基于强化学习的循环自动展开研究

Research on loop automatic unfolding based on reinforcement learning

近年来,芯片行业蓬勃发展,新的架构比以往更快地涌现出来。同时,现代计算机的应用场景变得越来越复杂,对计算机性能的要求逐渐增加。在编译优化中,循环展开发挥着承上启下的作用,且任务复杂,高度依赖经验,并需要大量人力和资源投入。为了减少编译器开发中循环展开的工作量,并适应芯片行业快速发展的环境,本文提出了一种基于强化学习的自动展开器。经过实验比较,该循环展开器性能优于Clang-O3,并且与蛮力搜索相比具有更快的编译速度。

In recent years,the chip industry has flourished,and new architectures have emerged at the faster pace than before.At the same time,the application scenarios of modern computers are becoming increasingly complex,and the requirements for computer performance are gradually increasing.Loop unrolling in compilation optimization plays a connecting role,and it is a complex task that highly relies on expert experience and requires a significant investment of manpower and resources.To reduce the workload of loop unrolling in compiler development and adapt to the rapidly developing environment of the chip industry,this article proposes an automatic unroller based on reinforcement learning.After experimental comparison,the proposed unroller performs better than Clang-O3,and it also has a faster compilation speed compared to brute force search.