GPU-based Out-of-Core HLBVH Construction

Abstract

Recently the GPU has been used extensively in building indexing structures for moderately complex scenes that fit inside the GPU core. However, only few methods have been developed for constructing indexing structures for massive models larger than GPU memory. In this paper, we present an out-of-core HLBVH algorithm, a new method for constructing spatial hierarchies suitable for massive models that cannot fit into GPU device memory. A key insight of our method is how to bring and process out-of-core data blocks that do not fit into available device memory. Results show that our approach can compete with HLBVH hierarchy builder for large models on CPU. We also demonstrate the value of our algorithms in a GPU-based out-of-core path tracer that brings tree nodes and geometry into GPU core as needed, and efficiently achieve complex global illumination effects for models up to hundred million triangles.