Research On Distributed-Memory Ray Tracing For Large-Scale Rendering

Recently,with the development of industrial,scientific and cultural fields,the demand for photo-realistic rendering increases rapidly,which puts forward a higher requirement for quality of ray tracing algorithm.That is,for any large-scale scenes and complex lighting conditions,the global illumination can be computed by ray tracing as fast as possible,even in commodity computers.However,the bigger the scene data get,the greater the amount of computation required for the same rendering quality,which typically exceeds the processing capacity of a single machine.Therefore,the research on the parallelization of ray tracing for general-purpose computing platform has important practical significance.This paper mainly studies the distributed-memory ray tracing for large-scale realistic rendering.The main work is as follows:First,with the analysis of the ray tracing algorithm flow,a new ray tracing task partitioning scheme based on scene data parallel was proposed,i.e.,each task processes only a part of scene data and a part of rays enter the same spatial region.The tracing tasks are scheduled to the machines where the scene data live rather than drags the data across the network,which avoids the bottleneck caused by communication overhead.Secondly,to select related rays for each scene split and reduce unnecessary intersection tests,a grid structure for region definition was designed.This spatial structure bounds the scene split closely and schedules the rays which have potential intersections with the split.By using the methods of heuristic subdivision,overlap reorganizing,path splitting and reordering,the result of intersection test of each tracing task is valid without combining judgment,and the computation of each ray is able to contribute to the final image without waiting for the result returned from the recursive path tracing,thereby the efficiency and reliability of the algorithm in distributed-memory environment are improved.Finally,to make the parallel processing of large-scale scene scalable and achieve high available and reliable of data processing,Hadoop was used for management of distributed rendering cluster.The improved algorithm was implemented by MapReduce interface and CUDA framework.By comparison with the previous algorithms in multiple indicators,the algorithm of this paper reduces the rendering time by more than 12%,and shows better in scalability,load balancing and fault tolerance than others.