ssahasra wrote:

>     1. Existing users already have to manually call waits on tensorcnt, so 
> making this intrinsic participate in the asyncmark scheme won't break their 
> code and

Ok, now some things clicked in place about this confusing exchange of emails. I 
have been assuming that tensor operations are also covered automatically. If 
they are not, then there is no compatibility issue with "intrinsics already out 
there".

> 2. The hardware isn't out yet - now is the time where a somewhat breaking 
> change like this is least unreasonable

Well, that depends on what "out" means. We do have the chance to rename the 
existing intrinsic with `.async`, and possibly auto-upgrade old occurrences. 
Renaming the equivalent Clang builtin will break source code that exists "out 
there somewhere". Eventually, this builtin will get listed [here](
https://llvm.org/docs/AMDGPUAsyncOperations.html#memory-accesses), but we want 
to reach a point where we can simply say "all `.async` intrinsics".




https://github.com/llvm/llvm-project/pull/200775
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to