TF_Status* status);
// The returned string remains valid throughout the lifetime of 'op'.
TF_CAPI_EXPORT extern const char* TFE_OpGetDevice(const TFE_Op* op,
                                                  TF_Status* status);

TF_CAPI_EXPORT extern void TFE_OpAddInput(TFE_Op* op, TFE_TensorHandle* input,

    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    const char* device_name = TFE_OpGetDevice(matmul, status);
    ASSERT_TRUE(strstr(device_name, "GPU:0") != nullptr);

    TFE_OpSetDevice(matmul, "CPU:0", status);
    ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    device_name = TFE_OpGetDevice(matmul, status);
    ASSERT_TRUE(strstr(device_name, "CPU:0") != nullptr);
  }

void ParallelDeviceExecute(const TFE_Op* original_op, int* num_outputs,
                           TFE_TensorHandle** outputs, TF_Status* status,
                           void* device_info) {
  const char* requested_placement = TFE_OpGetDevice(original_op, status);
  if (*requested_placement == '\0') {
    TF_SetStatus(
        status, TF_INTERNAL,
        "Ops must be placed on the parallel device explicitly, or their inputs "

}

void TFE_OpSetDevice(TFE_Op* op, const char* device_name, TF_Status* status) {
  status->status = tensorflow::unwrap(op)->SetDeviceName(device_name);
}

const char* TFE_OpGetDevice(const TFE_Op* op, TF_Status* status) {
  return tensorflow::unwrap(op)->DeviceName().c_str();
}

void TFE_OpAddInput(TFE_Op* op, TFE_TensorHandle* input, TF_Status* status) {

  // Method to execute an operation.
  //
  // Arguments provide enough information to reconstruct the original `TFE_Op`,
  // or construct a transformed version, by inspecting the passed `op`.
  //
  // TFE_OpGetDevice(op) records the original placement of the operation. It may
  // be an empty string if no device was explicitly requested, but will
  // otherwise be the name of this custom device. Ops are placed onto a custom