Seems like building a BVH in parallel needs to first sort the objects in a certain order, and each thread is responsible for some ranges of the objects.

http://devblogs.nvidia.com/parallelforall/thinking-parallel-part-iii-tree-construction-gpu/