You didn't read the thread. The OPs code used very small arrays and using numpy was slowing the code down by an order of magnitude. The pure python solution is 17x faster.