Mining effective parallelism from hidden coherence for GPU based path tracing

Abstract

As one of the essential global illumination algorithms, Monte Carlo path tracing has long been considered as a typical irregular problem that is less friendly for graphics hardware. To improve the efficiency of Monte Carlo path tracing, such techniques as ray reordering, ray compaction and wavefront formulation have been proposed to exploit the inherent coherence in processing different paths and materials for better SIMD efficiency on GPUs. In this paper, we develop a novel technique to extract extra parallelism in Monte Carlo path tracing applications by identifying hidden coherence. The basic idea is to perform a partial traversal in the fast on-chip memory of GPU and then identify coherent paths by analyzing the traversal results as well as other features of rays. Such coherence enables a higher level of parallelism that not only compensates the overhead of traversal, but also leads to improved performance. Experiments prove that our techniques deliver a traversal throughput higher than leading-edge results by up to 15%.