Coef was originally chosen to accomodate 12 bits/component with
4 extra precision intermediates + some under/overflow range, but
it turns out that 12 bits/component only uses 2 extra precision
intermediates, so we don't need coef.