Global optimization of numerical programs via prioritized stochastic algebraic transformations
Xie Wang,Huaijin Wang,Zhendong Su,Enyi Tang,Xin Chen,Weijun Shen,Zhenyu Chen,Linzhang Wang,Xianpei Zhang,Xuandong Li
DOI: https://doi.org/10.1109/ICSE.2019.00116
2019-01-01
Abstract:Numerical code is often applied in the safety-critical, but resource-limited areas. Hence, it is crucial for it to be correct and efficient, both of which are difficult to ensure. On one hand, accumulated rounding errors in numerical programs can cause system failures. On the other hand, arbitrary/infinite-precision arithmetic, although accurate, is infeasible in practice and especially in resource-limited scenarios because it performs thousands of times slower than floating-point arithmetic. Thus, it has been a significant challenge to obtain high-precision, easy-to-maintain, and efficient numerical code. This paper introduces a novel global optimization framework to tackle this challenge.
Using our framework, a developer simply writes the infinite-precision numerical program directly following the problem's mathematical requirement specification. The resulting code is correct and easy-to-maintain, but inefficient. Our framework then optimizes the program in a global fashion (i.e., considering the whole program, rather than individual expressions or statements as in prior work), the key technical difficulty this work solves. To this end, it analyzes the program's numerical value flows across different statements through a symbolic trace extraction algorithm, and generates optimized traces via stochastic algebraic transformations guided by effective rule selection. We first evaluate our technique on numerical benchmarks from the literature; results show that our global optimization achieves significantly higher worst-case accuracy than the state-of-the-art numerical optimization tool. Second, we show that our framework is also effective on benchmarks having complicated program structures, which are challenging for numerical optimization. Finally, we apply our framework on real-world code to successfully detect numerical bugs that have been confirmed by developers.