OpenACC到MIC平台上并行程序的自动翻译及优化.pdfVIP

2016 年4 月第4 期 小型微型计算机系统 Journal of Chinese Computer Systems Vol. 37 No.4 2016 OpenACC 到 MIC 平台上并行程序的自动翻译及优化 江霞，安虹，梁伟浩，张爱民，李丰 (中国科学技术大学计算机科学与技术学院，合肥 23∞27) E-mail:jiangxia@mail. 摘 要:随着异构系统在高性能计算领域的流行，各种为异构平台设计的编程模型被提出，比如 CUDA、OpenCL 等，而这些原 生的编程模型存在着编程复杂、优化困难和可移植性差等问题. OpenACC 标准通过添加编译制导语句指示编译器将相关的代 码段或循环卸载到加速器上执行，可以解决这些问题.本文设计并实现了OpenACC 到 MIC 平台上并行程序的自动翻译工具， 从而实现了OpenACC 对Intel 岛但C 协处理器的支持.同时采取了数据传输优化方法对翻译后的代码进行了优化.实验采用完整 的 NPB 测试包，实验结果表明，我们的翻译工具能够百分之百完成测试程序的翻译，并且针对不同规模的数据集翻译后的 O征­ load 代码的性能分别能够达到手工优化的Offload 代码性能的 74% ,76% ,80% . 关键词:异构系统;0严nACC;MIC; 自动翻译;优化 中图分类号:11号11 文献标识码:A 文章编号:1仪10-1220 ( 2016) 04-0824峭 Automatic OpenACC to Intel Omoad and Optimization for MIC JIANG Xia ,AN Hong ,LIANG Wei斗lao ， ZHANG Ai -min ,Ll Feng ( School of Computer Scienc理 and Technology ,University of Science and Technology of China ,Hefei 23∞27 ， αrina) Abs仕act:Wi曲曲e popularity of heterogeneous architectures in the field of High Performance Computi吨， a variety of progr组nming models was put forward ,such as CUDA ,OpenCL ,etc. 咀lese native programming models have some problems , such as programming complexity , difficulty of optimization ,poor portability ,etc. OpenACC can solve these problems by adding compiler directives to identi- fy which area of code or loop to accelerate. This paper implemen岱 at∞1 for transla由g 句enACC into optimized Intel Offload code automatically. Thus we can take advantage of 句enACC to program on Intel MIC coprocessor.ηle optimization technique for the translated code is data transmission optimization. Experiments on NPB benchmark with different input size shows that our tool can complete the translation of all test progr缸ns and achieve 749毛，76% , 80% performance of hand-written code in average respectively. Key words: het