CUDA内存资源池设计避免访问冲突.pdfVIP

2013 年 II 月第II 期 小型微型计算机系统 Journal of Chinese Compuler Syslems Yo l. 34 No. II 20日 CUDA 内存资源池设计避免访问冲突 于玉龙\王宇新2 ， 郭忠明1 ，唐 垄1 ， 郭禾l I( 大连理工大学软件学院.辽宁大连 116621 ) 2( 大连理工大学计n:机科学与技术学院，辽宁大连 11ω24 ) E.m创1: yuylong@ acm. org 摘 要: CUOA 内存资源往往存在给态申请与访问的约末，多 CPU 线毛主并发执行时会产生资源访问冲突.经建挨分析发况，消 除资源与设备核函数之间的静态关系可避免该问题.基于资源池模式设计 CUOA 内存资源池，整合多份资源1 本，增加相应调 皮管理机制，达到了消除资源访问冲突的目的.以纹Z室内存为案例，在确认纹皮肉存线4呈间访问冲突存在..发生频率的基础上. 设计并实现纹理内存资源池. 大量实验证明资源池不但能够保证程序正确性， 避免访问冲突，而且具有较小的性能损耗与负载 开销，较低的代码给护成本，同时适用于多 GPU 系统. 关键询: CUOA;多线程:纹理内存;访问冲突;内存资源池 中图分类号: TP3 11 文献标iR码: A 文章编号:1仪沁-1220(201 3) 11-2583{)5 Memory Resource Pool Devisal in CUDA to Avoid Acc臼sing Collision l 2 YU Yu.long ,WANG Yu也n ,GUO Zhong-ming ,TANG Kun , GUO He 。 (Sch∞1 of Softwar慧 Technology. Dalian University of Technology. Dalian 116621. Cbina) 2(Sch∞1 of Compuler Science and Technology. Daliωυniversity of Technology. Dalian 11ω24. China) Abstract: There are usually static a1location and accessing restrictions in using the CUOA device memories , which en∞町age the colli- sion 忧tween multiple CPU threads instances when concurrenl accessing. A method to avoid this issue is found by modeling and ana- Iyzing to d巳couple 山e static r芭lationship between the resource and the device kernel. The solution is proposed based on Resource P∞l Pattern with multiple resource copies and the managemenl mechanism 10 remove the accessing collision. Texture memory is selected for case s阳dy. 币le texture 阴沟1 is devised and implemented after 由e confrrmation of existence and frequency of the lexω陀 memory accessing collision between muJtiple threads . The experimental results i1Iustrate thal the reso山田 伊x1 brings accuracy , better code ma- intainability with IittJe overbead on 仰rformance and load , and cωbe used in multi-GPU syste