Hybrid Voxel Formats for Efficient Ray Tracing

TL;DR

It is shown that a hierarchical combination of voxel formats can achieve Pareto optimal trade-offs between memory consumption and rendering speed and achieve a new Pareto frontier in ray intersection performance and storage cost.

Abstract

Voxels are a geometric representation used for rendering volumes, multi-resolution models, and indirect lighting effects. Since the memory consumption of uncompressed voxel volumes scales cubically with resolution, past works have introduced data structures for exploiting spatial sparsity and homogeneity to compress volumes and accelerate ray tracing. However, these works don't systematically evaluate the trade-off between compression and ray intersection performance for a variety of storage formats. We show that a hierarchical combination of voxel formats can achieve Pareto optimal trade-offs between memory consumption and rendering speed. We present a formulation of "hybrid" voxel formats, where each level of a hierarchical format can have a different structure. For evaluation, we implement a metaprogramming system to automatically generate construction and ray intersection code for arbitrary hybrid formats. We also identify transformations on these formats that can improve compression and rendering performance. We evaluate this system with several models and hybrid formats, demonstrating that compared to standalone base formats, hybrid formats achieve a new Pareto frontier in ray intersection performance and storage cost.

Figures (10)

• Figure 1: Pseudo-C++ types describing the memory layout of base formats.
• Figure 2: An example usage of the metaprogramming system to compile the R(4, 4, 4) G(8) format with whole level de-duplication.
• Figure 3: Pseudo-code for generated level construction and intersection functions. Functions, methods, and fields abstract format specific behavior.
• Figure 4: A depiction of the Morton order of a $4 \times 4$ grid. The upper left $2 \times 2$ sub-grid (light gray) is traversed before the grid cells in the gray region are visited.
• Figure 5: Screenshots of the tested models from the camera positions used in experiments. The experiments are run without shading from bounces.