面向国产高性能加速器的LLVM编译器设计及优化

doi:10.19678/j.issn.1000-3428.0067000

摘要/Abstract

摘要： 国防科技大学自主研制的高性能加速器采用中央处理器(CPU)+通用数字信号处理器(GPDSP)的片上异构融合架构,使用超长指令集(VLIW)+单指令多数据流(SIMD)的向量化结构的GPDSP是峰值性能主要支撑的加速核。主流编译器在密集的数据计算指令排布、为指令静态分配硬件执行单元、GPDSP特有的向量指令等方面不能很好地支持高性能加速器。基于低级虚拟器(LLVM)编译框架,在前寄存器分配调度阶段,结合峰值寄存器压力感知方法(PERP)、蚁群优化(ACO)算法与GPDSP结构特点,优化代价模型,设计支持寄存器压力感知的指令调度模块;在后寄存器分配阶段提出支持静态功能单元分配的指令调度策略,通过冲突检测机制保证功能单元分配的正确性,为指令并行执行提供软件基础;在后端封装一系列丰富且规整的向量指令接口,实现对GPDSP向量指令的支持。实验结果表明,所提出的LLVM编译架构优化方法从功能和性能上实现了对GPDSP的良好支撑,GCC testsuite测试整体性能平均加速比为4.539,SPEC CPU 2017浮点测试整体性能平均加速比为4.49,SPEC CPU 2017整型测试整体性能平均加速比为3.24,使用向量接口的向量程序实现了平均97.1%的性能提升率。

关键词: 通用数字信号处理器, 低级虚拟器, 编译器, 指令调度, 向量指令接口

Abstract: National University of Defense Technology independently developed a high-performance accelerator that uses an on-chip heterogeneous fusion architecture of a Central Processing Unit(CPU) and General Purpose Digital Signal Processor(GPDSP). The GPDSP,with its Very Long Instruction Word(VLIW)+ Single Instruction Multiple Datastream(SIMD) vectorization structure,is the main support for the peak performance acceleration core. However,mainstream compilers cannot adequately support high-performance accelerators in intensive data calculation instruction layouts,static allocation of hardware execution units for instructions,and GPDSP-specific vector instructions. In this study,based on the Low Level Virtual Machine(LLVM) compilation framework,the PERP method,Ant Colony Optimization(ACO) algorithm,and GPDSP structural characteristics are combined to optimize the cost model in the pre-RA-sched stage,and the instruction scheduling module is designed to support register pressure awareness. This study proposes an instruction scheduling strategy that supports static functional unit allocation in the post-RA-sched stage,which guarantees correct functional unit allocation through a conflict detection mechanism,and provides a software basis for the parallel execution of instructions. Furthermore,a series of rich and regular vector instruction interfaces are encapsulated in the backend to support the GPDSP vector instructions. The experimental results demonstrate that the LLVM compilation architecture optimization method proposed in this study provides good support for the GPDSP in terms of both functionality and performance. Specifically,the overall performance average speedup ratio of GCC testsuite is 4.539,the overall performance average speedup ratio of SPEC CPU 2017 floating-point test is 4.49,and the overall performance average speedup ratio of SPEC CPU 2017 integer test is 3.24. Additionally,the vector program using vector interfaces achieves an average performance improvement ratio of 97.1%.

Key words: General Purpose Digital Signal Processor(GPDSP), Low Level Virtual Machine(LLVM), compiler, instruction scheduling, vector instruction interface

中图分类号:

TP314

宋强, 唐俊龙, 陈照云, 时洋, 谭期轩, 肖紫阳, 邹望辉. 面向国产高性能加速器的LLVM编译器设计及优化[J]. 计算机工程, 2024, 50(4): 321-331.

SONG Qiang, TANG Junlong, CHEN Zhaoyun, SHI Yang, TAN Qixuan, XIAO Ziyang, ZOU Wanghui. Design and Optimization of LLVM Compiler for Domestic High Performance Accelerator[J]. Computer Engineering, 2024, 50(4): 321-331.

https://www.ecice06.com/CN/Y2024/V50/I4/321

参考文献

[1] XU Z W,CHI X B,XIAO N.High-performance computing environment:a review of twenty years of experiments in China[J].National Science Review,2016,3(1):36-48.
[2] WANG H Q,PENG S L,ZHU X Q,et al.A method to accelerate GROMACS in offload mode on Tianhe-2 supercomputer[C]//Proceedings of the 15th IEEE/ACM International Symposium on Cluster,Cloud and Grid Computing.Washington D.C.,USA:IEEE Press,2015:781-784.
[3] PIÑEIRO C,PICHEL J C.A unified framework to improve the interoperability between HPC and Big Data languages and programming models[J].Future Generation Computer Systems,2022,134:123-139.
[4] YIN F,SHI F.A comparative survey of big data computing and HPC:from a parallel programming model to a cluster architecture[J].International Journal of Parallel Programming,2022,50(1):27-64.
[5] HEINECKE A,BREUER A,RETTENBERGER S,et al.Petascale high order dynamic rupture earthquake simulations on heterogeneous supercomputers[C]//Proceedings of International Conference for High Performance Computing,Networking,Storage and Analysis.Washington D.C.,USA:IEEE Press,2014:3-14.
[6] YAN D,WANG W,CHU X.An LLVM-based open-source compiler for NVIDIA GPUs[C]//Proceedings of the 27th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming.New York,USA:ACM Press,2022:448-449.
[7] SHOBAKI G,KERBOW A,PULIDO C,et al.Exploring an alternative cost function for combinatorial register-pressure-aware instruction scheduling[J].ACM Transactions on Architecture and Code Optimization,2019,16(1):1-30.
[8] CHEN S M,WANG Y H,LIU S,et al.FT-Matrix:a coordination-aware architecture for signal processing[J].IEEE Micro,2014,34(6):64-73.
[9] 荀长庆,陈照云,文梅,等.以编译为导向的Matrix-DSP程序分析与优化[J].计算机工程与科学,2020,42(10):1791-1800. XUN C Q,CHEN Z Y,WEN M,et al.Compilation-oriented code analysis and optimization for Matrix-DSP[J].Computer Engineering & Science,2020,42(10):1791-1800.(in Chinese)
[10] PANDEY M,SARDA S.LLVM cookbook[M].[S.l.]:Packt,2015:296.
[11] LOZANO R C,CARLSSON M,DREJHAMMAR F,et al.Constraint-based register allocation and instruction scheduling[C]//Proceedings of International Conference on Principles and Practice of Constraint Programming.Berlin,Germany:Springer,2012:750-766.
[12] SHOBAKI G,GORDON V S,MCHUGH P,et al.Register-pressure-aware instruction scheduling using ant colony optimization[J].ACM Transactions on Architecture and Code Optimization,19(2):23.
[13] DORIGO M,MANIEZZO V,COLORNI A.Ant system:optimization by a colony of cooperating agents[J].IEEE Transactions on Systems,Man,and Cybernetics,Part B:Cybernetics,1996,26(1):29-41.
[14] 刘胜,卢凯,郭阳,等.一种自主设计的面向E级高性能计算的异构融合加速器[J].计算机研究与发展,2021,58(6):1234-1237. LIU S,LU K,GUO Y,et al.A self-designed heterogeneous accelerator for exascale high performance computing[J].Journal of Computer Research and Development,2021,58(6):1234-1237.(in Chinese)
[15] GIESEMANN F,GERLACH L,PAYÁ-VAYÁ G.Evolutionary algorithms for instruction scheduling,operation merging,and register allocation in VLIW compilers[J].Journal of Signal Processing Systems,2020,92(7):655-678.
[16] LOZANO R C,CARLSSON M,BLINDELL G H,et al.Combinatorial register allocation and instruction scheduling[J].ACM Transactions on Programming Languages and Systems,41(3):17.
[17] MALEKI S,GAO Y Q,GARZAR'N M J,et al.An evaluation of vectorizing compilers[C]//Proceedings of International Conference on Parallel Architectures and Compilation Techniques.Washington D.C.,USA:IEEE Press,2011:372-382.
[18] 李嘉楠,韩林,柴赟达.面向国产平台的LLVM自动向量化移植与优化[J].计算机工程,2022,48(1):142-148. LI J N,HAN L,CHAI Y D.Automatic vectorization transplant and optimization of LLVM for domestic processors[J].Computer Engineering,2022,48(1):142-148.(in Chinese)
[19] 冯竞舸,贺也平,陶秋铭.自动向量化:近期进展与展望[J].通信学报,2022,43(3):180-195. FENG J G,HE Y P,TAO Q M.Auto-vectorization:recent development and prospect[J].Journal on Communications,2022,43(3):180-195.(in Chinese)
[20] MAMMADLI R,JANNESARI A,WOLF F.Static neural compiler optimization via deep reinforcement learning[C]//Proceedings of 2020 IEEE/ACM Workshop on the LLVM Compiler Infrastructure in HPC(LLVM-HPC) and Workshop on Hierarchical Parallelism for Exascale Computing(HiPar).Washington D.C.,USA:IEEE Press,2020:1-10.
[21] WU L,PEI J,TANG J,et al.Deep learning on graphs:methods and applications[C]//Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining.New York,USA:ACM Press,2022:4906-4907.
[22] WU Z H,PAN S R,CHEN F W,et al.A comprehensive survey on graph neural networks[J].IEEE Transactions on Neural Networks and Learning Systems,2021,32(1):4-24.
[23] FEY M,LENSSEN J E.Fast graph representation learning with PyTorch geometric[EB/OL].[2023-01-02].http://arxiv.org/pdf/1903.02428.
[24] WANG M J,YU L F,ZHENG D,et al.Deep graph library:towards efficient and scalable deep learning on graphs[EB/OL].[2023-01-02].http://arxiv.org/abs/1909.01315v1.
[25] 池昊宇,陈长波.基于机器学习的编译器自动调优综述[J].计算机科学,2022,49(1):241-251. CHI H Y,CHEN C B.Survey on automatic tuning of compilers by machine learning[J].Computer Science,2022,49(1):241-251.(in Chinese)

选择文件类型/文献管理软件名称

选择包含的内容