There seem to be three versions of FFDshow, regular, MMX, or SSE optimized. Which one should be used for which CPUs? Even my oldest computers (AthlonXP barton core 2500+) support both MMX and SSE (and my Athlon64 X2 supports SSE3 and MMX2 I think). What to get?

The regular FFDShow - compiled by drevil_xxl - autodetects the best optimizations. However, you may install what version you want, the differences aren't that major, anyway