I'd like to hear your opinion on the question above too, IanB.
Also, two bugs in the core filters:
1.
ConvertYV16ToYUY2 - return dst; line is missing. Each frame ends up being processed by both mmx and C versions making the filter slower than it should be.
2.
AFH_YV12_C - leftmost pixel is not processed. x should start with 0 just like it does everywhere else.