A Parallel Path Tracing Algorithm Based On Distributed Memory Of Large-Scale Scenes

The digital creative industry is an emerging economic form at the time when the modern information technology and cultural creativity and design services are gradually integrated.It is listed as one of the five strategic emerging industries that our country has focused on.Image rendering is one of the key supporting technologies in the digital creative industry and it is the most time-consuming part.Because of the need to deal with complex lighting models in realistic rendering,rendering a picture can take hours,and rendering an animation can take up to 10 million hours.As the industry develops,the complexity of the scenes used for rendering is getting higher and higher,and the fineness of rendering requirements is also increasing.How to make full use of existing digital simulation results and computer computing power to get rendering results better and faster has become an urgent need in the field of highly realistic renderingSupercomputers are computers that have far more computing power than personal computers.Supercomputers have been widely used in research,defense,and commercial fields that require a lot of computing.The use of supercomputers to accelerate the parallelization of rendering applications is a research direction.Because rendering computing has the characteristics of high computational load and large amount of data,rendering research on supercomputers also has great challenges.The research and development of this paper is carried out on the Sunway TaihuLight Supercomputer.The requirements for scene modeling accuracy and scene size,as well as the highly realistic rendering results and grand visual effects are also increasing.The path tracing algorithm needs to load all the scene data into the memory to detect the intersection of the light and the objects in the scene.The increasing model complexity poses a challenge to realize an efficient and scalable path tracing algorithm.To solve this problem,path tracing algorithm based on internal and external memory scheduling and path tracing algorithm based on distributed memory have been proposed.Among them,the internal and external storage scheduling loads the local data used for drawing from the external memory into the memory,and realizes the rendering of the complex model through the internal and external data exchange,which has the bottleneck of the data locality and the storage layer transmission bandwidth,and the expansion is insufficient;The distributed memory path tracing method realizes the drawing of large-scale scenes by distributing the scene data to different node memories in a distributed manner.Both of these methods aim to improve the utilization of computing resources while improving the localization of data,thereby improving parallel efficiency.At present,how to reduce the data acquisition overhead while ensuring the load balancing of the light computing task is still one of the difficult problems.Based on the high-performance computer Sunway TaihuLight,this paper proposes a large-scale scene parallel path tracing method based on distributed memory system.The efficient scene segmentation method is used to distribute the geometric data of the scene to multiple nodes for storage and this paper implements parallel lighting calculations.based on distributed storage.This paper proposes four scene partitioning and distributed storage methods,including scene randomization partitioning and storage method,scene geometry object size based scene balancing partitioning and storage method,Morton code based scene decentralized partitioning and storage method,Morton code based scene centralized balancing partitioning and storage method,and compares path tracing parallel efficiency for four distributed storage method.In addition,this paper uses non-blocking communication and double buffering strategy to optimize the communication between each node,which improves the weak scalability of the algorithm while ensuring the strong scalability of the path tracing algorithm.The experimental results show that the proposed method can effectively render ultra-large-scale scenes exceeding 100GB.Under 300,000 cores,the parallel efficiency of path tracing calculation is more than 50%compared with 50,000 cores.Among the four partition storage methods,Morton code based scene centralized balancing partitioning and storage method has the least time for path tracing,and the scene geometry object size based scene balancing partitioning and storage method takes the least time to read files and build KD trees.The method of this paper has good scalability in both the scene data scale and the ray computing task.The innovation of this paper is mainly to combine Morton coding with geometric model partitioning,and propose a method of scene geometry centralized partitioning based on Morton code,which improves the quality of the constructed KD tree and reduces the time of intersection calculation in path tracing.