(L) [2016/06/07] [ost
by mpeterson] [super fast parallel sbvh construction] Wayback!some month ago we presented early results of our mbvh4 traversal and intersection
routines for manycore cpus. these implementations outperform latest embree kernels
by 4x for coherent ray transport and > 2x for incoherent transport.
the underlying accl. structure is a mbvh4 built with spatial splits. this kind of
space subdivition typically gives the best traversal performance for 99% of all scenes
but a parallel, vectorized implementation is not that easy to do. some years ago we
already developed a highly parallel morton grid builder that is today as fast as any gpu
grid builder when running on a up2date dual socket system ([LINK https://www.youtube.com/watch?v=qtXs4APw1uI]).
in production environments low quality builders have no standing and therefore
a fast and scalable high quality bvh builder is needed. below we present some early results from our
latest developments in this direction. on complex scenes we outperform embree builders by up to 8x !
the paper and more informations can be found here:
[LINK http://rapt.technology/posts/part-ii-parallel-bvh-construction/]
[IMG #1 Image]
[IMG #1]:Not scraped:
https://web.archive.org/web/20210509163022im_/https://picload.org/image/rgpoaado/sbvh_results.png
(L) [2016/06/07] [ost
by atlas] [super fast parallel sbvh construction] Wayback!This is interesting, you either find guys with really good theory of spatial subdivision, or really good computer science theory with knowledge about vector units, memory hierarchy, etc. When the two collide, good things happen! Great work
(L) [2016/06/11] [ost
by MohamedSakr] [super fast parallel sbvh construction] Wayback!is this the raytracing timing or the building timing?
as far as I can see, this scales with more cores, this means it will be a monster on GPU?
(L) [2016/06/11] [ost
by atlas] [super fast parallel sbvh construction] Wayback!Don't forget that GPUs have a hard time dealing with branching, and coordinating the memory movement is challenging. Some BVH algorithms like LBVH map really well to GPUs, but I have a feeling this algorithm would be more challenging.
(L) [2016/07/14] [ost
by mpeterson] [super fast parallel sbvh construction] Wayback!>> spectral wrote:Is there some patent on it ... or we are free to use it or even add it to embree ?
Which Embree kernel do you use for the comparison ?
no patent. you are free to use it as you like. paper describes everything in detail.
we compared against all of them and took the best performing one.
mp
