超燃冲压发动机一维模型的GPU并行加速研究  

One-dimensional modeling of scramjet based on GPU parallel acceleration

作　　者：温思歆 苏承毅[2] 王东杰 孟万植 聂聆聪[2] 孙希明 WEN Sixin;SU Chengyi;WANG Dongjie;MENG Wanzhi;NIE Lingcong;SUN Ximing(School of Control Science and Engineering,Dalian University of Technology,Dalian 116000,China;Beijing Power Machinery Institute,Beijing 100074,China)

机构地区：[1]大连理工大学控制科学与工程学院,辽宁大连116000 [2]北京动力机械研究所,北京100074

出　　处：《推进技术》2024年第10期247-256,共10页Journal of Propulsion Technology

基　　金：国家科技重大专项(J2019-Ⅰ-0019-0018)。

摘　　要：发动机模型是控制计划优化、基于模型的控制和观测器设计等技术的基础,对控制系统的性能具有重要影响。然而,超燃冲压发动机一维模型由于依赖计算流体力学的网格计算,其计算量巨大,难以在机载控制器内实时运行。为解决这一问题,本文深入研究基于图形处理器(Graphics Processing Unit,GPU)的并行计算技术,探索了网格解耦与划分、串/并异构设计、内存优化、代码优化、编译指令优化、硬件模式优化等方法,综合设计了一个高效的中央处理器(Central Processing Unit,CPU)+GPU异构模型,并在基于虚拟路径交叉连接(Virtual Path Cross-Connect,VPX)总线的嵌入式控制器上进行验证。为充分验证所设计异构模型的有效性、高效性与实时性,本文依次开展基线测试、软硬件优化加速测试、并行计算测试,在测试中对比一维模型在CPU,单核GPU,多核GPU上的计算耗时与数据误差。最后根据数据、曲线、监控工具等方式,展示了所设计的异构模型在不损失模型精度的前提下,加速超过了6.7倍,运行时间均不超过25 ms,符合工程预期的实时性要求,具有良好的应用前景。The engine model serves as the foundation for various technologies such as control plan optimiza⁃tion,model-based control,and observer design,all of which significantly impact the performance of control sys⁃tems.However,the computational requirements of one-dimensional models for scramjets are immense,making real-time execution on onboard controllers challenging.To address this issue,this study delves into the research of GPU(Graphics Processing Unit)-based parallel computing techniques and explores methods such as grid de⁃coupling and partitioning,serial/parallel heterogeneous design,memory optimization,code optimization,compi⁃lation instruction optimization,and hardware mode optimization.By integrating these approaches,an efficient CPU(Central Processing Unit)+GPU heterogeneous model is designed and validated on the embedded controller based on VPX(Virtual Path Cross-Connect)bus.To adequately verify the effectiveness,efficiency,and realtime performance of the designed heterogeneous model,baseline tests,hardware and software optimization accel⁃eration tests,and parallel computing tests are conducted in this paper.In the tests,the time consumption and da⁃ta errors of the one-dimensional model on CPU,single-core GPU,and multi-core GPU are compared.Finally,leveraging data analysis,graphical representations,and monitoring tools,the study conclusively demonstrates that the designed heterogeneous model achieves an acceleration exceeding 6.7 times without compromising model accuracy.Importantly,none of the execution times surpass 25 ms,aligning with the real-time requirements es⁃sential for engineering applications.The methodologies investigated in this study showcase promising prospects for practical implementation.

关 键 词：超燃冲压发动机 并行计算 一维模型 嵌入式控制器 优化加速 

分 类 号：V231.3[航空宇航科学与技术—航空宇航推进理论与工程]