If anything AMD should follow NVIDIA in supporting OpenACC which is the way to go as far as true probability goes.
It doesn’t matter that the rocm libs are optimised for AMD.
In all effectiveness ROCm is AMD's version of CUDA it includes it's own run time and driver stack which effectively replaced the OpenCL frontend for AMD GPUs, basically it gives AMD the ability to do with their GPUs what CUDA allowed NVIDIA to do and it's not to be dependent on a camel (OpenCL).
This is what ROCm is: https://rocm.github.io/images/ROCm_Stack.png or at least this is what AMD is pushing it to be, currently a lot of things are still unfinished or non-existent.
The representation of it being open is misleading, while more of the source code is available than for CUDA, ROCm isn't vendor agnostic on any tangible level. In 2-3 years once it's somewhat in a production ready state it will lock you into a single vendor just like CUDA does.