Improving cache efficiency through parallel ray batching.

Abstract

Ray tracing is a rendering technique with an incoherent memory access pattern. In order to increase coherence and benefit more from caches in modern processors the technique of ray batching was created. This involves pausing ray traversal and resuming it later, when more cache hits can be expected. Ray batching proved successful but the technique was thus far limited to one processor core. This research aims to extend the technique of ray batching to work on multiple processor cores, in order to fully leverage the power of modern CPUs. A technique is presented to perform ray batching across multiple processor cores, however no performance increase is achieved over the current state-of-the-art techniques. Despite not achieving better performance than the current standard, insights are gained in what challenges need to be overcome for future work to improve upon the current standard. This thesis exposes the impact of inter-core communication overhead and shows promising results for batching in future work when that overhead is eliminated.