// Use other output
  %3:6 = "tf.FusedBatchNormV3"( %2#0, %arg1, %arg2, %arg3, %arg4) {T = "tfdtype$DT_FLOAT", U = "tfdtype$DT_FLOAT", data_format = "NHWC", epsilon = 0.001 : f32, is_training = false} : (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>) -> (tensor<8x8x8x8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>, tensor<8xf32>)

  %4 = "tfl.minimum"(%0, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>
  %5 = "tfl.minimum"(%1, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>
  %6 = "tfl.minimum"(%2, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>
  %7 = "tfl.minimum"(%3, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xf32>

}

// CHECK-LABEL: func @add_f32
func.func @add_f32(%arg0: tensor<3xf32>, %arg1: tensor<3xf32>) -> tensor<3xf32> {
  // CHECK-DAG: [[ADD:%.+]] = "tf.AddV2"(%arg0, %arg1) : (tensor<3xf32>, tensor<3xf32>) -> tensor<3xf32>
  %add = "tf.Add"(%arg0, %arg1) : (tensor<3xf32>, tensor<3xf32>) -> tensor<3xf32>
  func.return %add : tensor<3xf32>
}

// CHECK-LABEL: func @BiasAddGrad_NHWC

  %1 = "tf.Mul"(%arg1, %arg1) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
  %2 = "tf.Div"(%arg1, %arg1) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>

  // CHECK: "tf.If"{{.+}}else_branch = @tf.IfRegion_else{{.+}}then_branch = @tf.IfRegion_then
  %3 = "tf.IfRegion"(%arg0) ({
     %4 = "tf.Select"(%arg0, %0, %1) : (tensor<i1>, tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
     "tf.Yield"(%4) : (tensor<2xf32>) -> ()

// CHECK-LABEL: func @init
func.func @init(%arg0: tensor<4xf32>) -> tensor<7xf32> {
    %0 = builtin.unrealized_conversion_cast to tensor<7xf32>
    return %0 : tensor<7xf32>
}

// CHECK-LABEL: func @next
func.func @next(%arg0: tensor<7xf32>, %arg1: tensor<3xf32>) -> tensor<6xf32> {
    %0 = builtin.unrealized_conversion_cast to tensor<6xf32>
    return %0 : tensor<6xf32>
}

// CHECK-LABEL: func @finalize

func.func @fakeQuantPerChannelForActivation(%arg0: tensor<8x4xf32>) -> (tensor<8x4xf32>) {
  %arg1 = arith.constant dense<[0.0, -1.0, 1.0, 0.0]> : tensor<4xf32>
  %arg2 = arith.constant dense<[15.0, 14.0, 16.0, 1.0e-9]> : tensor<4xf32>
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) {num_bits = 3, narrow_range = false} : (tensor<8x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<8x4xf32>
  func.return %0 : tensor<8x4xf32>

      %outputs_6, %control_7 = tf_executor.island wraps "tf.BiasAdd"(%outputs_4, %outputs_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<16xf32>) -> tensor<*xf32>
      tf_executor.fetch %outputs_6 : tensor<*xf32>
    }
    func.return %0 : tensor<*xf32>
  }

  %1 = "tfl.add"(%0, %cst2) {fused_activation_function = "NONE"} : (tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4xf32>

  func.return %1 : tensor<4x4xf32>

  // CHECK-DAG: %[[bias:.*]] = arith.constant dense<{{.*}}> : tensor<4xf32>
  // CHECK-DAG: %[[filter:.*]] = arith.constant dense<{{.*}}> : tensor<4x4xf32>

}

// CHECK-LABEL: func @testPow
// CHECK-SAME:(%[[ARG_0:.*]]: tensor<4xf32>, %[[ARG_1:.*]]: tensor<4xf32>) -> (tensor<4xf32>, tensor<4xf32>, tensor<4xf32>)
func.func @testPow(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> (tensor<4xf32>, tensor<4xf32>, tensor<4xf32>) {

  %cst_zero = arith.constant dense<0.0> : tensor<f32>
  %cst_one = arith.constant dense<1.0> : tensor<f32>

func.func @fakeQuantPerChannelForActivation(%arg0: tensor<8x4xf32>) -> (tensor<8x4xf32>) {
  %arg1 = arith.constant dense<[0.0, -1.0, 1.0, 0.0]> : tensor<4xf32>
  %arg2 = arith.constant dense<[255.0, 254.0, 256.0, 1.0e-9]> : tensor<4xf32>
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) {num_bits = 5, narrow_range = false} : (tensor<8x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<8x4xf32>
  func.return %0 : tensor<8x4xf32>