面向多面体模型的非规则程序SIMD投机优化技术

SIMD Speculative Optimization for Irregular Program Based on Polyhedral Framework

多面体模型具有规范化、健壮性、灵活性等特点,被认为是最有前途的编译优化技术之一,SIMD优化是重要应用领域之一。由于静态编译技术的限制,大量非规则循环无法建立多面体模型,例如复杂依赖关系,导致无法利用该模型实现SIMD循环优化与代码生成。文章通过分析、利用SIMD优化所需的程序运行剖面信息,将静态编译未知情况转化为已知情况,为多面体模型建立创造条件。SPEC2000和PolyBench的测试结果显示,通过本文提出的方法,可将适用多面体模型的非规则循环个数提高2.3倍,经过SIMD优化后,平均加速比提高了1.53倍。

Polyhedral models have characteristics of normalization, robustness and flexibility, but the limitations of static compilation technology make most of irregular loops fail to model polyhedral framework because of aliasing, non-affine expressions, etc. To eliminate these limitations this paper applies the knowledge of the running profile information which is needed by SIMD Optimization . Evaluation on the SPEC2000 and PolyBench shows that the method is able to effectively increase the number of irregular loop which is amenable to polyhedral model by 2.3-fold, and average speed-up by 1.53-fold.