Thanks for the tip. By the way, line 84 introduces some unneeded serialization; you could get some more speed there by fetching the bits in parallel.