SIMD Packet Techniques for Photon Mapping

We present a novel photon mapping framework that uses single instruction, multiple data (SIMD) parallelism to accelerate the final gathering phase of photon mapping. By using SIMD instructions, four coherent tasks can be computed in parallel using almost the same memory traffic as it would cost to process one task alone. This approach has been very successful for real-time ray tracing, but until now it has been unclear how to effectively apply the same approach to final gathering. Our solution is to use sample-point density estimation instead of k-nearest neighbor density estimation, a technique drawn from reverse photon mapping. Sample-point estimation removes the overheads that make SIMD instructions impractical, while retaining the same benefits and image quality as traditional photon mapping. Additionally, an important question arises whether it is better to use forward or reverse photon mapping. In an interactive context, classical asymptotic algorithmic analysis is not enough to compare the two algorithms. We provide a novel asymptotic bandwidth analysis, which addresses more issues found in practice. The analysis motivates the use of forward photon mapping when using SIMD parallelism as well as partial reordering for improved scalability. The resulting framework can achieve interactive rates for photon mapping at low resolutions, including the time it takes to trace photons and build the photon map.