struct KernelSpec {
  // Scale constraint
  ScaleConstraintType type;

  // Custom function to derive the scales. Only available when the scale
  // constraint is `CustomScale`.
  ScaleFn scale_fn;
};

class KernelSpecs {
 public:
  using Signature = llvm::SmallVector<quant::AnyQuantizedType, 4>;

  // Returns the kernel specification for the kernel signature.

  switch (spec->type) {
    case ScaleConstraintType::OutputInputFreeScale: {
      // no propagation.
      *changed |= false;
      break;
    }
    case ScaleConstraintType::CustomScale: {
      if (failed(spec->scale_fn(this, op, new_items, changed))) {
        return failure();
      }
      break;
    }
    case ScaleConstraintType::OutputInputSameScale: {
      auto params = GetQuantParamsForSameScaleConstraint(op);

  Value scale_op = rewriter.create<TF::ConstOp>(
      loc, scale_type,
      DenseFPElementsAttr::get(scale_type,
                               {static_cast<float>(qtype.getScale())}));

  if (original_input_tensor_type.getElementType().isBF16()) {
    // Add bf16 cast op after scale to match with the next op's data
    // type.
    scale_op = rewriter.create<TF::CastOp>(

      return quantizeF32ToInt8(expressed_value);
    }

    bool lossy;
    expressed_value.convert(scale_.getSemantics(), round_mode_, &lossy);
    // fixed_point = clamp(clamp_min, clamp_max, (
    //   roundHalfToEven(expressed / scale) + zero_point))
    APFloat scaled = (expressed_value / scale_);
    scaled.roundToIntegral(round_mode_);
    scaled.add(zero_point_, round_mode_);

      return failure();
    }

    TF::ConstOp scale_op;
    TF::ConstOp zp_op;

    // Reads quantization parameters from the quantized type, and converts
    // them to constants.
    rewriter.setInsertionPoint(qparams_op);
    Location loc = qparams_op->getLoc();
    if (auto qtype = cast_qtype.dyn_cast<quant::UniformQuantizedType>()) {
      scale_op = rewriter.create<TF::ConstOp>(

  } else {  // float
    signature->push_back(AnyQuantizedType());
  }
}

LogicalResult DeviceTarget::RegisterKernel(
    llvm::StringRef kernel, const KernelSpecs::Signature& signature,
    const ScaleFn& fn, const ScaleDecomposeFn& dfn) {
  return specs_[kernel].Add(signature, {ScaleConstraintType::CustomScale, fn});
}

namespace ph = std::placeholders;

LogicalResult DeviceTarget::RegisterKernel(

	if ratio == 1 {
		return
	}
	ratios := make([]float64, len(p.SampleType))
	for i := range p.SampleType {
		ratios[i] = ratio
	}
	p.ScaleN(ratios)
}

// ScaleN multiplies each sample values in a sample by a different amount.
func (p *Profile) ScaleN(ratios []float64) error {
	if len(p.SampleType) != len(ratios) {
		return fmt.Errorf("mismatched scale ratios, got %d, want %d", len(ratios), len(p.SampleType))
	}

		return
	}
	ratios := make([]float64, len(p.SampleType))
	for i := range p.SampleType {
		ratios[i] = ratio
	}
	p.ScaleN(ratios)
}

// ScaleN multiplies each sample values in a sample by a different amount
// and keeps only samples that have at least one non-zero value.
func (p *Profile) ScaleN(ratios []float64) error {
	if len(p.SampleType) != len(ratios) {

			if sampleType[i] == nil {
				ratios[i] = 1
				continue
			}
			ratios[i], _ = Scale(1, st.Unit, sampleType[i].Unit)
			p.SampleType[i].Unit = sampleType[i].Unit
		}
		if err := p.ScaleN(ratios); err != nil {
			return fmt.Errorf("scale: %v", err)
		}
	}
	return nil
}

// CommonValueType returns the finest type from a set of compatible
// types.

	normScale := make([]float64, len(baseVals))
	for i := range baseVals {
		if srcVals[i] == 0 {
			normScale[i] = 0.0
		} else {
			normScale[i] = float64(baseVals[i]) / float64(srcVals[i])
		}
	}
	p.ScaleN(normScale)
	return nil
}

func isZeroSample(s *Sample) bool {
	for _, v := range s.Value {
		if v != 0 {
			return false
		}
	}
	return true
}

type profileMerger struct {