GPU-Hi:GPU RTL平台实现及效率分析  

GPU-Hi:GPU RTL platform implementation and efficiency analysis

作　　者：张立志 赵士彭 章隆兵[1,2,3] ZHANG Lizhi;ZHAO Shipeng;ZHANG Longbing(State Key Laboratory of Computer Architecture,Institute of Computer Technology,Chinese Academy of Sciences,Beijing 100190;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190;University of Chinese Academy of Sciences,Beijing 100049)

机构地区：[1]计算机体系结构国家重点实验室(中国科学院计算技术研究所),北京100190 [2]中国科学院计算技术研究所,北京100190 [3]中国科学院大学,北京100049

出　　处：《高技术通讯》2022年第5期484-492,共9页Chinese High Technology Letters

基　　金：国家自然科学基金(61521092);中国科学院重点部署项目(ZDRW-XH-2017-1)资助。

摘　　要：实现了寄存器传输级(RTL)图形处理器(GPU)研究平台——GPU-Hi。GPU-Hi支持OpenGL 2.0 API,支持统一着色器渲染架构,使用专用集成电路(ASIC)完成图形流水线的固定功能算法,使用单指令多线程(SIMT)架构流处理器完成图形流水线的可编程着色器模块。在使用28 nm工艺的情况下,该平台的物理设计面积为7.9μm^(2)。使用glmark2的测试集作为性能测试程序,完成了该平台的功能正确性验证,同时使用该测试集研究了3D图形应用的计算特性,并进行了GPU微结构级的性能分析。测试结果表明,图形应用的光栅化任务与像素着色任务不随图形应用分辨率等比例增大;同时GPU硬件的光栅化模块性能受着色程序处理能力与显存访问能力的影响。本平台的实现对GPU RTL平台的研究发展有重要的借鉴价值,本文中得到的结论对GPU性能优化具有重要参考意义,有力支持了GPU硬件研究的发展。GPU-Hi register-transfer level(RTL)graphics processing unit(GPU)research platform is established.GPU-Hi supports OpenGL 2.0 API(application programming interface),supports unified shader rendering architecture,uses application-specific integrated circuit(ASIC)to complete the fixed function algorithm of graphics pipeline,and uses single instruction multiple thread(SIMT)architecture to complete the programmable shader module of the graphics pipeline.In the case of using a 28 nm process,the physical design area of the platform is 7.9μm^(2).Using glmark2 test set as the performance test program,the functional correctness of the platform is verified.At the same time,the test set is used to study the computing characteristics of 3D graphics applications,and the performance analysis of GPU microstructure level is performed.The test results show that the rasterization tasks and pixel shading tasks of graphics applications do not increase in proportion to the resolution of graphics applications;at the same time,the performance of the rasterization module of GPU hardware is affected by the processing power of the shading program and the ability of video memory access.The implementation of this platform is important for the research and development of the GPURTL platform.The conclusions obtained in this article have important reference significance for GPU performance optimization and strongly support the development of GPU hardware research.

关 键 词：图形处理器(GPU) 性能分析 glmark2 流处理器集群 

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]