If you expect OpenCL to match CUDA in performance, that's unlikely to happen any time soon. Although there are more optimisations to come.