Downshift all the constants by one, and reduce the rounding shift by one.
This is mathematically equivalent since all constants are a multiple of two,
but allows for using 16-bit intermediates in the 1st pass of the 8-tap filter.