rocWMMA support?

🚀 The feature, motivation and pitch

Given how the competition has powerful pytorch extensions based on hardware-accelerated matrix FMA like tensorrt, and given how the impending gfx11 GPU releases have wmma support (similar to the competition's FMA-based tensorcores), I believe that it'd be really nice if pytorch-rocm is able to support GFX11's wmma instruction set (i.e. rocWMMA) via an extension similar to tensorrt or just natively via pytorch itself. Please refer to https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/issues/1899 for the relevant issue I created for the tensorflow software stack counterpart.

Alternatives

Not really as I believe gfx 11 wmma support is the way to go moving forward :)

Additional context

N/A