Biography
Enrollment Date: 2012
Graduation Date：2015
Degree：M.S.
Defense Date：2015.06.02
Advisors：Yangdong Deng
Department：Institute of Microelectronics，Tsinghua University
Title of Dissertation/Thesis：GPU Based Tiled Ray Tracing Algorithm
Abstract:
Ray tracing is one of the most import topic in academic of computer graphics. By simulating the behavior of light in the scene, ray tracing can provide excellent rendering results with photo realistic. Now ray tracing technique is used widely in film industry, computer aided design, visualization, etc. However, ray tracing algorithm requests massive computation, thus can run on the computer cluster for off-line rendering. With the improvement of hardware performance and the increasing demand for better rendering results from the customer, it is common interest to realize ray tracing on consumer hardware such as PC, tablet computer and mobile phone, both shared by the academia and industry.Nowadays, graphics process unit(GPU) is commonly used as the platform running ray tracing programs. However, the architecture of today’s GPU is designed for the rasterization render pipeline, not quit suitable for the ray tracing algorithm. Ray tracing tends to use hierarchical data structure such as BVH, KD tree and this leads to problems. Frostily, access pattern of these data structures are irregular and inconsistent as the GPU bandwidth is optimized for sequential access. Secondly, there are too many bifurcation of control and jumps in the traversal of the trees. This goes against the single-instruction, multiple-data execution mode on GPU. Finally, the random access to memory by massive threads in GPU will reduce the cache efficiency greatly, especially the threads are processing rays with great randomness such diffuse rays.Facing the problems above, this work proposes a tiled ray tracing algorithm to obtain the coherence of memory access pattern systematically. This algorithm can divide the ray tracing data structure into small tiles, which can be loaded into the GPU on-chip memory, as well as choose the rays with consistency to traverse the tiles accordingly. Our work can improve the memory access pattern of ray tracing and reduce the algorithm’s dependency on bandwidth. The experiment results reveals that this algorithm can speed up the primary ray traversable by 20%, in the mean time, the secondary ray traversal is accelerated significantly to the comparable speed of primary ray traversal.