jhuber6 wrote: > I did some experiment with RTX2060 wihch supports ITS. It seems > `__ballot_sync` does mask the result by the lame_mask argument so the > out-of-mask lanes won't affect result even if they reach the function and > vote. My concern is that if HIP does not mask the result, users may get > different behavior and get confused.
That's interesting, I have an `sm_89` locally and I did not observe this behavior. In fact I found if you intentionally restricted it you'd get warp illegal instructions. I don't know enough details about how this actually works unfortunately, but I didn't see any forced sub-masking. https://github.com/llvm/llvm-project/pull/176202 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
