Typical graphics processing units (GPU’s) are optimized for massively parallelised algorithms implemented as shaders, such as the rendering pipeline, which act on sequential memory access, and leverage SIMD capabilities. Path Tracing is a computationally expensive rendering technique that renders scenes in a photorealistic and physically accurate way, by projecting light rays into the scene and using phenomena such as reflection and refraction to determine the color of each pixel. Due to the unpredictable divergence of the rays after the first cast, there is an inherent amount of divergence in both memory access and instruction in determining what each ray collides with which manifests as a being a very expensive problem to solve for. Recent advancements in modern hardware have made certain types of realtime raytracing possible in games, namely the RT cores offered by Nvidia. While this is a major advancement in the field of realtime raytracing, we believe that the traversal of certain scenes and collections of entities in offline path tracers where such tree reconstruction is seldom, could be made more efficient through further development of certain acceleration structures, whilst leveraging specialized hardware. This project will investigate different methods for eliminating potential intersection candidates, both in their construction and traversal, to determine the feasibility of using specialized hardware to further boost performance of offline path tracers.
Undergraduate
An outcome of this project will be a fair comparison between path tracing on consumer hardware and a hardware implementation of an acceleration structure.
None
Parallel Computing (405.956, Lab)