SBVH has been described in this paper: http://www.nvidia.com/object/nvidia_res ... b_012.html. SBVH is an algorithm involving a different way to build accelerator structures so it is applicable to any BVH (including our SSE hand optimized QBVH). In fact, SQBVH and QBVH share the same traversal core.
Where QBVH uses a single strategy (i.e. object split) to build the BVH, SQBVH uses 2 (object split and space split) and pick the best one. Spatial split is extremely superior in handling triangles with large axis aligned bounding boxes. The result is that SQBVH is as fast as QBVH in the worst case but can be up to 2 time faster or more. The draw back is that the building process is slower and the end result can require slight more memory (i.e. up to a 30%).
This is an example, it is the Sponza model rotate by 45 degree (to not be axis aligned) rendered with QBVH (295 kC/s):
and rendered with SQBVH (420 kC/s):
+42% speed up is even too good to be true
Keep in mind, this is a best case scenario (and there are sometime subtle bugs that boost performances).Everything is highly untested and has required far more time than I expected to be implemented (some aspect is very tricky). SQBVH can be used like all other LuxRender accelerator with the name "sqbvh".
