All the internal math is float, and avoiding the conversion from double is much faster when using SSE math.