Terminating problem from parallel_reduce

[nuntawat]

See code snipet below:

[cpp]for(list::iterator iter = (*block_list).begin(); iter != (*block_list).end(); iter++)
{
  // Do someing
}[/cpp]

When for loop iterate all member inside block_list variable, ideally, it should be stopped and exited, but I found that the new thread has come and start iteration again! How do I terminate all threads when it ends iteration? Thanks very much.

[robert-reed]

Quoting - nuntawat

See code snipet below:

[cpp]for(list::iterator iter = (*block_list).begin(); iter != (*block_list).end(); iter++)
{
  // Do someing
}[/cpp]

When for loop iterate all member inside block_list variable, ideally, it should be stopped and exited, but I found that the new thread has come and start iteration again! How do I terminate all threads when it ends iteration? Thanks very much.

I fear there's just not enough context here to make much of a guess about what may be going on with your code. Could you provide more of the details, showing at least how you engaged your iterator in the context of a parallel_reduce call?

[nuntawat]

Quoting - Robert Reed (Intel)

I fear there's just not enough context here to make much of a guess about what may be going on with your code. Could you provide more of the details, showing at least how you engaged your iterator in the context of a parallel_reduce call?

There is a lot of codes, but I captured only codes used to determine condition and iterate loop inside.

[cpp]
public:
  static currentMutex_t mutex;
  static point_t current_point;



void operator() (const blocked_range& range) const
{
  // Begin Queuing Mutex Lock
  currentMutex_t::scoped_lock lock(mutex, false);
  int block_num = 0;

  for(list::iterator iter = (*block_list).begin(); iter != (*block_list).end(); iter++)
  {
    local_work(*iter, &block_num, lock);
  }
}

void local_work(unsigned char **xc, int *block_num, currentMutex_t::scoped_lock &lock) const
{
  int tmp_x_pos, tmp_y_pos;

  tmp_x_pos = current_point.x_pos;
  tmp_y_pos = current_point.y_pos;

  for(int i = tmp_y_pos; i < tmp_y_pos + pixel_per_block; i++)
  {
    for(int j = tmp_x_pos; j < tmp_x_pos + pixel_per_block; j++)
    {
      // Do sometasks...
    }
  }



  if((current_point.x_pos + pixel_per_block) < ncols)
  {
    if(((current_point.y_pos + pixel_per_block) <= nrows) || (current_point.y_pos == nrows))
    {
      lock.scoped_lock::upgrade_to_writer();

      current_point.x_pos += pixel_per_block;
      current_point.y_pos -= pixel_per_block;

      lock.downgrade_to_reader();
    }
  }
  else if (current_point.y_pos != nrows)
  {
    lock.scoped_lock::upgrade_to_writer();

    current_point.x_pos = 0;
    (*block_num)++;
			
    lock.downgrade_to_reader();
  }
}
[/cpp]

[RafSchietekat]

I have only briefly glanced at the code, but it seems highly problematic that you want to use blocked_range without actually using the parameter range. That's after I never understood what a linear list was doing here, and I still don't see it. Plus there's the current_point navigation: are you certain this wouldn't work better with blocked_range2d on the pixels?