面向龙芯处理器SIMD扩展的编译器内在函数优化.pdfVIP

 面向龙芯处理器SIMD 扩展的编译器内在函数优化 凌坤, 胡士文，连瑞琦 ( 中国科学院计算技术研究所 系统结构重点实验室 北京 100190) Optimization of the Intrinsic Support in Loongcc Compiler for Loongson CPU ’s * SIMD Extensions LING Kun, HU Shi-Wen, LIAN Rui-Qi (Key Laboratory of Computer System and Architecture, Institute of Computing Technology CAS, Beijing China 100190 ) Corresponding author: Phn: +86 -10, E-mail: lingkun@ Abstract : Single Instruction Multiple Data (SIMD ）extension has been proven to be a highly effective approach for enhancing the performance on applications that exhibit massive data parallelism. Therefore, it is widely favored by manufacturers, including Loongson processors. However, exploring the full performance power of SIMD units is often a challenging task. We have implemented a robust, flexible and easy-to-use programming interface with the intrinsic support of Loongson SIMD extensions. To pursue automatic performance optimization, we have modularized the techniques developed along the SIMD interface as a compiler featur e. To evaluate the effectiveness of our development, we have chosen a highly optimized video decoder with a assembly coded core to use Loongson 2F SIMD -- Mplayer as the test harness. The experimentation result is satisfactory that after replacing the original assembly core by the new SIMD intrinsic support, we gain 2% performance improvement . Our technique is capable to not only make programmers use C level language for SIMD programming on Loongson platforms, but also make use of advanced modern compiler optimizations for intrinsic. Key words: Loongson CPU; SIMD; compiler intrinsic; compiler optimization 摘 要: 通用微处理器中的单指令多数据(SIMD)指令扩展因能有效提升并行处理能力而一直受到处理器生产厂 商的亲睐，龙芯处理器也增加了这种SIMD 扩展。如何为程序员提供高效、好用的SIMD 编程接口，是工业界和学 界研究的