session_options.config.intra_op_parallelism_threads()),
      use_multiple_streams_(options.use_multiple_streams),
      shape_determination_fns_(options.shape_determination_fns),
      allowed_devices_(options.allowed_devices),
      use_global_compute_stream_(options.use_global_compute_stream) {
  if (options.shape_determination_fns.empty()) {
    LOG(ERROR) << "shape_representation_fns must be non-empty.";
  }

  // platform will have resources allocated. For GPUs this will be
  // filled from visible_gpu_devices list from session configuration.
  std::optional<std::set<int>> allowed_devices_;

  const bool use_global_compute_stream_;

  // A static vector with device_ordinal as its index, describing the global
  // compute streams used in each XLA device. It is only used if

    %0 = "tf.VarHandleOp"() {_class = ["loc:@dense/kernel"], allowed_devices = [], container = "", device = "/job:worker/replica:0/task:1/device:CPU:0", shared_name = "var1"} : () -> tensor<!tf_type.resource<tensor<100x50xf32>>>
    %1 = "tf.ReadVariableOp"(%0) {device = ""} : (tensor<!tf_type.resource<tensor<100x50xf32>>>) -> tensor<100x50xf32>

  attributes {tf.entry_function = {control_outputs = "", inputs = "", outputs = "dense_2/Add:0"}, tf_saved_model.exported_names = ["serving_default"]} {
    %0 = "tf.VarHandleOp"() {_class = ["loc:@dense/kernel"], allowed_devices = [], container = "", device = "/job:worker/replica:0/task:1/device:CPU:0", shared_name = "var1"} : () -> tensor<!tf_type.resource<tensor<100x50xf32>>>

  TF_RETURN_IF_ERROR(op_ptr->SetAttrType("dtype", dtype));
  TF_RETURN_IF_ERROR(op_ptr->SetAttrShape("shape", shape));
  TF_RETURN_IF_ERROR(op_ptr->SetAttrStringList("allowed_devices", allowed_devices));
  int num_retvals = 1;
  return op_ptr->Execute(absl::MakeSpan(resource, 1), &num_retvals);
}

// Op: RestoreV2()
// Summary:
//
// Description:

    %VarHandleOp, %ctl = VarHandleOp device("/CPU:0") name("x") {_output_shapes = [#tf_type.shape<>], allowed_devices = [], container = "a", dtype = i64, shape = #tf_type.shape<>, shared_name = "x"} : () -> (tensor<!tf_type.resource<tensor<i64>>>)...

  "tf.TPUCompileSucceededAssert"(%compilation_status) {device = "/CPU:0"} : (tensor<!tf_type.string>) -> ()
  %0 = "tf.VarHandleOp"() {_xla_inferred_shapes = [#tf_type.shape<>], allowed_devices = [], container = "", device = "/CPU:0", shared_name = "y"} : () -> tensor<!tf_type.resource<tensor<3x1xf32>>>
  %1 = "tf.ReadVariableOp"(%0) {device = "/CPU:0"} : (tensor<!tf_type.resource<tensor<3x1xf32>>>) -> tensor<3x1xf32>

//
Status VarHandleOp(AbstractContext* ctx, AbstractTensorHandle** resource, DataType dtype, const PartialTensorShape shape, const char* container = "", const char* shared_name = "", const char* debug_name = "", absl::Span<string const> allowed_devices = {}, const char* name = nullptr, const char* raw_device_name = nullptr);

//

    options.device_name = DEVICE_XLA_GPU;
    options.device_ordinal = i;
    options.compilation_device_name = DEVICE_GPU_XLA_JIT;
    options.use_multiple_streams = true;
    options.allowed_devices = gpu_ids;
    XlaShapeLayoutHelpers::ShapeDeterminationFns shape_representation_fns{
        UseNoPreferenceLayoutFn(), IdentityShapeRepresentationFn()};
    options.shape_determination_fns = {shape_representation_fns};

}

// CHECK-LABEL: test_expensive_ops
func.func @test_expensive_ops(%arg: tensor<?x512xf32>) -> tensor<?x512xf32> {
    // expected-remark@+1 {{Cost: 1}}
    %0 = "tf.VarHandleOp"() {allowed_devices = [], container = "", device = "/job:localhost/replica:0/task:0/device:CPU:0", shared_name = "var"} : () -> tensor<!tf_type.resource<tensor<512x512xf32>>>
    // expected-remark@+1 {{Cost: 2}}