Windows Vista, FireStream(RV770).
I am calling my kernel with work_group_size set to NULL, which is supposed to allow the compiler to pick a "good" work group size. My kernel runs fine as long as the global size is below 256 x 64. Once it reaches this size, the results are completely bogus. (It actually fails just a bit below this size, 248 x 62 works, though). I then tried calling with work_group_size set to (4, 16) and everything works fine.
Should NULL have worked? That's how I read the standard.