Christian Trott
Christian Trott
Not clear that OpenMP is fixable. Note: the OpenMP runtime has its own semantics, and Legion would need to use only OpenMP threads to make it possible to have concurrent...
I see. I think that might work already since the 4.0 release. Essentially we made multiple OpenMP instances work for cases where you enable nested OpenMP. Something along these lines...
Oh I should add, I don't know exactly whether its a requirement for the nested stuff to see that it is proper standards compliant "nested OpenMP" my guess is it...
I don't think we use the work_range stuff for OpenMP at all anymore or? Should it hit the native OpenMP parallel for clause?
Also what is the performance impact you have seen from this not being theoretically optimal? I mean some threads would likely still work as much as before. Essentially you would...
Do you have a simple reproducer which can expose the issue say on a typical Dual Socket machine. We may want to look at this a little bit - I...
I bet the cray compiler things the combo of "-fopenmp" and "--offload-arch" implies OpenMP Target on, and that conflicts with building HIP.
I updated my XCode now. Gonna close the issue, nothing we can do about this really - or at least its not worth it.
Christian needs to do unit test.
No it is kinda weird that it does. I can somewhat understand why since the unordered map grows somewhat dynamically internally, but the idea is to make it behave more...