Description
Since you guys already compute the BVHs on the GPU, what is the barrier to radeon rays creating meshes from GPU-sourced data?
It is a little bit bothersome that for initially GPU resident data would have to be pulled down to the CPU and re-uploaded to the GPU with the current API (or mapped, but there is often little mappable memory, its slower and it will still rely on the driver+CPU or DMA engine to move it across).
The fact that the createMeshes call is blocking is also bothersome, could we get a non-blocking version?
Finally, such an approach could save memory as a duplicate set of vertex positions would not need to be kept between i.e. OpenGL/Vulkan and Radeon Rays.
Note: I understand that unless your BHV approach sorts the actual vertices according to the hierarchy instead of some index list.