Closed aprokop closed 4 years ago
We should wait to merge this until ECP-CoPA/Cabana#308 is merged. Then we can either wait and implement maxNeighbor
in Cabana, or merge this for now and add the automatic increase later.
The 10% is just to try to avoid counting for small fluctuations in max neighbor count - it should be checked and could be improved
I think the main challenge with ArborX is a lack of a simple way to determine whether buffer optimization succeeded, i.e. did not go through the second pass. The only (I think) current way of determining that is to do the extra cost computation of explicitly computing max using the offset
array. While comparatively not expensive, it is still an extra cost we would unnecessarily incur, and moves us further away from the Verlet baseline.
So far, just the basic implementation. Things that are to be checked: