This version explicitly performs the +0.5 step using double, which should avoid losing the low bits. -This version explicitly performs the +0.5 step using double, which should avoid losing the low bits.