It seems there is possibility to port CUDA-based NV apps to AMD hardware w/o re-writing them.

http://gpuopen.com/tag/hip/

Also it seems it's mostly Linux's part of compute world.
Anyone had some experience with all this stuff?

Looked at some similar stuff a little for options (though a different route), though I'd consider going via OpenCL the better choice as you do already, at least for now.

They're a good idea IMO, but always seem to fall a little short: In that way I look at these high level ecosystems as more framework attempts like OpenMP, Intel TBB, OpenACC and things like that meant to ease porting where the designs weren't made to be cross-platform/device-independent in the first place.

So it's what we need, but more domain specific to our particular style of problems, where something in how they try to be all generic and everyone's solution, can make them "jack of all trades, master of none"

The way the architectures and our own apps seem to be pointing, including changing DX12 and Vulkan underneath, is toward domain specific languages (DSLs) and plugin architectures.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.