// will be preserved since it has uses.
  %2 = "tf.MyCustomOp"(%1, %0) {fused_activation_function = "RELU", int_attr = 2 : i32}  : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  // will be preserved since it has side-effect.
  "tf.MyCustomOp"(%1, %0) {fused_activation_function = "RELU", int_attr = 2 : i32}  : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  func.return %2 : tensor<4xf32>

  if act == 'RELU':
    return tf.raw_ops.Relu(features=res)
  elif act == 'RELU6':
    return tf.raw_ops.Relu6(features=res)
  elif act == 'TANH':
    return tf.raw_ops.Tanh(x=res)
  else:
    return res


@tf.RegisterGradient('NewConv2D')
def _conv_add_relu_grad(op: ops.Operation, grad):
  act = op.get_attr('act')
  y = op.outputs[0]
  if act == 'RELU':
    grad = gen_nn_ops.relu_grad(grad, y)

func.func @main(%arg0: tensor<?xi32>) -> (tensor<?xi32>, tensor<?xi32>) {
    %0 = "tf.Relu"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
    %1 = "tf.Relu"(%0) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
    %2 = "tf.Add"(%0, %1) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf.Relu"(%2) : (tensor<?xi32>) -> tensor<?xi32>
    %4 = "tf.Relu"(%1) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>

// CHECK: %[[cast:.*]] = "tf.Cast"(%[[sub]]) <{Truncate = false}> : (tensor<1x3x2x2xi32>) -> tensor<1x3x2x2xf32>
// CHECK: %[[dequant1:.*]] = "tf.Mul"(%[[cast]]
// CHECK: %[[relu:.*]] = "tf.Relu"(%[[dequant1]]
// CHECK: %[[clamped:.*]] = "tf.Minimum"(%[[relu]]

// CHECK: %[[rescale1:.*]] = "tf.Mul"(%[[cast]]
// CHECK: %[[add2:.*]] = "tf.AddV2"(%[[rescale1]]
// CHECK: %[[maximum2:.*]] = "tf.Maximum"(%[[add2]]

// CHECK-LABEL: @main

  op: "custom_embedding_matmul"
}
library {
  function {
    signature {
      name: "custom_relu"
    }
    attr {
      key: "_implements"
      value {
        func {
          name: "tensorflow.relu"
        }
      }
    }
  }
  function {
    signature {
      name: "custom_embedding_matmul"
    }
    attr {
      key: "_implements"
      value {
        func {

  %14 = "tf.AddV2"(%10, %12) : (tensor<?x256x56x56xf32>, tensor<?x256x56x56xf32>) -> tensor<?x256x56x56xf32>
  %15 = "tf.Relu"(%14) : (tensor<?x256x56x56xf32>) -> tensor<?x256x56x56xf32>

  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%[[BATCH_NORM1]], %[[BATCH_NORM2]])
  // CHECK: %[[RELU:[0-9]*]] = "tf.Relu"(%[[ADD]])

  // Reduce spatial dimensions
  %16 = "tf.Mean"(%15, %1) : (tensor<?x256x56x56xf32>, tensor<2xi32>) -> tensor<?x256xf32>

  // CHECK: %[[TANH:[0-9]*]] = "tf.Tanh"(%[[ARG_TRANSPOSE]]) {{.*}} tensor<1x8x4x4xf32>
  // CHECK: %[[RELU:[0-9]*]] = "tf.Relu"(%[[TANH]]) {{.*}} tensor<1x8x4x4xf32>
  // CHECK: return %[[RELU]]

  %0 = "tf.Tanh"(%arg0) : (tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  %1 = "tf.Relu"(%0) : (tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>

  %2 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>

  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>
  %4 = "tf.BiasAdd"(%3, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  %5 = "tf.Relu"(%4) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>


  %6 = "tf.Conv2D"(%arg0, %arg1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],

// CHECK-SAME: f = @composite_conv3d_fn_1}>
// CHECK-NOT: {_tfl_quant_trait = "fully_quantizable"
// CHECK: %[[RELU:.*]] = "tf.Relu"(%[[PARTITIONEDCALL_0]])
// CHECK: return %[[RELU]]

// CHECK-LABEL: private @composite_conv3d_fn_1

// WEIGHTONLY-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<2x3x3x3x2xf32>
// WEIGHTONLY: %[[PARTITIONEDCALL_0:.*]] = "tf.PartitionedCall"(%arg0, %[[CST]])