func.return %1: tensor<*xf32>
  }
  func.func private @composite_matmul_fn_1(%arg0: tensor<2x12xf32>, %arg1: tensor<12x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    return %0 : tensor<*xf32>
  }
}

// PerTensor-LABEL: func @matmul

  }
  func.func private @composite_conv2d_with_bias_and_relu6_fn_1(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>, %arg2: tensor<2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {
      attr_map = "0:strides,1:padding,2:explicit_paddings,3:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1]

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
    return %0 : tensor<2x2xf32>
  }
  func.func private @composite_matmul_fn_2_0(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> tensor<2x2xf32> attributes {tf_quant.composite_function} {

  func.func private @composite_conv2d_with_bias_and_relu6_fn_2(%arg0: tensor<1x2x2x3xf32>, %arg1: tensor<2x2x3x2xf32>, %arg2: tensor<2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

            auto range = arg_ranges.at(arg.name());
            values.push_back(range.second - range.first);
          }
          auto attr_value =
              mlir::DenseI32ArrayAttr::get(tf_op->getContext(), values);
          tf_op->setAttr(attr_name, attr_value);
        };
    if (tf_op->hasTrait<mlir::OpTrait::AttrSizedOperandSegments>() ||
        tf_op->hasTrait<mlir::OpTrait::AttrSizedResultSegments>()) {

// CHECK-SAME: attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations"
// CHECK-NEXT: return %[[CONV2D_0]]

// CHECK-LABEL: private @composite_conv2d_fn_1
// CHECK-NEXT: %[[CONV2D_0:.*]] = "tf.Conv2D"(%arg0, %arg1)
// CHECK-SAME: attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations"

    %quant_input, %quant_filter, %input_scale, %input_zp,
    %filter_scale, %filter_zp, %accum_scale, %accum_zp
  ) {
    Tin = "tfdtype$DT_QINT8", Tout = "tfdtype$DT_QINT32",
    attr_map = "", batch_group_count = 1 : i64,
    dimension_numbers = "\10\03\1A\02\01\02 \02(\032\02\00\01@\03J\02\01\02",
    explicit_padding = [], feature_group_count = 1 : i64, lhs_dilation = [1, 1],

  (*op)->setAttr("data_format", StringAttr::get(context, target_data_format));

  for (auto pair : shuffle_attrs) {
    StringRef attr_name = pair.first;
    ArrayAttr attr_value = pair.second;
    (*op)->setAttr(attr_name,
                   ShuffleArrayAttr(attr_value, reverse_permutation));
  }

  auto fold = cast<FoldOperandsTransposeInterface>(op->getOperation());
  for (unsigned idx : fold.GetLayoutDependentResults()) {

    %2 = "tf.Cast"(%identity) {Truncate = false} : (tensor<*xi8>) -> tensor<*xi32>
    %3 = "tf.Sub"(%2, %weight_zp) : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>

    %5 = "tf.MatMul"(%1, %3) {
      attr_map = "transpose_a:0,transpose_b:1"
    } : (tensor<*xi32>, tensor<*xi32>) -> tensor<*xi32>
    func.return %5 : tensor<*xi32>
  }

  // Conv2D with int32 accumulation
  func.func private @internal_conv2d_fn(