一种面向55 nm工艺的可扩展统一架构图形处理器设计与实现

摘要/Abstract

摘要：

现代3D图形处理器已从固定渲染管线发展成可编程渲染管线，且其并行度越来越高，研究并设计高性能的3D图形处理器对3D图形处理具有重要意义。着色器是实现3D图形处理器的核心，因此开发性能高、面积小、功耗低又易于扩展的着色器对3D图形处理器的开发具有重要作用。提出的统一架构图形处理器基于单指令多线程和单指令多数据，单指令多线程可以提高图形处理的并行度，从而提高图形处理性能；单指令多数据可以降低设计复杂度，从而实现面积小、功耗低又易于扩展的着色器。实验结果表明，提出的统一架构图形处理器在面积较小、功耗较低的情况下实现了较高的性能，且设计可扩展性较好。

关键词: 图形处理器, 统一架构, 着色器, 单指令多线程, 单指令多数据

Abstract:

Modern 3D Graphics Processing Unit (GPU) has evolved from a fixed-function rendering pipeline to a programmable rendering pipeline with much more parallelism.Research and design of highperformance 3D GPU is important for 3D graphics processing.Shader is crucial to implement 3D GPU,so developing a shader with high performance,small area,low power and ease of extension is of great significance.Based on Single Instruction Multiple Threads (SIMT) and Single Instruction Multiple Data (SIMD),an extendable unified architecture GPU is proposed. SIMT can achieve more parallelism,resulting in higher performance.SIMD simplifies the design complexity,achieving a shader with less area,lower power and ease of extension.Experimental results show that the proposed extendable unified architecture GPU provides higher performance,smaller area,lower power,and better extension.

Key words: GPU;unified architecture;shader;single instruction multiple threads (SIMT);single instruction multiple data (SIMD)

黄亮，秦信刚，武玲娟，熊庭刚. 一种面向55 nm工艺的可扩展统一架构图形处理器设计与实现[J]. J4, 2014, 36(12): 2418-2423.

HUANG Liang，QIN Xingang，WU Lingjuan，XIONG Tinggang. Design and implementation of an extendable unified
architecture GPU based on 55nm technology [J]. J4, 2014, 36(12): 2418-2423.

参考文献［16］

［1］	AMDNEWS ［EB/OL］. ［20110628］. http://www.amdnews.cn/201 1/0628/1747_3.html.
［2］	Lindholm E, Nickolls J, Oberman S, et al. Nvidia tesla:A unified graphics and computing architecture［J］. IEEE Micro, 2008, 28(2):3955.
［3］	Ati radeon hd 2000/3000 series［EB/OL］. ［20140606］. http://en.m.wikipedia.org/wiki/Radeon_R600.
［4］	Woo J H, Sohn J H, Nam B G, et al. Mobile 3d graphics soc from algorithm to chip［M］. Singapore:John Wiley & Sons(Asia) Pte Ltd, 2010.
［5］	Nam B G, Kim H, Yoo H J. A lowpower unified arithmetic unit for programmable handheld 3d graphics systems［J］. IETable 1Performance comparison of different shaders表1不同着色器的性能对比（V表示顶点着色器，F表示片段着色器）文献［14］文献［11］文献［13］文献［12］文献［15］文献［10］文献［16］本文工艺/nm1801801801801301806555功能单元VVVV+FV+FVV+FV+F数据类型FloatFloatFloatFloatFloatFloatFloatFloat核心频率/MHz100110100200166200400600顶点渲染性能/(M顶点/s)1201861205050141300600片段渲染性能/(M片段/s)---5050-400600功耗/mW231.8161157155407153248238面积/(M与非门)1.52.01.52.0-1.571.171.02Table 2Comparison of power efficiency and area efficiency表2功耗效率与面积效率对比(V表示顶点着色器，F表示片段着色器)文献［14］文献［11］文献［13］文献［12］文献［15］文献［10］文献［16］本文功耗效率(V)0.5181.1150.7640.3230.1230.9221.2092.521功耗效率(F)---0.3230.123-1.6132.521硬件效率(V)80938025-89.8256.4588.2硬件效率(F)---25--341.9588.2EE Journal of Solid State Circuits, 2009, 42(8):17671778.
［6］	Kim W Y, Lee B H, Lee K Y, et al. Design of a fully programmable shader processor for low power mobile devices［C］∥Proc of IEEE TENCON, 2009:15.
［7］	Woo J H, Sohn J H, Kim H, et al. A 152 mw mobile multimedia soc with fully programmable 3d graphics and mpeg4/h.264/jpeg［J］. IEEE Transactions On Very Large Scale Integration(VLSI) Systems, 2009, 17(9):12601266.
［8］	Sohn J H,Park Y H,Yoon C W,et al.Low power 3d graphics processors for mobile terminals［J］. IEEE Communication Magazine, 2005,43(12):9099.
［9］	Woo R. A 210 mw graphics lsi implementing full 3d pipeline with 264mtexels/s texturing for mobile multimedia applications［J］. IEEE Journal of Solid State Circuits, 2004, 39(2):358367.
［10］	Nam B G, Lee J, Kim K. A 52.4 mw 3d graphics processor with 141m vertices/s vertex shader and three power domains of dynamic voltage and frequency scaling［C］∥Proc of IEEE International Solid State Circuits Conference, 2007:278279.
［11］	Yu C H, Chung K, Kim D. A 186m vertices/s 161 mw floatingpoint vertex processor with optimized datapath and vertex caches［J］. IEEE Transactions on Very Large Scale Integration(VLSI) Systems, 2009, 17(9):13691382.
［12］	Sohn J H, Woo J H, Lee M W, et al. A 155 mw 50m vertices/s graphics processor with fixedpoint programmable vertex shader for mobile applications［J］. IEEE Journal of Solid State Circuits, 2006, 41(5):10811091.
［13］	Yu C H, Chung K, Kim D, et al. A 120m vertices/s multithreaded vliw vertex processor for mobile multimedia applications［C］∥Proc of ISSCC, 2006:408409.
［14］	Yu C H, Chung K, Kim D. An energyefficient mobile vertex processor with multithread expanded vliw architecture and vertex caches［J］. IEEE Journal of Solid State Circuits, 2007, 42(10):22572269.
［15］	Kim D, Chung K, Yu C H, et al. An soc with 1.3g texels/s 3d graphics full pipeline engine for consumer applications［J］. IEEE Journal of Solid State Circuits, 2006, 41(1):7184.
［16］	Sun Gangde.Design and research of unified architecture shader based on automatic threading and vliw［D］. Hangzhou:Zhejiang University, 2012. (in Chinese)
	附中文参考文献：
	孙纲德. 基于自动线程和超长指令的统一架构着色器的设计研究［D］. 杭州:浙江大学, 2012.

[1]	王智广, 张腾畅, 吴相锦, 鲁强, . 基于GPU的海量离散点高程并行插值算法[J]. 计算机工程与科学, 2021, 43(04): 614-619.
[2]	陈虎, 韩建国. GPU上典型存储器难散列函数的优化[J]. 计算机工程与科学, 2020, 42(10高性能专刊): 1905-1912.
[3]	贾朝阳, 张敦博, 王琼, 沈立. 一种高效的压缩Page Walk Cache结构[J]. 计算机工程与科学, 2020, 42(09): 1521-1528.
[4]	方程，邢座程，陈顼颢，张洋. 一种基于GPU的高性能稀疏卷积神经网络优化[J]. 计算机工程与科学, 2018, 40(12): 2103-2111.
[5]	晏敏1，何欣1，李沙1，祝龙1，赵丽2. 基于一阶泰勒级数查表法单精度倒数的设计与实现[J]. 计算机工程与科学, 2017, 39(07): 1269-1272.
[6]	刘旭江，徐圆，齐宏亮，洪虹，周凌宏. 基于高斯拉普拉斯的层次剥离体绘制[J]. J4, 2014, 36(06): 1148-1153.
[7]	黄立波，王志英，沈立，马胜. 一种低成本128位高精度浮点SIMD乘加单元的设计与实现[J]. J4, 2012, 34(9): 71-76.
[8]	张连伟1，2，刘大学2，刘肖琳2，李焱2，徐昕2，贺汉根2. 基于图形处理器的点云快速光顺[J]. J4, 2011, 33(4): 86-92.
[9]	刘耀林邱飞岳王丽萍. 基于GPU的图像快速旋转算法的研究及实现[J]. J4, 2008, 30(6): 48-50.