InterfaceMethod<
      [{Returns the fixed output range.}],
      "UniformQuantizedType", "GetFixedOutputRange",
      (ins "bool":$sign, "int":$bit_width)
    >,
  ];
}

def AffineQuantizedOpInterface : OpInterface<
  "AffineQuantizedOpInterface"> {
  let description = [{
    Interface for affine quantized ops (conv2d, fully_connected, etc.)
  }];

  let methods = [
    InterfaceMethod<

class TFL_ConvOp<string mnemonic, string opSummary, int index,
                 list<Trait> additional_traits = []> :
    TFL_Op<mnemonic,[Pure,
                     AccumulatorUniformScale<2, 0, 1>,
                     AffineQuantizedOpInterface,
                     AffineOpCoefficient<index, 1>,
                     QuantizableResult,
                     TFL_SparseOp] # additional_traits> {
  let summary = opSummary # " operator";

    int bit_width = quant_specs_.GetQuantizationTypeWidth();

    Operation* quantize_op = quant_op.first;
    int quantize_operand_num = quant_op.second;

    auto affine_user = dyn_cast<AffineQuantizedOpInterface>(quantize_op);

    bool op_with_per_axis_support = false;

    if (!llvm::dyn_cast_or_null<CustomOp>(quantize_op)) {
      bool op_with_narrow_range =
          affine_user &&

  // FC and Conv* ops. Restriction for the weight can be relaxed if there are
  // needs for adjusting scale of variable weights.
  auto affine_op = dyn_cast<AffineQuantizedOpInterface>(op);
  auto bias_op = op->getOperand(bias_index).getDefiningOp<arith::ConstantOp>();
  if (!affine_op || !bias_op || input_indices.size() != 2) return false;

    }
    for (auto& use : value.getUses()) {
      Operation* user = use.getOwner();
      if (user->hasTrait<mlir::OpTrait::IsTerminator>()) continue;

      auto affine_user = llvm::dyn_cast<mlir::AffineQuantizedOpInterface>(user);
      if (affine_user &&
          affine_user.GetAffineOperandIndex() == use.getOperandNumber() &&
          affine_user.RequiredNarrowRangeAffineOperand())
        continue;