return %3 : tensor<?x?x?x2xf32>
  }

// CHECK-LABEL: func @conv_with_dynamic_shape
// The Conv2D should not be quantized since it has dynamic channel.
// CHECK: "tf.Conv2D"
// CHECK-SAME: (tensor<?x?x?x?xf32>, tensor<2x3x3x2xf32>) -> tensor<?x?x?x2xf32>

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])

  // CHECK: %[[CONV2D:[0-9]*]] = "tf.Conv2D"(%[[ARG_TRANSPOSE]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: dilations = [1, 4, 2, 3]
  // CHECK-SAME: explicit_paddings = [1, 2, 7, 8, 3, 4, 5, 6]
  // CHECK-SAME: padding = "EXPLICIT"

  // cuDNN prefers NCHW data format for spatial convolutions.
  // CHECK: "tf.Conv2D"(%[[INPUT_TRANSPOSE:[0-9]*]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2D"(%input, %filter)
       {
         data_format = "NHWC",
         padding = "VALID",
         strides = [1, 1, 1, 1]

  %filter = arith.constant dense<2.0> : tensor<3x3x3x16xf32>
  %bias = arith.constant dense<3.0> : tensor<16xf32>
  %value = arith.constant dense<4.0> : tensor<16xf32>
  %0 = "tf.Conv2D"(%arg, %filter) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])

  // CHECK: %[[CONV2D:[0-9]*]] = "tf.Conv2D"(%[[ARG_TRANSPOSE]], %arg1)
  // CHECK-SAME: data_format = "NHWC"
  // CHECK-SAME: dilations = [1, 3, 4, 2]
  // CHECK-SAME: explicit_paddings = [1, 2, 5, 6, 7, 8, 3, 4]
  // CHECK-SAME: padding = "EXPLICIT"

        mul_op.getLoc(), conv_op.getType(), conv_op.getLhs(), new_filter,
        conv_op.getWindowStridesAttr(), conv_op.getPaddingAttr(),
        conv_op.getLhsDilationAttr(), conv_op.getRhsDilationAttr(),
        conv_op.getWindowReversalAttr(), conv_op.getDimensionNumbers(),
        conv_op.getFeatureGroupCount(), conv_op.getBatchGroupCount(),
        conv_op.getPrecisionConfigAttr());

  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>
  %1 = "tf.Conv2D"(%0, %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<1x3x4x3xbf16>, tensor<2x3x3x2xbf16>) -> tensor<1x3x2x2xbf16>

// -----

func.func @conv_with_non_constant_filter(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

// CHECK: return %0 : tensor<*xf32>
}

// -----

module {
  func.func @conv2d(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

      %outputs_2, %control_3 = tf_executor.island(%control_1) wraps "tf.Const"() {value = dense<0.000000e+00> : tensor<5x5x32x16xf32>} : () -> tensor<5x5x32x16xf32>
      %outputs_4, %control_5 = tf_executor.island wraps "tf.Conv2D"(%outputs, %outputs_2) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<5x5x32x16xf32>) -> tensor<*xf32>