Results: Yes, if I'm reading things right (and assuming the unittest caught any issues) you are much faster at large cardinalities! Your approach seems to be a constant .7 cycle/op, while Wojciech's varies from about .4 for card=1 to 4.0 for card=64. The breakeven point is card=7, with your approach faster at all higher numbers. Nice improvement!
Send me email if you'd like to pursue this further tomorrow. Also, you probably want to read dragontamer's post here closely and mentally consider whether his approach might make even more sense. I haven't thought about it yet, but if he thinks he knows the best way, he may well be right. Good night!
Edit: Just noticed that my numbers seem slightly faster than the ones Wojciech reported for large cardinalities. My guess would be that this is because I added "-march=native" to both compilations, although maybe there's something more subtle happening too. He's in Europe, so probably asleep now, but will likely be by to offer his thoughts in a few hours.