogram mid values that minimize MSE error.
    // This is very slow algorithm because it computes all errors for all
    // histogram mid value pairs. Therefore the value of num_bins is recommended
    // to be 256 or less.

Regions are grouped by the region type and the point at which the region started.
The rightmost column of the table contains a latency histogram for each region group.
Note that this histogram only counts regions that began and ended within the traced
period.
However, the "Count" column includes all regions, including those that only started
or ended during the traced period.

	// Insert random uniforms and check histogram mass and
	// cumulative sum approximations.
	rnd := rand.New(rand.NewSource(42))
	mass := 0.0
	var mud mud
	for i := 0; i < 100; i++ {
		area, l, r := rnd.Float64(), rnd.Float64(), rnd.Float64()
		if rnd.Intn(10) == 0 {
			r = l
		}
		t.Log(l, r, area)
		mud.add(l, r, area)
		mass += area

		// Check total histogram weight.
		hmass := 0.0

	// trackSum >= trackMass, trackBucket must be recomputed.
	trackSum float64

	// hist is a hierarchical histogram of distribution mass.
	hist [mudDegree]float64
}

const (
	// mudDegree is the number of buckets in the MUD summary
	// histogram.
	mudDegree = 1024
)

type edge struct {
	// At x, the function increases by y.
	x, delta float64

    .SetIsStateful()
    .Doc(R"doc(
Aggregates and saves the calibration statistics data.

This op collects outputs of multiples CustomAggregator ops, which includes
`min`, `max` and `histogram`. Then it aggregates them according to the
calibration method and save the result to the given file path as a binary
proto file.)doc");

class CalibrationStatisticsSaverOp : public OpKernel {
 public:

	{"GREATER_THAN_512_MB", humanize.MiByte * 512, math.MaxInt64},
}

// ObjectsHistogramIntervals is the list of all intervals
// of object sizes to be included in objects histogram.
// Note: this histogram expands 1024B-1MB to incl. 1024B-64KB, 64KB-256KB, 256KB-512KB and 512KB-1MiB
var ObjectsHistogramIntervals = [dataUsageBucketLen]objectHistogramInterval{
	{"LESS_THAN_1024_B", 0, humanize.KiByte - 1},

		VariableLabels:       make(map[string]string, len(m.VariableLabels)),
		Histogram:            make(map[string]uint64, len(m.Histogram)),
	}
	for k, v := range m.StaticLabels {
		metric.StaticLabels[k] = v
	}
	for k, v := range m.VariableLabels {
		metric.VariableLabels[k] = v
	}
	for k, v := range m.Histogram {
		metric.Histogram[k] = v
	}
	return metric
}