Early evaluation of directive-based GPU programming models for productive exascale computing