当 CUDA 线程处于同一循环中并且我们需要同步它们以仅执行有限的部分时，如何同步它们

Question

我已经编写了一些代码，现在我想在 CUDA GPU 上实现它，但我对同步还不熟悉。下面我展示了代码，我希望 LOOP1 由所有线程执行（因此我希望这部分利用 CUDA，而其余部分（LOOP1 之外的部分）仅由单线程。

 do{
  point_set = master_Q[(*num_mas) - 1].q;
  List* temp = point_set;
  List* pa = point_set;
  if(master_Q[num_mas[0] - 1].max)
   max_level =  (int) (ceilf(il2 * log(master_Q[num_mas[0] - 1].max)));
  *num_mas = (*num_mas) - 1;
  while(point_set){
   List* insert_ele = temp;
   while(temp){
    insert_ele = temp;
    if((insert_ele->dist[insert_ele->dist_index-1] <= pow(2, max_level-1)) || (top_level == max_level)){
     if(point_set == temp){
      point_set = temp->next;
      pa  = temp->next;
     }
     else{
      pa->next = temp->next;
     }
     temp = NULL;
     List* new_point_set = point_set;
     float maximum_dist = 0;
     if(parent->p_index != insert_ele->point_index){
      List* tmp = new_point_set;
      float *b = &(data[(insert_ele->point_index)*point_len]);
 **LOOP 1:**     while(tmp){
       float *c = &(data[(tmp->point_index)*point_len]);
         float sum = 0.;
         for(int j = 0; j < point_len; j+=2){
        float d1 = b[j] - c[j];
        float d2 = b[j+1] - c[j+1];
        d1 *= d1;
        d2 *= d2;
        sum = sum + d1 + d2;
           }
          tmp->dist[tmp->dist_index] = sqrt(sum);
       if(maximum_dist < tmp->dist[tmp->dist_index])
           maximum_dist = tmp->dist[tmp->dist_index];
          tmp->dist_index = tmp->dist_index+1;
          tmp = tmp->next;
      }
      max_distance = maximum_dist;
     }
     while(new_point_set || insert_ele){
      List* far, *par, *tmp, *tmp_new;
      far = NULL;
      tmp = new_point_set;
      tmp_new = NULL;
      float level_dist = pow(2, max_level-1);
      float maxdist = 0, maxp = 0;
      while(tmp){
       if(tmp->dist[(tmp->dist_index)-1] > level_dist){
        if(maxdist < tmp->dist[tmp->dist_index-1])
         maxdist = tmp->dist[tmp->dist_index-1];
        if(tmp == new_point_set){
         new_point_set = tmp->next;
         par = tmp->next;
        }
        else{
         par->next = tmp->next;
        }
        if(far == NULL){
         far = tmp;
         tmp_new = far;
        }
        else{
         tmp_new->next = tmp;
         tmp_new = tmp;
        }
        if(parent->p_index != insert_ele->point_index)
         tmp->dist_index = tmp->dist_index - 1;
        tmp = tmp->next;
        tmp_new->next = NULL;
       }
       else{
        par = tmp;
        if(maxp < tmp->dist[(tmp->dist_index)-1])
         maxp = tmp->dist[(tmp->dist_index)-1];
        tmp = tmp->next;
       }

      }
      if(0 == maxp){    
       tmp = new_point_set;
       aloc_mem[*tree_index].p_index = insert_ele->point_index;
       aloc_mem[*tree_index].no_child = 0;
       aloc_mem[*tree_index].level = max_level--;
       parent->children_index[parent->no_child++] = *tree_index;
       parent = &(aloc_mem[*tree_index]);
       tree_index[0] = tree_index[0]+1;
       while(tmp){
        aloc_mem[*tree_index].p_index = tmp->point_index;
        aloc_mem[(*tree_index)].no_child = 0;
        aloc_mem[(*tree_index)].level = master_Q[(*cur_count_Q)-1].level;
        parent->children_index[parent->no_child] = *tree_index;
        parent->no_child = parent->no_child + 1;
        (*tree_index)++;
        tmp = tmp->next;
       }
         cur_count_Q[0] = cur_count_Q[0]-1;
         new_point_set = NULL;
      }
      master_Q[*num_mas].q = far;
      master_Q[*num_mas].parent = parent; 
      master_Q[*num_mas].valid = true;
      master_Q[*num_mas].max = maxdist;
      master_Q[*num_mas].level = max_level;

             num_mas[0] = num_mas[0]+1;
      if(0 != maxp){
       aloc_mem[*tree_index].p_index = insert_ele->point_index;
       aloc_mem[*tree_index].no_child = 0;
       aloc_mem[*tree_index].level = max_level;
       parent->children_index[parent->no_child++] = *tree_index;
       parent = &(aloc_mem[*tree_index]);
       tree_index[0] = tree_index[0]+1;
       if(maxp){
        int new_level = ((int) (ceilf(il2 * log(maxp)))) +1;
        if (new_level < (max_level-1))
         max_level = new_level;
        else
         max_level--;
       }
       else
        max_level--;
      }
      if( 0 == maxp )
       insert_ele = NULL;
     }

    }
    else{
     if(NULL == temp->next){
      master_Q[*num_mas].q = point_set;
      master_Q[*num_mas].parent = parent; 
      master_Q[*num_mas].valid = true;
      master_Q[*num_mas].level = max_level;
      num_mas[0] = num_mas[0]+1;
     }
     pa = temp;
     temp = temp->next;
    }
   }
   if((*num_mas) > 1){
    List *temp2 = master_Q[(*num_mas)-1].q;
    while(temp2){
     List* temp3 = master_Q[(*num_mas)-2].q;
     master_Q[(*num_mas)-2].q = temp2;
     if((master_Q[(*num_mas)-1].parent)->p_index != (master_Q[(*num_mas)-2].parent)->p_index){
      temp2->dist_index = temp2->dist_index - 1;
     }
     temp2 = temp2->next;
     master_Q[(*num_mas)-2].q->next = temp3;
    }
    num_mas[0] = num_mas[0]-1;
   }
   point_set = master_Q[(*num_mas)-1].q;
   temp = point_set;
   pa = point_set;
   parent = master_Q[(*num_mas)-1].parent;
   max_level = master_Q[(*num_mas)-1].level;
   if(master_Q[(*num_mas)-1].max)
    if( max_level > ((int) (ceilf(il2 * log(master_Q[(*num_mas)-1].max)))) +1)
        max_level = ((int) (ceilf(il2 * log(master_Q[(*num_mas)-1].max)))) +1;
   num_mas[0] = num_mas[0]-1;
  }

 }while(*num_mas > 0);

Answer 1

如果您希望单个内核仅在块内的单个线程上执行其部分代码，请使用 if 语句仅针对单个 threadIdx 执行部分代码，然后进行屏障同步。也许您应该尝试编写一个内核并将其发布以供人们查看。