So you just have to find somebody who can modificate your client or the oclvanitygen... Am I right?
Just the oclvanitygen. The Go-generator is probably already 90% of what can be done with it (I guess).
"My client" is the Perl stuff that holds all this together and there are no (not yet) time critical pieces of code in there.
Very probably the hash check (hash/map/asociative array of funds) would also have to be moved in the GPU as the CPU wouldn't catch up checking the generated hash160s.
Maybe similar to brainflayer with bloom filters, and only in case of a hit, the CPU would check it. If the rate of false positives would be low enough, the CPU should have no problems catching up.
From what I read in the vanitygen/oclvanitygen docs, it seems the code is already 99% there, maybe even 99,5% as it does the checking and also does sequential PK increments for speed. I would only need a means to be able to steer the offset where these sequential PK increments start from.