Real-time global illumination on distributed systems

Abstract

Realistic computer-generated visuals captivate users and provide them with immersive experiences. Synthesising these images in real time requires significant computational power and is out of reach for a wide range of commodity hardware, particularly for mobile devices. Remote rendering solves this problem by computing frames on the Cloud and streaming the results to the client device as video. This solution provides good image quality but introduces latency, which may make applications appear unresponsive, degrading user experience. It may also require significant bandwidth. This work investigates an alternative distributed rendering strategy, where the computational power of the local device is not discarded but instead used to eliminate or reduce latency. Three methods are presented, all using a client-server architecture that splits the rendering pipeline between a powerful remote endpoint and a weaker local device. The first makes use of sparse irradiance sampling on a voxelised representation of the scene. It supports multiple clients and is highly configurable, allowing the use of different interpolation schemes according to the capability of the device; image quality can be reduced to lower reconstruction cost. The second stores radiance in a megatexture and communicates it to the client device, where rendering is performed at a low cost by sampling the megatexture. A coarse megatexture that fits into GPU memory is maintained on the client device and used to provide temporary low-quality output until high-quality server data are received. The third uses the double warping image-based rendering technique to produce novel views from two reference views. The client device also receives irradiance data which it caches in a coarse megatexture. The cache is used in a fallback mechanism that mitigates visual artefacts due to holes in the data. The results show that input lag can be eliminated and bandwidth requirements can be kept low, while retaining a measure of fault tolerance and decent image quality. It is envisaged that distributed methods similar to the ones proposed will gain more traction as the computational capability of commodity devices increases and they can be assigned larger workloads and use more sophisticated algorithms.