"""
        scale = [1.0] * self.out_channel_size
        offset = [0.5] * self.out_channel_size
        mean, variance = scale, offset
        out = nn_ops.depthwise_conv2d_native(
            input_tensor,
            self.filters,
            strides=[1, 2, 2, 1],
            dilations=[1, 1, 1, 1],
            padding='SAME',
            data_format='NHWC',
        )

  %cst = "tf.Const"() {value = dense<2.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<3xf32>} : () -> tensor<3xf32>
  %0 = "tf.DepthwiseConv2dNative"(%arg0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<*xf32>, tensor<2x3x3x1xf32>) -> tensor<?x?x?x3xf32>

  %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1]
  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<*xf32>
  %1 = "tf.Relu6"(%0) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>


  %3 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {

   // OK
   %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>) -> tensor<256x30x30x12xf32>
   // Unsupported data format

        %cst_2 = "tf.Const"() {value = dense<3.0> : tensor<f32>} : () -> tensor<f32>
        %w = "tf.Mul"(%cst, %cst_2) : (tensor<2x3x3x1xf32>, tensor<f32>) -> tensor<2x3x3x1xf32>
        %0 = "tf.DepthwiseConv2dNative"(%arg0, %w) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<*xf32>, tensor<2x3x3x1xf32>) -> tensor<?x?x?x3xf32>

  %fq = "tf.FakeQuantWithMinMaxVars"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<f32>, tensor<f32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.DepthwiseConv2dNative"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x30x30x16xf32>

  %fq = "tf.FakeQuantWithMinMaxVars"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<f32>, tensor<f32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.DepthwiseConv2dNative"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x30x30x16xf32>

    if (!function_name.starts_with("composite_")) {
      return failure();
    }

    if (function_name.contains("depthwise_conv2d")) {
      // Uniform Quantized op requires weights of tf.DepthwiseConv2dNative to
      // be transformed from [H,W,C,M] to [H,W,1,CxM] where
      // H=height,W=width,C=channel,M=multiplier. Therefore, a reshape op is
      // inserted between the constant op and the function op so that the

  %2 = "tf.DepthwiseConv2dNative"(%1, %0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<1x520x520x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x520x520x1xf32>
  func.return %2 : tensor<1x520x520x1xf32>

  // CHECK: tf.Const
  // CHECK-NOT: tf.DepthwiseConv2dNative
}

// CHECK-LABEL: DontFoldNoConstantFold

    // Compared to tfl.conv_2d, tfl.depthwise_conv_2d has an additional
    // 'depth_multiplier' attribute. However, tf.DepthwiseConv2dNative does not
    // have a corresponding 'depth_multiplier' attribute; the multiplier is the
    // fourth dimension in the 4-D filter tensor. We query the multiplier from
    // tf.DepthwiseConv2dNative and set it as the attribute value accordingly.
    auto multiplier =