We used the SSE reciprocal square root instruction to vectorize the serch rather
than compare one at a time with multiplies. Speeds up the entire encoder by 8-10%.