Avoid handling NaN when scaling float samples for output