[BACKEND] Deprecate MMAv1; fallback to FMA for computeCapability < 80

[Jokeren]

No description provided.

[ThomasRaoux]

Can you update the readme?

[Jokeren]

Updated

[Isotr0py]

Hi @Jokeren @ThomasRaoux, I noticed that the supported Nvidia GPU in README was bumped to requiring cc 8.0+ in this PR.

Did it mean the Volta and Turing support including FMA fallback will be removed in the future? Or did it just mean MMA will be disabled on pre-Ampere GPUs with FMA fallback?

I used T4 to run the triton core tests and it passed with latest triton 0.3.2, but I'm not sure if it just still worked by coincidence since it's no longer in the supported hardware list.

[Jokeren]

Did it mean the Volta and Turing support including FMA fallback will be removed in the future? Or did it just mean MMA will be disabled on pre-Ampere GPUs with FMA fallback?

IMO, community contribution to pre ampere architectures with MMA is welcome if it doesn't break backward compatibility. The FMA code path will still be maintained in triton/main.

[Isotr0py]

@Jokeren Got it! Thanks for your explanation!

[ThomasRaoux]

yes we are not going to break support for older platform but the support will be best effort and won't be tested on our side. As Keren mentioned contributions are welcome to fix older platform however we are unlikely to accept complex performance related patches for those.