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

Algorithm 7

The master algorithm for multi-GPU programming in OpenACC.

(1)     function  ENQUEUE_TASK_XXXX

(2)        Allocate memory and populate the task argument

(3)        Allocate memory and populate the task

(4)        pthread_mutex_lock(&cur_thread queue_lock)

(5)        if    then     ⊳push the task into the FIFO queue

(6)            cur_thread queue_head = cur_task;

(7)            cur_thread queue_tail = cur_task;

(8)            pthread_cond_signal(&cur_thread queue_ready); ⊳signal the worker that the queue is not empty and the task is ready

(9)        else

(10)          cur_thread queue_tail next = cur_task;

(11)           cur_thread queue_tail = cur_task;

(12)      end if

(13)      cur_thread queue_size++;

(14)      pthread_mutex_unlock(&cur_thread queue_lock);

(15)      if    then         ⊳if the task is synchronous

(16)          pthread_mutex_lock(&cur_thread queue_lock);

(17)          while    do    ⊳wait until this task is done

(18)              pthread_cond_wait(&cur_thread work_done, &cur_thread queue_lock);

(19)          end while

(20)         pthread_mutex_unlock(&cur_thread queue_lock);

(21)      end if

(22) end function