%0 = "tf.Conv2D"(%arg0, %arg1) {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<1x2x2x3xf32>, tensor<2x2x3x2xf32>) -> tensor<*xf32>

    %0 = "tf.Conv2D"(%arg0, %arg1) {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 = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>
    return %0 : tensor<*xf32>
  }

    %0 = "tf.Conv2D"(%arg0, %cst_1) {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 = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>
    // Dequantize added before BiasAdd.

  %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>
  %6 = "tf.TPUReplicatedOutput"(%5) {device = ""} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xbf16>

  %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 = ""} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>

  %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>
  %3 = "tf.BiasAdd"(%2, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>

    %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>
    %6 = "tf.Mul"(%arg2, %arg4) : (tensor<f32>, tensor<512xf32>) -> tensor<512xf32>

    %0 = "tf.Conv2D"(%arg0, %arg1) {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<*xf32>, tensor<*xf32>) -> tensor<*xf32>

      %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>

    %6 = "tf.Conv2D"(%5, %arg1) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true} : (tensor<2x230x230x3xf32>, tensor<7x7x3x64xf32>) -> tensor<2x112x112x64xf32>
    // CHECK: %[[BACKPROP:.*]] = "tf.Conv2DBackpropFilter"
    // CHECK-SAME: strides = [1, 1, 1, 1]