Massively parallel implementation of algorithms for computer graphics

Abstract

Computer graphics, since its inception in the 1960s, has made great progress. It has become part of everyday life. We can see it all around us, from smartwatches and smartphones, where graphic accelerators are already part of the chips and can render not only interactive menus but also demanding graphic applications, to laptops and personal computers as well as to high-performance visualization servers and supercomputers that can display demanding simulations in real time. In this dissertation we focus on one of the most computationally demanding area of computer graphics and that is the computation of global illumination. One of the most widely used methods for simulating global illumination is the path tracing method. Using this method, we can visualize, for example, scientific or medical data. The path tracing method can be accelerated using multiple graphical accelerators, which we will focus on in this work. We will present a solution for path tracing of massive scenes on multiple GPUs. Our approach analyzes the memory access pattern of the path tracer and defines how the scene data should be distributed across up to 16 GPUs with minimal performance impact. The key concept is that the parts of the scene that have the highest number of memory accesses are replicated across all GPUs. We present two methods for maximizing the performance of path tracing when dealing with partially distributed scene data. Both methods operate at the memory management level, and therefore the path tracing data structures do not need to be redesigned. We implemented this new out-of-core mechanism in the open-source Blender Cycles path tracer, which we also extended with technologies that support running on supercomputers and can take advantage of all accelerators allocated on multiple nodes. In this work, we also introduce a new service that uses our extended version of the Blender Cycles renderer to simplify sending and running jobs directly from Blender.