The oclvanitygen I'm currently using checks both the uncompressed and the positive and negative compressed versions of each key it computes.  Since the majority of the processing time is getting the keys to check this results in about 3x the addresses checked for the same amount of processing power.  At some point a while back I posted the .cl file I was using in the vanitygen thread but I never did completely modify oclvanitygen to do the rest of the calculations properly, I just run it in verbose mode with hardware error checking disabled and use a helper script to pull out the proper keys.

Those GPU numbers were measured at the original clockrate (725 core, 1000 mem) although usually I run it slightly underclocked for efficiency.  I generally point it at one of the vanity pools (vanitypool.thruhere.net which accepts compressed solutions or vanitypool.appspot.com which doesn't but sometimes has more work) when there is work that is worth more than bitcoin mining.