Support for explicit matmul instructions / tensor core instructions

[thomasdullien]

Hey all,

if the hope that WebGPU can possibly provide a vendor-agnostic replacement for things like CUDA over the medium run should play out, WebGPU needs to support instructions for Tensor cores / Matrix cores.

This case was inspired by reading the comments under https://news.ycombinator.com/item?id=35800988

This case depends on #3950 as far as I understand.

Happy to flesh out the issue in more detail (as I learn more about the subject).

[alan-baker]

We're definitely interested interested in exposing ML primitives as they become standardized more widely.

[dneto0]

Thank you for raising this.
I'd like to redirect discussion to #4195

Agree this is dependent on portability of reconvergence, which is being worked out in subgroup discussions. #4306.
Therefore I'm moving the milestone to 2, to match subgroups.

[dneto0]

Agreed. Google thinks this is very attractive.
Agreed this depends on subgroup semantics. That's why I moved it to M2 at least.