Don't assume we can do a clipped negation in 16 bit before the
multiplication (as it might affect the end result), but do the
multiplication first and negate in 32 bit, just like in the reference.