func.func @device_test(%arg0: tensor<3x1xf32>) -> (tensor<3x3xf32>) {

  // CHECK: device = "gpu"
  %0 = "tf.Const"() {value = dense<[[1.0, 2.0, 3.0]]> : tensor<1x3xf32>} : () -> tensor<1x3xf32>
  // CHECK: device = "gpu"
  %1 = "tf.MatMul"(%arg0, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>
  // CHECK: device = "cpu"

  %1 = "tf.ReadVariableOp"(%0) {device = "/CPU:0"} : (tensor<!tf_type.resource<tensor<1x1xf32>>>) -> tensor<1x1xf32>
  %2:2 = "tf.Split"(%cst, %arg0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x4xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)
  %3:2 = "tf.Split"(%cst, %2#0) {device = "/CPU:0"} : (tensor<i32>,  tensor<1x2xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>)

// CHECK: %0 = "tfl.pseudo_const"() <{value = dense<2> : tensor<2xi32>}> : () -> tensor<2xi32>
// CHECK: %1 = "tfl.reshape"(%arg0, %0) : (tensor<1x4xf32>, tensor<2xi32>) -> tensor<2x2xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "RESHAPE"
    }
  ],
  "subgraphs": [
    {
      "tensors": [
        {
          "shape": [1, 4],

# CHECK:      func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME:   [[ARG0:%.*]]: tensor<3x1xf32> {tf_saved_model.index_path = ["x"]},
# CHECK-SAME:   [[ARG1:%.*]]: tensor<!tf_type.resource<tensor<1x3xf32>>> {tf_saved_model.bound_input = @[[VAR]]})
# CHECK-SAME:             -> (tensor<3x3xf32> {tf_saved_model.index_path = ["r"]})

  func.return
}

// CHECK-LABEL: func @basic
func.func @basic(
    %arg0: tensor<3x1xf32>,
    %arg1: tensor<!tf_type.resource<tensor<1x3xf32>>>) -> (tensor<3x3xf32>) {
  %1 = "tf.ReadVariableOp"(%arg1) {_output_shapes = ["tfshape$dim { size: 1 } dim { size: 3 }"], device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>

  func.return %2 : tensor<3x3xf32>

    %11 = "tf.Einsum"(%8, %10) {device = "", equation = "ab,bc->ac"} : (tensor<2x3xi32>, tensor<3x4xi32>) -> tensor<2x4xi32>
    %12 = "tf.Cast"(%11) {Truncate = false, device = ""} : (tensor<2x4xi32>) -> tensor<2x4xf32>
    %13 = "tf.Mul"(%12, %cst_0) {device = ""} : (tensor<2x4xf32>, tensor<f32>) -> tensor<2x4xf32>
    %14 = "tf.Relu"(%13) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>

  // CHECK-NEXT: [[r1:%.*]] = tfrt_fallback_async.executeop {{.*}} "tf.BiasAdd"([[r0]], [[result]])
  %3 = "tf.BiasAdd"(%2, %0) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], data_format = "NHWC", device = "/device:CPU:0"} : (tensor<3x3xf32>, tensor<3xf32>) -> tensor<3x3xf32>

For example, if we have the code

```mlir
  %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
```

then running this pass with 'default-device=foobar', we get:

```mlir

  %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>