基于图元光栅化触发的高效GPU深度数据预取  

An efficient GPU depth data pre-fetching strategy triggered by primitive rasterizing

作　　者：田泽 张骏[1,2,3] 许宏杰 TIAN Ze;ZHANG Jun;XU Hongjie(Xi’an XiangTeng Micro-Electronic Technology Co.Ltd,Xi’an 710068;AVIC Computing Technique Research Institute,Xi’an 710068;Key Laboratory of Aviation Science and Technology on Integrated Circuit and Micro-System Design,Xi’an 710068)

机构地区：[1]西安翔腾微电子科技有限公司,西安710068 [2]航空工业西安航空计算技术研究所,西安710068 [3]集成电路与微系统设计航空科技重点实验室,西安710068

出　　处：《高技术通讯》2022年第11期1126-1133,共8页Chinese High Technology Letters

基　　金：核高基重大专项(2016ZX01012101-004)资助。

摘　　要：数据预取技术已经广泛应用在各类中央处理器(CPU)设计领域,取得了很好的效果。而图形处理器(GPU)对存储带宽的需求更为巨大,与图形渲染流水线和图形算法直接相关,且数据访问模式与通用CPU有显著差异,需要更有针对性的有效数据预取机制。针对GPU深度测试关键功能,本文提出一种图元光栅化触发的高效深度数据预取机制——DPRT,通过图元光栅化过程中实时扫描到的片段块地址来触发Z缓存(Z Cache)的深度数据预取,同时为了适应不同实现中流水线处理延迟,为Z Cache数据块增加访问一次标志(OTT),保证深度数据预取有效性。实验结果表明,DPRT使深度测试时Z Cache访问命中率平均提升9.51%,深度测试延迟平均降低40.43%。Data pre-fetching technique has already been widely applied to many kinds of central processing unit(CPU)design,and obtained very good results.While,graphic processing unit(GPU)has even larger requirement for memory bandwidth,directly interrelates with graphic rendering pipeline structure and graphic algorithm,and has extraordinarily differences in memory accessing pattern in contrast to general CPU,which needs more well-directed and effective data pre-fetching strategy.Aiming at the GPU depth test function,this paper proposes a depth data pre-fetching based on primitive rasterizing triggering,called DPRT,which triggers Z Cache depth data pre-fetching by using current scanned pixel fragment tile address.To adapt graphic rendering pipeline latency in different implementation,a once touching tag(OTT)is added in Z Cache tag array to assure the pre-fetching validation.Experiment result indicates that,comparing to base graphic rendering pipeline structure,DPRT increases Z Cache hit rate by average 9.51%and reduces depth test latency by average 40.43%.

关 键 词：图形处理器(GPU) 光栅化 扫描 深度测试 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP332[自动化与计算机技术—计算机科学与技术]