%4 = "tfr.cast"(%concat) : (!tfr.tensor) -> tensor<3xf32>
  func.return %4 : tensor<3xf32>

// CHECK: %[[concat:.*]] = "tf.RiscConcat"(%arg0, %arg1, %arg2) : (tensor<f32>, tensor<f32>, tensor<f32>) -> tensor<*xf32>
// CHECK: %[[es:.*]] = "tf.EnsureShape"(%[[concat]]) <{shape = #tf_type.shape<3>}> : (tensor<*xf32>) -> tensor<3xf32>
// CHECK: return %[[es]] : tensor<3xf32>
}

// CHECK-LABEL: decompose_tf_split

func.func @multiply(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "vhlo.multiply_v1"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  func.return %0 : tensor<1xf32>
}

// CHECK:func.func private @multiply(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
// CHECK: %0 = "vhlo.multiply_v1"(%arg0, %arg1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
// CHECK: return %0 : tensor<1xf32>
// CHECK:}

        none, none, none,
        tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>,
        none, none,
        tensor<1x3xf32>, tensor<1x3xf32>,
        none, none, none, none) -> tensor<1x2x3xf32>
  %17 = "quantfork.stats"(%16) {layerStats = dense<[-0.1, 0.1]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  func.return %17 : tensor<1x2x3xf32>

    %0 = "tfl.pseudo_const"() {value = dense<0.962260901> : tensor<1xf32>} : () -> tensor<1xf32>
    %1 = func.call @func_0_GPU_FLOAT(%arg0, %0) {tac.device = "GPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<1x200x200x200xf32>, tensor<1xf32>) -> tensor<1x200x200x200xf32>
    %2 = "tfl.pseudo_const"() {value = dense<0.895973444> : tensor<1xf32>} : () -> tensor<1xf32>

    # CHECK: "tf_saved_model.global_tensor"() <{is_mutable, {{.*}} type = tensor<*xf32>, value = dense<0.000000e+00> : tensor<1xf32>}> {tf_saved_model.exported_names = ["v0"]} : () -> ()
    # CHECK: "tf_saved_model.global_tensor"() <{is_mutable, {{.*}} type = tensor<?xf32>, value = dense<[0.000000e+00, 1.000000e+00]> : tensor<2xf32>}> {tf_saved_model.exported_names = ["v1"]} : () -> ()
    self.v0 = tf.Variable([0.], shape=tf.TensorShape(None))

    %0 = "tf.Div"(%arg1, %cst_0) : (tensor<?x?xf32>, tensor<f32>) -> tensor<?x?xf32>
    %1 = "tf.AddV2"(%0, %cst_1) : (tensor<?x?xf32>, tensor<f32>) -> tensor<?x?xf32>
    %2 = "tf.Floor"(%1) : (tensor<?x?xf32>) -> tensor<?x?xf32>
    %3 = "tf.ClipByValue"(%2, %cst_5, %cst_6) : (tensor<?x?xf32>, tensor<f32>, tensor<f32>) -> tensor<?x?xf32>
    %4 = "tf.Cast"(%3) {Truncate = false} : (tensor<?x?xf32>) -> tensor<?x?xi8>

    %cst = "tf.Const"() <{value = dense<true> : tensor<i1>}> {device = ""} : () -> tensor<i1>
    %cst_0 = "tf.Const"() <{value = dense<[0.117216609, 0.933735609, 0.0728900209]> : tensor<3xf32>}> {device = ""} : () -> tensor<3xf32>

    %26 = "tf.Cast"(%25) {Truncate = false} : (tensor<2xi64>) -> tensor<2xf32>
    %27 = "tf.Equal"(%14, %26) {incompatible_shape_error = true} : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>
    %28 = "tf.Cast"(%27) {Truncate = false} : (tensor<2xi1>) -> tensor<2xf32>
    %29 = "tf.Sum"(%28, %6) {keep_dims = false} : (tensor<2xf32>, tensor<1xi32>) -> tensor<f32>

  %0 = "tfr.cast"(%arg0) : (tensor<1x3x!quant.uniform<i8:f32:1, {0.1:1, 0.2:2, 0.3:3}>>) -> !tfr.tensor
  %scale, %zp = tfr.quant_qparam(%0) : (!tfr.tensor) -> (!tfr.tensor, !tfr.tensor)
  %1 = "tfr.cast"(%scale) : (!tfr.tensor) -> tensor<3xf32>
  %2 = "tfr.cast"(%zp) : (!tfr.tensor) -> tensor<3xi32>
  func.return %1, %2 : tensor<3xf32>, tensor<3xi32>

  // CHECK-SAME:      <[-1.000000e+00, 1.000000e+00]> : tensor<2xf32>}>
  // CHECK-SAME:      : (tensor<40x37xf32>) -> tensor<40x37xf32>
  // CHECK-NEXT:    %[[cst:.*]] = "tfl.pseudo_const"() <{value = dense<
  // CHECK-SAME:      1.000000e+00> : tensor<40xf32>}> : () -> tensor<40xf32>
  // CHECK-NEXT:    %[[fc:.*]]:2 = "tfl.fully_connected"(%[[stat]], %arg1,