Isn't cuDNN a much better case for reimplementing than CUDA? It has much more choice in how things actually happen and cuDNN itself chooses different implementations at runtime + does fusing. It seems way more generic and the reimplementation would allow using the best AMD-targeted kernel rather than one the original has.
And that thing is left for unreleased on windows for almost a whole year for unknown reason. Even though there is activity on github and build fix frequently. There is just no .exe or .msi for you to download. In fact, the rocm for linux is on major 6 release (which includes miopen). But somehow windows is still on major 5 (don't have miopen) for almost a whole year.
It almost make me wonder. Is there a shady trade somewhere to ask amd never release sdk for Windows to hike the price of nvidia card higher? Why they keep developing these without release it at all?
Since they cancelled the work on zluda and absolutely fail to do anything about other options, I really believe there's some "don't do it or you'll get sued to hell and back" agreement. They can't be so dumb they just miss it by accident.
