2024, Leria, E., Mäkitalo, M., Jääskeläinen, P., Sjöström, M. and Zhang, T., In the 30th ACM Symposium on Virtual Reality Software and Technology (VRST).
Abstract: Path tracing combined with multi-view displays enables progress towards achieving ultrarealistic virtual reality. However, multi-view displays based on light field technology impose a heavy workload for real-time graphics due to the large number of views to be rendered. In order to achieve low latency performance, computational effort can be reduced by path tracing only some views (source views), and synthesizing the remaining views (target views) through spatial reprojection, which reuses path traced pixels from source views to target views. Deciding the number of source views with respect to the computational resources is not trivial, since spatial reprojection introduces dependencies in the otherwise trivially parallel rendering pipeline and path tracing multiple source views increases the computation time. In this paper, we demonstrate how to reach near-perfect linear multi-GPU scalability through a coarse-grained distribution of the light field path tracing workload. Our multi-source method path traces a single source view per GPU, which helps decreasing the number of dependencies. Reducing dependencies reduces the overhead of image transfers and G-Buffer rasterization used for spatial reprojection. In a node of 4x RTX A6000 GPUs, given 4 source views, we reach a light field rendering frequency of 3-19 Hz, which corresponds to interactive rate. On four test scenes, we outperform state-of-the-art multi-GPU light field path tracing pipelines, achieving a speedup of 1.65x up to 4.63x for 1D light fields of dimension 100 x 1, each view having a resolution of 768 x 432, and 1.51x up to 3.39x for 2D stereo near-eye light fields of size 12 x 6 (left eye: 6 x 6 views and right eye: 6 x 6 views), 1024 x 1024 per view.