int main ()
{
  struct Cascade *cascade;
  int barrier_view[wsz];
  int barrier_stream __attribute__((stream));
    
  #omp task cache (W:owm_pool[:SZ])
  {
    cascade = _malloc (sizeof (struct Cascade));

    foo (barrier_stream, cascade, ...);
    
    #omp task cache (R:owm_pool[:SZ]) \
              input (barrier_stream >> barrier_view[scaleIndexMax]
    {
      ........
    }

  // we need a taskwait here... taskwait doesn't work with COTSon
  // and this task wait couldn't be replaced by stream...
  #pragma omp taskwait
  }
}

void foo (int barrier_stream __attribute__((stream)), matrix, cascade,...)
{
    matrix[:] = ...;
    for (i = 0; i < scaleIndexMax; i++)
    {
      #pragma task omp output (barrier_stream) cache (R:matrix[:SZ])
      { 
        //task
      }
    }
    //other computations...
}