Why not just output 64 timed pulses? Some special reason? You have to time *CS and U/*D anyway, because CS should be made inactive when you're not switching. At the very least, you'd have to know when it's safe to change direction. In other words, know when the fade operation is finished. I assume that a failure to reach the end point would not be good. Wouldn't it be better to "know" by controlling it rather than to "estimate" the timing?