I need to calculate convolution about a sparse 3d point clouds, so I need to split the space into different zones. However, as for the points at the border between zones, I need to use atomicadd to ensure they are calculated correctly. Is there any way to avoid the use of atomicadd? (can’t calculate the border points totally in one zone, because of the tensor cores shape requirement)
Perhaps make the zones overlapping and do some work twice?
Or keep them separated, but sum it up into different target memory, which you combine in a second step?