%4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  func.return %1 : tensor<1xf32>
}

func.func @cond_true(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
  %0 = tfl.add %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

func.func @cond_false(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {

// CHECK-NOT: tf.BatchFunction
// CHECK: %[[ADD0:.*]] = "tf.AddV2"(%[[ARG0]], %[[ARG1]])
// CHECK: return %[[ADD0]] : tensor<1xf32>

func.func private @batched_func(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Identity"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
  %1 = "tf.Identity"(%arg1) : (tensor<1xf32>) -> tensor<1xf32>

    //body
    ^bb0(%arg_body: tensor<1xf32>):
    %result_body = "tfl.add"(%arg_body, %arg_body) { fused_activation_function = "NONE" } : (tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>)
    "tfl.yield"(%result_body) : (tensor<1xf32>) -> ()
  }) : (tensor<1xf32>) -> (tensor<1xf32>)
  %tmp5 = "tfl.add"(%tmp4, %tmp2) { fused_activation_function = "NONE" } : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

// -----

func.func @testAddReluPack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) {
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %fill = "tf.Fill"(%cst_0, %cst) : (tensor<1xi64>, tensor<f32>) -> tensor<1xf32>
  tf_device.replicate([%0, %fill] as %arg_r0: tensor<1xf32>) {n = 2 : i32} {
    %1 = "tf_device.launch"() <{device = "TPU_REPLICATED_HOST_0"}> ({
      %2 = "tf.Identity"(%arg_r0) : (tensor<1xf32>) -> tensor<1xf32>
      tf_device.return %2 : tensor<1xf32>
    }) : () -> tensor<1xf32>

  func.func @testFunctionSkiped(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) {
    // CHECK: tfl.add
    // CHECK-SAME: tac.skip_target_annotation
    %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    // CHECK: tfl.add
    // CHECK-SAME: tac.skip_target_annotation
    %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %2 = "tfl.add"(%arg0, %arg3) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>

  # Note that semantically in Python, multiple return values are equivalent
  # to returning a tuple/list.
  #
  # CHECK:      func {{@[a-zA-Z_0-9]+}}() -> (
  # CHECK-SAME:   tensor<1xf32> {tf_saved_model.index_path = [0]},
  # CHECK-SAME:   tensor<2xf32> {tf_saved_model.index_path = [1]})
  # CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["f0001_multiple_results_no_punctuation"]
  @tf.function(input_signature=[])

// CHECK-LABEL: func @main
func.func @main(%arg0: tensor<1xf32>) -> tensor<1xf32>
attributes  {tf.entry_function = {inputs = "input", outputs = "output_node"}} {
  %0 = arith.constant dense<2.000000e+00> : tensor<f32>
  %1 = arith.constant dense<1.000000e+00> : tensor<f32>
  %2 = "tf.AddV2"(%arg0, %1) {T = "tfdtype$DT_FLOAT", device = "", name = "StatefulPartitionedCall/add"} : (tensor<1xf32>, tensor<f32>) -> tensor<1xf32>