It'll be interesting to see how things develop with HIP (http://gpuopen.com/compute-product/hip-convert-cuda-to-porta...)