// CHECK: tfl.add
    // CHECK-SAME: tac.skip_target_annotation
    %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    // CHECK: tfl.relu
    // CHECK-SAME: tac.skip_target_annotation
    %2 = "tfl.relu"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
    func.return
  }
}

// -----

  // 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>

    rewriter.replaceOp(batch_norm, op->getResults());

    // Depending on the case, we may fuse the add, the relu, or both.
    if (!add_op || add_op.getZ().hasOneUse()) {
      // We fuse the Relu only if the add has a single use, otherwise we only
      // fuse the add itself.
      op->setAttr("activation_mode", rewriter.getStringAttr("Relu"));
      rewriter.replaceOp(relu_op, op->getResult(0));
    }
    if (add_op) {

  %cpu = corert.get_op_handler %ch "cpu"
  %0 = corert.executeop(%cpu) "tf.Relu"(%arg0) { T = f32 } : 1
  %arg1 = tfrt_fallback_async.corert_tensorhandle_to_fallback_tensor %arg1_th {_tfrt_cost = 1 : i64, device = "/CPU:0"} : (!corert.tensorhandle) -> (!tfrt_fallback.tf_tensor)
  %1 = tfrt_fallback_async.executeop key(0) cost(100) device("/CPU:0") "tf.Relu"(%arg1) { T = f32 } : 1
  tfrt.return

// 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>