面向大规模场景的多片元效果高效绘制  

Efficient Rendering of Multi-Fragment Effects for Massive Scene

作　　者：周果[1,2,3] 朱登明 王兆其[1,2] ZHOU Guo;ZHU Deng-Ming;WANG Zhao-Qi(Advanced Computing Research Laboratory, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190;Beijing Key Laboratory of Mobile Computing and Pervasive Device, Beijing 100190;University of Chinese Academy of Sciences, Beijing 100049)

机构地区：[1]中国科学院计算技术研究所前瞻研究实验室,北京100190 [2]移动计算与新型终端北京市重点实验室,北京100190 [3]中国科学院大学,北京100049

出　　处：《计算机学报》2017年第11期2606-2618,共13页Chinese Journal of Computers

基　　金：国家自然科学基金(61173067;61379085;61532002);国家"八六三"高技术研究发展计划项目基金(2015AA016401)资助

摘　　要：多片元效果具有实时透明等重要应用,它需要每个像素按深度顺序遍历对应的所有片元.深度剥离法将场景重复绘制多次来满足这个需求,故对显存带宽提出了很高的要求.该文针对大规模场景图元分布稀疏的特点,使用类体素八叉树在物体空间将场景近似剖分以减少图元读取总量.这允许场景能够按可见性顺序被分块加载、精确剖分和逐个绘制:通过光栅化对应的八叉树网格构造深度桶列表,在屏幕空间将场景块与网格布尔求交;通过为整个场景构造块的深度直方图,在剥取时利用它来避免硬件遮挡查询操作.由于为每块采取了不同的迭代次数,该文的两阶段剖分方法能够适应物体空间变化的深度复杂度.由于不依赖于面片的邻接信息,该文方法能够支持非流形网格和外存绘制.相比已有工作,该文方法在每次剥取一层时绘制效率有30%以上的提升.The multi fragment effects have many important applications such as real time transparency,where all the fragments per pixel are required to be visited in depth order.The depth peeling is a fundamental method committed to this task.It resorts to drawing the whole scene multiple times until those fragments are enumerated exhaustively.However,each time only a small portion of the scene contributes to the framebuffer.Most of the primitives and fragments are discarded after visibility tests such as depth comparison,thus wasting too much bandwidth and computation resource.Building upon the sparsity of distributed primitives,we propose a2step method based on the divide and conquer strategy.Our goal is to effectively reduce the total amount of accessed primitives and generated fragments given a large scene with complex geometry.In the first step,the scene is approximately decomposed into parts by using a voxel like octree in object space.This octree and its accompanying mesh is constructed in a depth first manner,to shatter scene primitives from the top down and create scene parts from the bottom up.It enables the results to be serialized as a stream,thereby the peak memory usage is alleviated.With the aid of the octree mesh,all the parts can be loaded and processed in visibility order.We render a frame by batching those parts with respect to the depth complexity of corresponding sub mesh.Every time the largest possible batch is extracted and submitted to the graphics pipeline.A predefined lookup table is introduced to resolve the visibility order of some or even all parts.Although it only encodes the visibility order for a sub mesh template,the full solution can be obtained on the fly by applying this table recursively during traversal.In the second step,all the parts are exactly decomposed by applying bucket sort in screen space.To render each batch of parts in visibility order,we rasterize their corresponding sub mesh beforehand.A per pixel list of depth intervals are constructed in parallel.Then they serve as bucket

关 键 词：八叉树 体素化 深度剥离 桶排序 构造实体几何 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]