Multi-GPU Support on Single Node Using Directive-Based Programming Model

Algorithm 8

typedef struct _task_s

{

  int type; //task type (e.g. memory allocation and kernel launch, etc.)

  void  (routine)(void); // the task routine

  _work_args args; // point to the task argument

  int work_done; // indicate whether the task is done

  int async; // whether the task is asynchronous

  struct _task_s next; // next task in the task queue

}  _task;

typedef struct

{

  int destroyed; // whether this thread is destroyed

  int queue_size; // the task queue size

  pthread_t thread; // the thread identity

  context_t context; // the GPU context associated with this thread

  int context_id; // the GPU context id

  _task queue_head; // head of the FIFO task queue

  _task queue_tail; // tail of the FIFO task queue

  pthread_mutex_t queue_lock;

  pthread_cond_t queue_ready; // the task queue is not empty and ready

  pthread_cond_t work_done;

  pthread_cond_t queue_empty;

}  _gpu_thread;