shiltian wrote: Looking at this now, it doesn't seem the same as it was at the start. For the latest changes, I think @arsenm made it clear that waves-per-eu values should always give way to the values calculated from flat-workgroup-size if they differ, since flat-workgroup-size is ABI and waves-per-eu is just a hint. I'm also dealing with something similar https://github.com/llvm/llvm-project/pull/155246, so however this is handled will also affect what we do in the attributor.
On the other hand, our backend seems surprisingly sensitive to these values. Previous changes to how these attributes were handled in the attributor have already caused a lot of performance regressions in downstream testing and even in past ROCm releases, with some showing as much as a 20x drop. I think we need to be very careful here. https://github.com/llvm/llvm-project/pull/138284 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
