Fix loadStateField evaluator to address #1092 #1096
Conversation
Apparently linear access of Kokkos dynamic rank views is no longer working
| Kokkos::RangePolicy<ExecutionSpace>(0,data.size()), | ||
| KOKKOS_CLASS_LAMBDA(const int i) { | ||
| dataVec[i] = (i < stateToLoad_size) ? stateData(i) : 0.0; | ||
| Kokkos::MDRangePolicy<ExecutionSpace, Kokkos::Rank<3>>({0,0,0},{stateData.extent(0),stateData.extent(1),stateData.extent(2)}), |
There was a problem hiding this comment.
This works cause we use DynRankView, so usince 3 indices is always ok (we just have extent(1) and extent(2) that may be 1), right?
There was a problem hiding this comment.
Thanks for pointing this out. That's correct, however this code would fail in debug mode. To avoid the debug check, we have to use access() instead of operator(). Fix coming up soon.
There was a problem hiding this comment.
kind of weird that this works for rank<3 but nice that it does since it seems a lot simpler.
|
An alternative would be to deep_copy the state view into the MDfiled view, but this is a bit more fragile and it expects the views to have same properties (e.g. Kokkos layouts). |
I can confirm this doesn't work, I tried this awhile back before using an MDField iterator. |
| Kokkos::RangePolicy<ExecutionSpace>(0,data.size()), | ||
| KOKKOS_CLASS_LAMBDA(const int i) { | ||
| dataVec[i] = (i < stateToLoad_size) ? stateData(i) : 0.0; | ||
| Kokkos::MDRangePolicy<ExecutionSpace, Kokkos::Rank<3>>({0,0,0},{stateData.extent(0),stateData.extent(1),stateData.extent(2)}), |
There was a problem hiding this comment.
kind of weird that this works for rank<3 but nice that it does since it seems a lot simpler.
The DynRankView that is causing problems is the one in stateData, I'm pretty sure. MDFieldVectorRight is designed to be accessed linearly. |
|
Yes, I agree, stateData was the issue. |
Apparently linear access of Kokkos dynamic rank views is no longer working.
Not sure if this is a bug (DynRankView is not as widely tested as View) or expected behavior.
Fixes #1092