}

  // Returns a device-to-device stream, in round-robin fashion.
  se::Stream* GetDeviceToDeviceStream();

  Status ThenExecute(Device* device, stream_executor::Stream* stream,
                     std::function<void()> func) override;

 private:
  bool UseMultipleStreams() const { return stream_ != host_to_device_stream_; }

  // The main compute stream of the device, used to synchronize the transfer

    %1 = "tf.Relu"(%0) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
    %2 = "tf.Add"(%0, %1) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf.Relu"(%arg1) {device = "tpu1"} : (tensor<?xi32>) -> tensor<?xi32>
    %4 = "tf.Add"(%3, %arg1) {device = "tpu1"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %5 = "tf.Relu"(%4) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>

  %0 = "tf.Const"() {device = "/device:CPU:0", value = dense<2> : tensor<i32>} : () -> tensor<i32>
  %1 = "tf.ReadVariableOp"(%arg) {device = "/device:CPU:0", dtype = i32} : (tensor<!tf_type.resource<tensor<i32>>>) -> tensor<i32>
  %2 = "tf.Add"(%1, %0) {device = "/device:CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>

                 SP_Event event, TF_Status* const status) {}
void WaitForEvent(const SP_Device* const device, SP_Stream stream,
                  SP_Event event, TF_Status* const status) {}
void CreateTimer(const SP_Device* const device, SP_Timer* timer,
                 TF_Status* const status) {}
void DestroyTimer(const SP_Device* const device, SP_Timer timer) {}

  %0 = "tf.Conv2D"(%arg0, %cst_1) {
    data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [],
    padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true
  } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>
  %1 = "tf.BiasAdd"(%0, %cst) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  %2 = "tf.Relu6"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>

func.func @conv2d_unmatching_loc_pattern(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>

  // CHECK: {{%.*}} = tfrt_fallback_async.executeop {{.*}} device("/device:CPU:0") "tf.MatMul"
  // CHECK-SAME: {T = f32, transpose_a = false, transpose_b = false}

    %2:3 = "tf.UnpackHyp"(%arg1) {_tpu_replicate = "cluster_Fn", _xla_outside_compilation = "0", device = "", max_seq_length = 16 : i64} : (tensor<*x!tf_type.string>) -> (tensor<*xi32>, tensor<*xi32>, tensor<*xf32>)
    %3 = "tf.Reshape"(%2#2, %1) {_tpu_replicate = "cluster_Fn", _xla_outside_compilation = "0", device = ""} : (tensor<*xf32>, tensor<*xi32>) -> tensor<*xf32>

  %dense_default_1 = "tf.Const"() {device = "/device:CPU:0", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>
  %dense_keys = "tf.Const"() {device = "/device:CPU:0", dtype = !tf_type.string, value = dense<""> : tensor<2x!tf_type.string>} : () -> tensor<2x!tf_type.string>

    %0 = "tf.Const"() {device = "/device:CPU:0", value = dense<0> : tensor<i64>} : () -> tensor<i64>
    %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<5> : tensor<i64>} : () -> tensor<i64>
    %2 = "tf.Const"() {device = "/device:CPU:0", value = dense<1> : tensor<i64>} : () -> tensor<i64>