[Bug target/105873] [amdgcn][OpenMP] task reductions fail with "team master not responding; slave thread aborting"

ams at gcc dot gnu.org gcc-bugzilla@gcc.gnu.org
Wed Jun 8 09:09:15 GMT 2022


--- Comment #4 from Andrew Stubbs <ams at gcc dot gnu.org> ---
I think unused threads should be given a no-op function to run, not a null
pointer. The GCN implementation cannot tell the difference between a null
pointer and an unset pointer (which is what happens when the master thread

There's also a potential issue of what happens when barriers occur within an
active thread when there are also inactive threads. GCN barrier instructions
are unconditional, meaning that all the live threads must respond. The inactive
threads can do so, in a harmless way, as long as they allowed to spin, but we
don't want them spinning forever when the master dies.

I believe the current barrier implementation skips the barrier instruction when
the team's thread count is 1. This is how we avoid issues with nested teams and
tasks. I don't know why that doesn't help here?

More information about the Gcc-bugs mailing list