PTX seems quite reasonable to me, but I don't think it's really intended to be a programmer-facing abstraction. NVIDIA wants you to target CUDA, and they'll (try to) ensure that CUDA has sufficient performance advantages to make it worth your while to do so.