%cst = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x512x512xf32>} : () -> tensor<2x3x512x512xf32>
  %0 = "tf.Conv2D"(%arg0, %cst) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true
  } : (tensor<1x3x4x512xf32>, tensor<2x3x512x512xf32>) -> tensor<*xf32>
  func.return %0 : tensor<*xf32>

  %conv = "tf.Conv2D"(%dq_input, %dq_weight) {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

                                     Value input, Value filter,
                                     int8_t input_zp_value, ArrayAttr strides,
                                     ArrayAttr dilations,
                                     StringAttr conv_padding,
                                     ArrayAttr explicit_paddings,
                                     Value &padding, int num_dims = 4);

  %6 = "tf.Cast"(%5) {Truncate = false} : (tensor<1x100x100x1xi8>) -> tensor<1x100x100x1xf32>

  %7 = "tf.Sub"(%6, %cst_4) : (tensor<1x100x100x1xf32>, tensor<f32>) -> tensor<1x100x100x1xf32>
  %8 = "tf.Conv2D"(%7, %cst_3) {dilations = [1, 1, 1, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x100x100x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x98x98x1xf32>
  %9 = "tf.AddV2"(%8, %cst_6) : (tensor<1x98x98x1xf32>, tensor<1xf32>) -> tensor<1x98x98x1xf32>

  %2 = "tf.Conv2D"(%1, %0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>

  %4 = "tf.Transpose"(%3, %cst_1) {_tpu_replicate = "cluster", device = ""} : (tensor<1x3x3x4xbf16>, tensor<4xi32>) -> tensor<1x3x4x3xbf16>
  %5 = "tf.Conv2D"(%4, %cst) {_tpu_replicate = "cluster", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xbf16>, tensor<2x3x3x2xbf16>) -> tensor<1x3x2x2xbf16>

    %3 = "tf.Cast"(%2) {Truncate = false} : (tensor<960x960x3x512xi8>) -> tensor<960x960x3x512xi32>
    %4 = "tf.Sub"(%3, %arg5) : (tensor<960x960x3x512xi32>, tensor<512xi32>) -> tensor<960x960x3x512xi32>
    %5 = "tf.Conv2D"(%1, %4) {dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x2240x2240x3xi32>, tensor<960x960x3x512xi32>) -> tensor<1x2240x1120x512xi32>

      %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>