My question is, is it feasible for AMD to build an ahead of time compiler that transparently translates CUDA instructions into whatever AMD could have, so thing

My question is, is it feasible for AMD to build an ahead of time compiler that transparently translates CUDA instructions into whatever AMD could have, so things Just Work™? They'd be heavily incentivized to do so. Or even put hardware CUDA translators directly into the cards? Or am I misunderstanding CUDA? I think of it as something like OpenGL/DirectX.