high performance computing on graphics processing units: hgpu.org

Photon mapping is a popular extension to the classic ray tracing algorithm in the field of realistic image synthesis. Moreover, it benefits from the massive parallelism computational power brought by recent developments in graphics processor hardware and programming models. However rendering the scenes with dynamic lights still greatly limits the performance due to the re-construction at each rendered frame of a kd-tree for the photons. We developed a novel approach based on the idea that storing the photons data along with the kd-tree leaf nodes data and implemented new incremental update scheme to improve the performance for dynamic lighting. The implementation is GPU-based and fully parallelized. A series of benchmarks with the prevalent existing GPU photon mapping technique is carried out to evaluate our approach. Our new technique is shown to be faster when handling scenes with dynamic lights than the existing technique while having the same image quality.