namespace tensorflow {
namespace calibrator {

using ::stablehlo::quantization::CalibrationOptions;

// AverageMinMax calibration calculates the average of min and max values.
// average of min = sum of min values / number of samples
// average of max = sum of max values / number of samples
class CalibrationStatisticsCollectorAverageMinMax
    : public CalibrationStatisticsCollectorBase {
 public:

var (
	replicationAverageActiveWorkersMD = NewGaugeMD(replicationAverageActiveWorkers,
		"Average number of active replication workers")
	replicationAverageQueuedBytesMD = NewGaugeMD(replicationAverageQueuedBytes,
		"Average number of bytes queued for replication since server start")
	replicationAverageQueuedCountMD = NewGaugeMD(replicationAverageQueuedCount,
		"Average number of objects queued for replication since server start")

StringAttr GetAvgPoolOpPadAttr(Builder& builder, mhlo::CompositeOp op);

// Get dense attr for a matrix that corrects the over counting of divisors when
// casting an average pool with ceil mode on in terms of average pool with it
// off.
DenseFPElementsAttr GetCorrectionMatrix(Builder& builder, mhlo::CompositeOp op);

}  // namespace odml
}  // namespace mlir

  // for any `x`, `ceil(x) - floor(x) <= 1`.

  // Consider that we pad the input of a average pool with floor rounding to the
  // appropriate size and switch the rounding mode to ceil. When computing the
  // average of a given window, the elements which exist in the newly padded
  // zones will be counted as present elements. Therefore in some windows we

	}
	if _, ok := m1.(*image.Gray); !ok {
		t.Errorf("got %T, want *image.Gray", m1)
	}
	// Compare the average delta to the tolerance level.
	want := int64(2 << 8)
	if got := averageDelta(m0, m1); got > want {
		t.Errorf("average delta too high; got %d, want <= %d", got, want)
	}
}

// averageDelta returns the average delta in RGB space. The two images must
// have the same bounds.
func averageDelta(m0, m1 image.Image) int64 {

	m2Distance := n.NUMADistances.CalculateAverageFor(m2)
	// If average distance is the same, take bitmask with more lower-number bits set.
	if m1Distance == m2Distance {
		if m1.IsLessThan(m2) {
			return m1
		}
		return m2
	}

	// Otherwise, return the bitmask with the shortest average distance between NUMA nodes.
	if m1Distance < m2Distance {
		return m1
	}

	return m2
}

	case sizeLessThan1GiB:
		return "LESS_THAN_1_GiB"
	case sizeGreaterThan1GiB:
		return "GREATER_THAN_1_GiB"
	default:
		return "unknown"
	}
}

// AccElem holds information for calculating an average value
type AccElem struct {
	Total int64
	Size  int64
	N     int64
}

// Add a duration to a single element.
func (a *AccElem) add(dur time.Duration) {
	if dur < 0 {
		dur = 0
	}

		allDriveLabels...)
	driveWaitingIOMD = NewGaugeMD(driveWaitingIO,
		"Total waiting I/O operations on a drive", allDriveLabels...)
	driveAPILatencyMD = NewGaugeMD(driveAPILatencyMicros,
		"Average last minute latency in µs for drive API storage operations",
		append(allDriveLabels, apiL)...)
	driveHealthMD = NewGaugeMD(driveHealth,
		"Drive health (0 = offline, 1 = healthy, 2 = healing)", allDriveLabels...)

	show("Float32", r.Float32(), r.Float32(), r.Float32())
	show("Float64", r.Float64(), r.Float64(), r.Float64())

	// ExpFloat64 values have an average of 1 but decay exponentially.
	show("ExpFloat64", r.ExpFloat64(), r.ExpFloat64(), r.ExpFloat64())

	// NormFloat64 values have an average of 0 and a standard deviation of 1.
	show("NormFloat64", r.NormFloat64(), r.NormFloat64(), r.NormFloat64())

	show("Float32", r.Float32(), r.Float32(), r.Float32())
	show("Float64", r.Float64(), r.Float64(), r.Float64())

	// ExpFloat64 values have an average of 1 but decay exponentially.
	show("ExpFloat64", r.ExpFloat64(), r.ExpFloat64(), r.ExpFloat64())

	// NormFloat64 values have an average of 0 and a standard deviation of 1.
	show("NormFloat64", r.NormFloat64(), r.NormFloat64(), r.NormFloat64())