int pout, pin;
  pout = 0;
  pin = 1;
  // Load input into shared memory.
  // This is exclusive scan, so shift right by one
  // and set first element to 0
  if (idx <= n) {
    buf[pout*dim+tid] = (idx > 0) ? in[idx-1] : 0;
    barrier(CLK_LOCAL_MEM_FENCE);
    for (int offset = 1; offset < dim; offset *= 2) {
      pout = 1 - pout; // swap double buffer indices
      pin = 1 - pout;
      if (offset <= pin*dim+tid) {
        if (tid >= offset) {
          buf[pout*dim+tid] += buf[pin*dim+tid - offset];
        } else {
          buf[pout*dim+tid] = buf[pin*dim+tid];
        }
      }
      barrier(CLK_LOCAL_MEM_FENCE);
    }
    if (idx < pout*dim+dim+1) {
      out[idx] = buf[pout*dim+tid+1]; // write output
    }
  }