func.func @main(%arg0: tensor<3x2xf32>) -> tensor<3x2xi32> {
  // CHECK-LABEL: @main
  // CHECK: "tfl.bucketize"(%arg0) <{boundaries = [0.000000e+00 : f32, 1.000000e+01 : f32, 1.000000e+02 : f32]}> : (tensor<3x2xf32>) -> tensor<3x2xi32>
  %0 = "tfl.bucketize"(%arg0) {boundaries = [0.0 : f32, 10.0 : f32, 100.0 : f32]} : (tensor<3x2xf32>) -> tensor<3x2xi32>
  func.return %0 : tensor<3x2xi32>

  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %3 : tensor<2x1xf32>
}

// CHECK: %[[CST:.*]] = arith.constant dense<1> : tensor<4xi32>

  }

  // CHECK-LABEL: pack
  func.func @pack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
    func.return %0 : tensor<2x1xf32>
    // CHECK: %[[VAL_0:.*]] = arith.constant dense<[2, 1]> : tensor<2xi32>

  // CHECK: %[[RES_INT_2:.*]] = mhlo.convert %[[RES_INT_1]] : tensor<3x2xi32>
  %result_int = "tf.Cast"(%result) {Truncate = false} : (tensor<3x2x!tf_type.qint32>) -> tensor<3x2xi32>
  // CHECK: return %[[RES_INT_2]] : tensor<3x2xi32>
  func.return %result_int : tensor<3x2xi32>

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x2xbf16>) -> tensor<1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  return %3 : tensor<1x2xf32>
}

// CHECK: func @cast_bf16_matmul_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>
// CHECK: %[[matmul:.*]] = "tf.MatMul"(%arg0, %[[cst]])

func.func @TfBatchMatMulV2(%arg0: tensor<4x128x2xf32>, %arg1:  tensor<2x1xf32>) -> tensor<4x128x1xf32> {
  %0 = "tfl.custom"(%arg0, %arg1) {
    custom_code = "FlexBatchMatMulV2",
    custom_option = #tfl<const_bytes : "0x0D42617463684D61744D756C56320038120D42617463684D61744D756C56321A001A002A070A0154120230012A0B0A0561646A5F78120228002A0B0A0561646A5F791202280032000002493B1414042801">
  } : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>

  %cst2 = arith.constant dense<3.0> : tensor<23x2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<1x112x112x3xf32>, tensor<1x3x3x2xf32>) -> tensor<1x28x23x2xf32>
  %1 = "tf.Mul"(%0, %cst2) : (tensor<1x28x23x2xf32>, tensor<23x2xf32>) -> tensor<1x28x23x2xf32>

  func.return %1 : tensor<1x28x23x2xf32>

  %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>
  func.return %3 : tensor<2x1xf32>
})";
  const std::string kExpectedFB = CreateRuntimeMetadata();
  mlir::DialectRegistry registry;
  registry.insert<mlir::TFL::TensorFlowLiteDialect, mlir::arith::ArithDialect,

// CHECK-LABEL: dont_move_transpose_different_ranks
func.func @dont_move_transpose_different_ranks(%arg0:tensor<1x1x2x3xf32>, %arg1:tensor<2x3xf32>) -> tensor<1x2x1x3xf32> {
  %cst = "tf.Const"() {value = dense<[0, 2, 1, 3]> : tensor<4xi32>} : () -> tensor<4xi32>
  %0 = "tf.AddV2"(%arg0, %arg1) {device = ""} : (tensor<1x1x2x3xf32>, tensor<2x3xf32>) -> tensor<1x1x2x3xf32>
  %1 = "tf.Transpose"(%0, %cst) {device = ""} : (tensor<1x1x2x3xf32>, tensor<4xi32>) -> tensor<1x2x1x3xf32>

  %cst0 = mhlo.constant dense<[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]> : tensor<2x3xf32>
  %cst1 = mhlo.constant dense<[[[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], [[7.0, 8.0, 9.0], [10.0, 11.0, 12.0]]]]> : tensor<1x2x2x3xf32>
  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<[1, 3]> : tensor<2xi64>}> : (tensor<2x3xf32>) -> tensor<1x2x2x3xf32>
  %1 = mhlo.multiply %0, %cst1 : tensor<1x2x2x3xf32>