Пример #1
0
static int
work_group_scheduler (kernel_run_command *k,
                      struct pool_thread_data *thread_data)
{
  void *arguments[k->kernel->num_args + k->kernel->num_locals];
  struct pocl_context pc;
  unsigned i;
  unsigned start_index;
  unsigned end_index;
  char last_wgs = 0;

  if (!get_wg_index_range (k, &start_index, &end_index,  &last_wgs))
    return 0;

  setup_kernel_arg_array ((void**)&arguments, k);
  memcpy (&pc, &k->pc, sizeof (struct pocl_context));
  do
    {
      if (last_wgs)
        {
          PTHREAD_LOCK (&scheduler.wq_lock, NULL);
          LL_DELETE (scheduler.kernel_queue, k);
          PTHREAD_UNLOCK (&scheduler.wq_lock);
        }
      for (i = start_index; i <= end_index; ++i)
        {
          translate_wg_index_to_3d_index (k, i, (size_t*)&pc.group_id);
#ifdef DEBUG_MT
          printf("### exec_wg: gid_x %d, gid_y %d, gid_z %d\n",
                 pc.group_id[0],
                 pc.group_id[1], pc.group_id[2]);
#endif
          k->workgroup (arguments, &pc);
        }

    }while (get_wg_index_range (k, &start_index, &end_index,  &last_wgs));


  free_kernel_arg_array (arguments, k);

  return 1;
}
Пример #2
0
int pthread_scheduler_get_work (thread_data *td, _cl_command_node **cmd_ptr)
{
  _cl_command_node *cmd;
  kernel_run_command *run_cmd;
  // execute kernel if available
  PTHREAD_LOCK (&scheduler.wq_lock, NULL);
  if ((run_cmd = scheduler.kernel_queue))
    {
      ++run_cmd->ref_count;
      PTHREAD_UNLOCK (&scheduler.wq_lock);

      work_group_scheduler (run_cmd, td);

      PTHREAD_LOCK (&scheduler.wq_lock, NULL);
      if (!(--run_cmd->ref_count))
        {
          PTHREAD_UNLOCK (&scheduler.wq_lock);
          finalize_kernel_command (td, run_cmd);
        }
      else
        PTHREAD_UNLOCK (&scheduler.wq_lock);
    }
  else
    PTHREAD_UNLOCK (&scheduler.wq_lock);

  // execute a command if available
  PTHREAD_LOCK (&scheduler.wq_lock, NULL);
  if ((cmd = scheduler.work_queue))
    {
      DL_DELETE (scheduler.work_queue, cmd);
      PTHREAD_UNLOCK (&scheduler.wq_lock);
      *cmd_ptr = cmd;
      return 0;
    }
  PTHREAD_UNLOCK (&scheduler.wq_lock);
  *cmd_ptr = NULL;
  return 1;
}
Пример #3
0
void pthread_scheduler_release_host ()
{
  PTHREAD_LOCK (&scheduler.cq_finished_lock, NULL);
  pthread_cond_signal (&scheduler.cq_finished_cond);
  PTHREAD_UNLOCK (&scheduler.cq_finished_lock);
}
Пример #4
0
/**
 * \brief Function executed by worker threads.
 */
static void* threadqueue_worker(void* threadqueue_opaque)
{
  threadqueue_queue_t * const threadqueue = (threadqueue_queue_t *) threadqueue_opaque;

  PTHREAD_LOCK(&threadqueue->lock);

  for (;;) {
    while (!threadqueue->stop && threadqueue->first == NULL) {
      // Wait until there is something to do in the queue.
      PTHREAD_COND_WAIT(&threadqueue->job_available, &threadqueue->lock);
    }

    if (threadqueue->stop) {
      break;
    }

    // Get a job and remove it from the queue.
    threadqueue_job_t *job = threadqueue_pop_job(threadqueue);

    PTHREAD_LOCK(&job->lock);
    assert(job->state == THREADQUEUE_JOB_STATE_READY);
    job->state = THREADQUEUE_JOB_STATE_RUNNING;
    PTHREAD_UNLOCK(&job->lock);
    PTHREAD_UNLOCK(&threadqueue->lock);

    job->fptr(job->arg);

    PTHREAD_LOCK(&threadqueue->lock);
    PTHREAD_LOCK(&job->lock);
    assert(job->state == THREADQUEUE_JOB_STATE_RUNNING);
    job->state = THREADQUEUE_JOB_STATE_DONE;

    PTHREAD_COND_SIGNAL(&threadqueue->job_done);

    // Go through all the jobs that depend on this one, decreasing their
    // ndepends. Count how many jobs can now start executing so we know how
    // many threads to wake up.
    int num_new_jobs = 0;
    for (int i = 0; i < job->rdepends_count; ++i) {
      threadqueue_job_t * const depjob = job->rdepends[i];
      // The dependency (job) is locked before the job depending on it.
      // This must be the same order as in kvz_threadqueue_job_dep_add.
      PTHREAD_LOCK(&depjob->lock);

      assert(depjob->state == THREADQUEUE_JOB_STATE_WAITING ||
             depjob->state == THREADQUEUE_JOB_STATE_PAUSED);
      assert(depjob->ndepends > 0);
      depjob->ndepends--;

      if (depjob->ndepends == 0 && depjob->state == THREADQUEUE_JOB_STATE_WAITING) {
        // Move the job to ready jobs.
        threadqueue_push_job(threadqueue, kvz_threadqueue_copy_ref(depjob));
        num_new_jobs++;
      }

      // Clear this reference to the job.
      PTHREAD_UNLOCK(&depjob->lock);
      kvz_threadqueue_free_job(&job->rdepends[i]);
    }
    job->rdepends_count = 0;

    PTHREAD_UNLOCK(&job->lock);
    kvz_threadqueue_free_job(&job);

    // The current thread will process one of the new jobs so we wake up
    // one threads less than the the number of new jobs.
    for (int i = 0; i < num_new_jobs - 1; i++) {
      pthread_cond_signal(&threadqueue->job_available);
    }
  }

  threadqueue->thread_running_count--;
  PTHREAD_UNLOCK(&threadqueue->lock);
  return NULL;
}