func.func @broadcast_to_bf16(%arg0: tensor<3xbf16>, %arg1: tensor<2xi64>) -> tensor<3x3xbf16> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<3xbf16>, tensor<2xi64>) -> tensor<3x3xbf16>
  func.return %0: tensor<3x3xbf16>

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>

  %0 = "tf.Const"() {value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : () -> tensor<4xui64>
  // CHECK: tfrt_fallback_async.const_dense_tensor  dense<1.000000e+00> : tensor<1xbf16>
  %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<[1.0]> : tensor<1xbf16>} : () -> tensor<4xbf16>
  // CHECK: corert.executeop({{.*}}) "tf.Const"() {dtype = ui64, value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : 1

// CHECK:  return %[[V1]] : tensor<32x1x256xbf16>
}

// CHECK-LABEL: Relu_bf16
func.func @Relu_bf16(%arg0: tensor<2x3xbf16>) -> tensor<2x3xbf16> {
  %cst = arith.constant dense<0.0> : tensor<2x3xbf16>
  %0 = "tfl.maximum"(%arg0, %cst) : (tensor<2x3xbf16>, tensor<2x3xbf16>) -> tensor<2x3xbf16>
  func.return %0 : tensor<2x3xbf16>

  // CHECK: %[[RESULT:.*]] = "tfl.relu"(%arg0)
  // CHECK: return %[[RESULT]]
}

    // CHECK: %[[CAST_1:.*]] = "tf.Cast"(%[[ARG_1]]) <{Truncate = false}> : (tensor<10xf32>) -> tensor<10xbf16>
    // CHECK: %[[RESULT:.*]]:2 = "tf.XlaCallModule"(%[[CAST_0]], %[[CAST_1]], %[[ARG_2]])
    // CHECK-SAME: (tensor<10xbf16>, tensor<10xbf16>, tensor<6xi32>) -> (tensor<10xbf16>, tensor<6xi32>)

  %8 = "tfl.mul"(%2, %3) {fused_activation_function = "NONE"} : (tensor<4xbf16>, tensor<4xbf16>) -> tensor<4xbf16>

  func.return %5, %6, %7, %8 : tensor<bf16>, tensor<4xbf16>, tensor<4xbf16>, tensor<4xbf16>
}

// CHECK-LABEL: @mul_f16
func.func @mul_f16() -> (tensor<f16>, tensor<4xf16>, tensor<4xf16>, tensor<4xf16>) {
  %0 = arith.constant dense<4.5> : tensor<f16>

  // CHECK: %[[DQ_8:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_9:.*]] = "tfl.dequantize"({{.*}}) : (tensor<3xf16>) -> tensor<3xf32>
  // CHECK: %[[DQ_10:.*]] = "tfl.dequantize"({{.*}}) : (tensor<3xf16>) -> tensor<3xf32>
  // CHECK: %[[DQ_11:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x3xf16>) -> tensor<1x3xf32>

    %1 = "tf.Cast"(%0) <{truncate = false}> : (tensor<3x1xf32>) -> tensor<3x1xbf16>
    %2 = "tf.VarHandleOp"() <{container = "", shared_name = "y"}> : () -> tensor<!tf_type.resource<tensor<3x1xbf16>>>
    "tf.AssignVariableOp"(%2, %1) : (tensor<!tf_type.resource<tensor<3x1xbf16>>>, tensor<3x1xbf16>) -> ()
    return
  }

  func.return %0 : tensor<4xcomplex<f64>>
}

func.func @f16() -> tensor<4xf16> {
  // CHECK-LABEL: @f16
  // CHECK: value = dense<[1.000000e+00, 2.000000e+00, 3.000000e+00, 4.000000e+00]> : tensor<4xf16>
  %0 = "tfl.pseudo_const"() { value = dense<[1.0, 2.0, 3.0, 4.0]> : tensor<4xf16> } : () -> tensor<4xf16>
  func.return %0 : tensor<4xf16>
}

func.func @f32() -> tensor<4xf32> {
  // CHECK-LABEL: @f32

//
module {
  func.func @serving_default() -> tensor<1x3xbf16> {
    %cst = "tf.Const"() <{value = dense<[1]> : tensor<1xi32>}> : () -> tensor<1xi32>
    %0 = "tf.VarHandleOp"() <{container = "", shared_name = "Variable"}> : () -> tensor<!tf_type.resource<tensor<2x3xbf16>>>
    %1 = "tf.ResourceGather"(%0, %cst) <{batch_dims = 0 : i64, validate_indices = true}> : (tensor<!tf_type.resource<tensor<2x3xbf16>>>, tensor<1xi32>) -> tensor<1x3xbf16>

    func.return %0, %1, %2 : tensor<3xf16>, tensor<3xf32>, tensor<3xf64>
  }

  // CHECK-LABEL: bessel_i1e
  func.func @bessel_i1e(%arg0: tensor<3xf16>, %arg1: tensor<3xf32>, %arg2: tensor<3xf64>) -> (tensor<3xf16>, tensor<3xf32>, tensor<3xf64>) {
    // CHECK-NOT: tf.BesselI1e
    %0 = "tf.BesselI1e"(%arg0) : (tensor<3xf16>) -> (tensor<3xf16>)
    %1 = "tf.BesselI1e"(%arg1) : (tensor<3xf32>) -> (tensor<3xf32>)