RyanRio wrote:
> -1 on adding async variants
>
> These intrinsics, just like `global.{load,store}.async.{to,from}.lds.*` on
> gfx1250, should only use asyncmark / asyncwait and don't need a new "async"
> variant.
>
> The only reason that gfx9 has "sync" and "async" variants is because we tried
> to infer the memory accesses patterns for those DMAs on gfx9 before we
> realized that that was hard to do for software pipelined/high-performance
> workloads.
>
> Because the tensor DMAs already have their own separate counter, they
> naturally slot into the asyncmark/asyncwait system and will never have a
> "sync" variant.
>
> Therefore, this patch can be simplified significantly.
Yea agreed, but I'm thinking that these should actually be renamed to represent
that. So only have the async variant, in name and function.
https://github.com/llvm/llvm-project/pull/200775
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
