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[tid] = 0;
    buf[dim+tid] = 0;
    buf[pout*dim+tid] = in[idx];
    barrier(CLK_LOCAL_MEM_FENCE);
    for (int offset = 1; offset < dim; offset *= 2) {
      pout = 1 - pout; // swap double buffer indices
      pin = 1 - pout;
      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);
    }
    out[idx] = buf[pout*dim+tid]; // write output
    barrier(CLK_LOCAL_MEM_FENCE);
    if (tid == n-1) {
      bout[gid] = buf[pout*dim+tid]; // write output
    }
  }