[llvm] [NVPTX] Select bfloat16 add/mul/sub as fma on SM80 (PR #121065)

[Artem Belevich via llvm-commits]

Artem-B wrote:

> Why would PTX >= 7.8 mean that sm_80 doesn't benefit from this patch?

Ugh. Indeed. You're absolutely right. sm_80 still does not have mul/add, even in newer PTX versions. I was concentrating on sm_90 and somehow ignored that sm_80 is still affected. 

OK, the patch is still useful. Sorry about the noise.

https://github.com/llvm/llvm-project/pull/121065