Project: Hardware acceleration of light transport algorithms
Description
Photon mapping is a physically-based light transport algorithm. It consists of two stages: in the first stage, energy packets called photons are emitted from light sources and traced through the scene. When the photons intersect with scene geometry, their position and energy are recorded and stored in a data structure called a photon map. In the second stage of the algorithm, rays are shot from the camera into the scene. At the intersection of the scene and the ray, the photon map is used to find all photons that lie in a certain radius around the intersection. Using these photons, the radiance is estimated, from which a color value is computed.
In 2018, Nvidia announced a line of graphics cards containing hardware acceleration for ray tracing, called RTX [1]. RTX graphics cards contain so-called RT cores, which perform bounding volume hierarchy (BVH) traversal and ray-triangle intersection testing in hardware. A BVH is a hierarchical representation of the geometry of a scene, allowing to speed up finding with which objects a ray intersects. Ray-triangle intersection testing is beneficial since objects are typically represented by a collection of triangles. These two components together speed up the process of ray tracing significantly.
While RT cores are originally only intended for ray tracing, Wald et al. show that the RT cores can also be used to accelerate point location for tetrahedron meshes [2]. Point location is a problem where, given a point, the goal is to find out on which geometric object this point lies.
The construction and querying of the photon map is a costly part of the photon mapping algorithm. Inspired by the work of Wald et al., the goal of this project is to investigate whether the RTX-accelerated BVH traversal can be leveraged to speed up the photon map queries. This would be done by storing the photons in an RTX-accelerated BVH and formulating the point query as a ray query. The performance of this hardware-accelerated photon map should be compared to existing data structures, and the strategy may also be applied to advanced photon-mapping techniques.
In the event of unforeseen complications that prevent such an implementation using RTX BHV traversal from succeeding, alternatives will be explored.
- [1] https://www.nvidia.com/en-us/geforce/news/geforce-rtx-20-series-turing-architecture-whitepaper/
- [2] https://www.willusher.io/publications/rtx-points
Details
- Student
-
NMNol Moonen
- Supervisor
-
Andrei Jalba
- Link
- Thesis