Open lanwatch opened 3 years ago
@lanwatch thank you for reporting this and apologies for the delay in picking it up. Quick question, do you happen to have a repro handy? I suspect that this manifests for a particular situation, involving multiple __syncthreads()
calls in the same kernel? Thanks.
Since HIP-CPU uses libco fiber mechanism to mimic GPU thread behaviour, the stack gets constantly re-written and that leads to ASan reporting false positives all over the place.
Seems like there is a way to annotate fibers that other coroutine libraries use, see this, this or this.
Maybe the solution would be to have different backends for other coroutine libraries? Eventually we would want to use the
std
version, if that makes sense.