%val1 = "tf.Identity"(%val0) : (tensor<96x!tf_type.f32ref>) -> tensor<96xf32>
  %val2 = "tf.Identity"(%val0) : (tensor<96x!tf_type.f32ref>) -> tensor<96xf32>
  func.return
}

// -----

// Test case: No follow-up ReadVariable case.

func.func @f() {
  // CHECK-NOT: "tf.VariableV2"
  // CHECK-NOT: "tf.VarHandleOp"

  func.return %0 : tensor<256x32x32x16xf32>
}

func.func @testConv2DDynamicShape(tensor<?x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<?x32x32x16xf32> {
^bb0(%arg0: tensor<?x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: _arithmetic_count = -1 : i64

  %1 = "tfl.pseudo_const"() { value = dense<1.0> : tensor<16xf32>} : () -> tensor<16xf32>
  %2 = "tfl.conv_3d"(%arg0, %0, %1) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, dilation_d_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, stride_d = 1 : i32} : (tensor<1x32x32x32x8xf32>, tensor<1x1x1x8x16x!quant.uniform<i8:f32, 1.0>>, tensor<16xf32>) -> tensor<1x32x32x32x16xf32>

  %1 = "tfl.sum"(%0, %cst) {keep_dims = false} : (tensor<2xf32>, tensor<1xi32>) -> tensor<f32>
  %2 = "tfl.add"(%1, %cst_1) {fused_activation_function = "NONE"} : (tensor<f32>, tensor<1xf32>) -> tensor<1xf32>
  %3 = "tfl.rsqrt"(%2) : (tensor<1xf32>) -> tensor<1xf32>
  %4 = "tfl.mul"(%arg0, %3) {fused_activation_function = "NONE"} : (tensor<2xf32>, tensor<1xf32>) -> tensor<2xf32>
  func.return %4: tensor<2xf32>

  // 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>)

  func.return %0 : tensor<256x32x32x16xf32>
}

func.func @testConv2DGPU(tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x32x32x16xf32> {
^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: tac.cost = 0x4C300000

      %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>
  }

// -----

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>

// CHECK-LABEL: func @testLeakyRelu(%arg0: tensor<16xf32>)
func.func @testLeakyRelu(tensor<16xf32>) -> tensor<16xf32> {
^bb0(%arg0: tensor<16xf32>):
  %0 = "tf.LeakyRelu"(%arg0) {alpha = 0.2 : f32} : (tensor<16xf32>) -> tensor<16xf32>
  func.return %0 : tensor<16xf32>
}

// -----
func.func @testLeakyWrongAlphaType(tensor<16xf32>) -> tensor<16xf32> {
^bb0(%arg0: tensor<16xf32>):

        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20xf32>, tensor<20xf32>, tensor<20xf32>, tensor<20xf32>,
        tensor<20x20xf32>, none,
        tensor<1x20xf32>, tensor<1x20xf32>,
        none, none, none, none) -> tensor<1x28x20xf32>
    %1 = "quantfork.stats"(%0) {layerStats = dense<[-1.0, 2.0]> : tensor<2xf32>} : (tensor<1x28x20xf32>) -> tensor<1x28x20xf32>