func.return %add : tensor<*xf32>
  }

  func.func @composite_conv2d_with_relu6_fn(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

  // Input: [N, H, W, C] for Conv2D or [N, D, H, W, C] for Conv3D.
  dnums.set_input_batch_dimension(0);
  dnums.set_input_feature_dimension(num_dims - 1);
  // Kernel: [K, K, I, O] for Conv2D or [K, K, K, I, O] for Conv3D.
  dnums.set_kernel_input_feature_dimension(num_dims - 2);
  dnums.set_kernel_output_feature_dimension(num_dims - 1);
  // Output: [N, H, W, C] for Conv2D or [N, D, H, W, C] for Conv3D.

        scale, offset = [1.0] * 2, [0.5] * 2
        mean, variance = scale, offset
        out = nn_ops.conv2d(
            q_input,
            q_filter,
            strides=[1, 1, 2, 1],
            dilations=[1, 1, 1, 1],
            padding='SAME',
            data_format='NHWC',
            name='sample/conv2d',
        )
        if has_bias:
          out = nn_ops.bias_add(out, bias, data_format='NHWC')

              (UpdateShapeWithAxis<-1> $qtype, $old_value))),
          [(CanUpdateShapeWithAxis<-1> $qtype, $old_value)]>;

// The axis is set to 0 because the transpose is from the legalization of
// tf.conv2d and the new channel axis is the first dimension.
def ReorderTransposeDequantQuantUsedByConv :
      Pat<(TF_TransposeOp:$old_value
              (TFL_DequantizeOp (TFL_QuantizeOp $input, $qtype)), $perm),

      if (!input_type || !input_type.hasStaticShape()) {
        return false;
      }
      total_count += input_type.getNumElements();
    }
    *count = total_count;
    return true;
  }

  // For conv2d/depthwise_conv/fully_connected ops.
  // This algorithm actually comes from TOCO tooling_util.cc
  static bool GetArithmeticCountForConvAndFullyconnectedOp(mlir::Operation* op,

  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

  %0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device...

  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

      %outputs_6, %control_7 = tf_executor.island wraps "tf.Const"() {device = "", value = dense<[-1, 16384]> : tensor<2xi32>} : () -> tensor<2xi32>
      %outputs_8, %control_9 = tf_executor.island wraps "tf.Conv2D"(%arg0, %outputs_0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true} : (tensor<16x224x224x3xf32>, tensor<*xf32>) -> tensor<16x112x112x?xf32>

  // Canonicalization includes const folding, which is utilized here to optimize
  // away ops that can't get constant folded after PrepareTF pass. For example,
  // tf.Conv2D is split into tf.Transpose and tfl.Conv2D.
  pass_manager->addNestedPass<mlir::func::FuncOp>(
      mlir::createCanonicalizerPass());
  pass_manager->addNestedPass<mlir::func::FuncOp>(mlir::createCSEPass());