Research On Key Technologies Of Parallel Optimization Based On Physics For Ray Tracing Algorithms

Physically based ray tracing is the process of producing a photo-realistic image from the description of a 3D scene,which uses principles of physics to model the interaction of light and matter.Its primary goal is to attempt to simulate reality,and is widely available for fields of entertainment(movies,special-effects industry and animations)and high-end design.In contrast to interactive or real-time ray tracing,it involves complex physical processes and mathematical model.Meanwhile,random length of light sample and irregular memory access pattern lead to low scalability and severe load imbalance.And the complexity of the scene and the high computational cost of the dense sampling light thrown into the scene bring great challenges to the design a set of multi-node distributed parallel imaging algorithm.In addition,most of existing works are running on single processor that usually need several hours or even days to completed rendering,which are far from the requirement of people.In this paper,by means of analysis the visibility computing of 3D scene,physical essence of bidirectional path tracing of light transport equation,we propose the corresponding methods with multi-nodes parallelism to fast rendering.The main works are as follows:(1)As a technology to accelerate the speed of rendering,visibility computing of 3D scene got great attention in recent years.However,traditional visibility computing method is based on view field of cone,which often loss some visible primitives.In addition,the calculation of intersection between view of cone and primitives is often very complicated,which further increases the computational burden.Therefore,we propose a new method called hemispheric visibility computing to overcome deficiencies in traditional method in this paper.In addition,we provide the distributed parallel framework based on Master-Worker paradigm for hemispheric visibility computing of 3D scene.Then,we also analyze the key factors such as the scale of scene,assign granularity and so on that affects the performance in detail.Finally,we select a complex scene to test on Tianhe-2 supercomputer and got the linearly speedup with the scale of ten-thousand cores.(2)Compared with path tracking,bidirectional path tracking algorithm overcomes the shortcomings of image noise when the light source is small,but the computational complexity is also increased by 3-5 times due to the random combination of viewpoint path and light source path to generate new light paths.By analyzing the physical essence of the propagation equation of light and bidirectional path tracing algorithm,the concepts of final image and sub-image are proposed,which provide a theoretical basis for large-scale distributed parallel rendering.We propose a set of imaging algorithms based on bidirectional path tracing of light transport equation,which divides the tasks of inter compute node and intra compute node in detail.It achieves load balance through dynamic and static combination of task scheduling.Typical yet challenging 3D scenes are used to test.The experimental results on Tianhe-2 Supercomputer show that the algorithms achieve nearly linear speedups on ten-thousand cores scale.