Reducing the chunksize for the diff routines very slightly reduces the
amount of data copied; it also should not change performance for any
of the SIMD optimized routines, since the instructions executed and
branches taken should not change.