Hybrid photon map on a multi-core CPU

Abstract

Photon mapping is one of the most popular global illumination algorithms. The main operations of photon mapping consist of ray tracing and the k-nearest neighbor(KNN) search. Among the two operations, more research has been investigated to improve the performance ray tracing. As a result, the bottleneck of photon mapping becomes the KNN search. The KNN search in photon mapping is used to estimate the radiance for each ray. In addition to reducing the cost of the KNN search, reducing the construction time of photon maps gets more important because dynamic scenes and lights are widely used in various applications such as games and computer-generated movies. This thesis presents a hybrid data structure that combines a hash-grid and kd-trees to reduce both the construction time and the runtime KNN search time. We compute a uniform grid on the input scene. Then, we construct a photon kd-tree for each cell of the grid. For the KNN search at runtime, we can remove photons that do not contribute to the radiance estimate using the hash-grid. More specifically, we only consider gird cells that are in close proximity to the grid cell that contains the query point of the KNN search. Then, kd-trees in these grid cells are traversed for finding the exact k-nearest-neighbor photons. We compare our representation against the original kd-tree representation that is constructed from all the photons. By using our hybrid representation, we can reduce the traversal time and thus improve the performance of the KNN search by up to 2 times. Also, due to the simplicity of our representation, we reduce the construction time by up to 7 times. Finally, we have found that our method shows a higher scalability on a multi-core CPU, since we access small kd-trees of gird cell instead of accessing the big single kd-tree during the KNN search.