Online Autotuning for Interactive Raytracing

点击翻译按钮获取中文摘要

This project aims to combine and synergistically develop two separate areas, raytracing and autotuning. Raytracing is an established and widely used technique for generating photorealistic images.The combination of raytracing and autotuning holds significant promise, because image synthesis is performance critical, but offers a large range of tuning opportunities, for example the choice of acceleration structures and heuristics, the partitioning of work among CPU and GPU, the number of threads employed for individual tasks, or the use of ray sorting to achieve coherent accesses. Autotuning can speed up raytracing significantly by finding the optimal operating point among the numerous tuning parameters available.However, the optimal operating point may shift over time, for example by changes in viewpoint, light source, geometry, or material properties. Thus, a performant raytracer must be re-optimized continually.The aim of the project is twofold. First, we wish to develop new capabilites for autotuning, in particular the capability of dynamic, or online, autotuning. For this, we plan to explore suitable search procedures, machine learning techniques, and dynamic modeling, and evaluate these approaches on interactive raytracing-based methods. Second, these methods will be extended to be tuneable to current conditions. For this, they must be made adaptable und deliver and aggregate important information about their runtime behavior.We expect results to include not only adaptive image synthesis techniques that perform well over a large range of scene variations and user interactions, but also new software architectures, for instance autotuning frameworks, libraries, and machine learning procedures that drastically simplify the programming tasks involved in achieving high performance, applicable beyond raytracing.