[PATCH] D157437: AMDGPU: Expand remaining system atomic operations

[Yaxun Liu via Phabricator via llvm-commits]

yaxunl added a comment.

In D157437#4644367 <https://reviews.llvm.org/D157437#4644367>, @arsenm wrote:

> In D157437#4644366 <https://reviews.llvm.org/D157437#4644366>, @arsenm wrote:
>
>> In D157437#4644295 <https://reviews.llvm.org/D157437#4644295>, @yaxunl wrote:
>>
>>> In D157437#4614101 <https://reviews.llvm.org/D157437#4614101>, @arsenm wrote:
>>>
>>>> ping, I think we just need to go with this and add in the new thing later. Also would like clarification on whether add/xchg at system scope are supposed to work
>>>
>>> This will cause perf regression for XGMI users. We cannot afford that.
>>
>> Nobody complained when the last batch were fixed
>
> Also correctness trumps speed. We should establish a working baseline

If the XGMI users complain, do we have an existing approach to recover the performance by modifying their code?

How about making the default always work but adding an option -amdgpu-unsafe-integer-ops to allow integer atomic rmw ops? Then at least they have a way to recover the performance.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D157437/new/

https://reviews.llvm.org/D157437