Remember that even though these operations are atomic, the overhead of using things that have critical sections in CUDA is high, and usually will cause a large slowdown