Thanks runrig, I didn't even try the hash method. I thought it would be too slow. I was surprised how quick it was.

l_runrig: 17 wallclock secs (17.23 usr + 0.00 sys = 17.23 CPU) @ 5943.12/s (n=102400)
leftover: 16 wallclock secs (15.86 usr + 0.00 sys = 15.86 CPU) @ 6456.49/s (n=102400)