国产DCU加速卡与MPS方法结合高效模拟带障碍物溃坝流动问题  

Efficient Simulation of Obstructed Dam-break Flows Using the MPS Method on Domestic DCU Accelerator Card

作　　者：黄聪祎 赵伟文 万德成[1] Huang Congyi;Zhao Weiwen;Wan Decheng(Computational Marine Hydrodynamics Lab(CMHL),School of Naval Architecture,Ocean and Civil Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学船舶海洋与建筑工程学院船海计算水动力学研究中心(CMHL),上海200240

出　　处：《水动力学研究与进展（A辑）》2024年第2期187-195,共9页Chinese Journal of Hydrodynamics

基　　金：国家自然科学基金(52131102)。

摘　　要：移动粒子半隐式方法(Moving Particle Semi-implicit, MPS)因其拉格朗日特性而具有独特优势,尤其适合用于模拟具有自由表面大变形特征的强非线性问题。然而,邻居粒子搜索和压力泊松方程求解的高计算开销限制了该方法在大规模应用场景中的实用性。图形处理器(GPU)因其多核架构,非常适合用于并行模拟处理这种相似且大量的问题。该文将自主开发的基于GPU加速的MPS算法程序移植到基于HIP编程环境的国产自主类GPU平台DCU (Deep Computing Unit,深度计算单元)上运行,开发了MPSDCU-SJTU求解器。利用MPSDCU-SJTU模拟了三维溃坝流动,其模拟结果与已公开发表的实验结果吻合良好。进一步比较了DCU平台加速前后的计算时长,证明MPSDCU-SJTU求解器能显著提升MPS方法的计算效率。The moving particle semi-implicit method(MPS)has unique advantages due to its Lagrangian characteristics,making it particularly suitable for simulating strong nonlinear problems with large deformation characteristics of free surfaces.However,the high computational cost of neighbor particle searching and solving the Pressure Poisson Equation limits the practicality of this method in large-scale application.Graphics processing unit(GPU)is highly suitable for parallel simulation of similar and numerous problems due to its multi-core framework.In this paper,our in-house GPU accelerated MPS algorithm program MPSGPU-SJTU is migrated to the domestic GPU-like accelerator deep computing unit(DCU)based on the HIP programming environment,and an MPSDCU-SJTU solver is developed.The three-dimensional dam break flow is simulated using MPSDCU-SJTU,and the results are in good agreement with the published results.Further comparison is made between the calculation time before and after DCU acceleration,demonstrating that the MPSDCU-SJTU solver can significantly improve the computational efficiency.

关 键 词：移动粒子半隐式方法(MPS) 自由表面流动 深度计算单元(DCU) 异构计算 

分 类 号：TU476[建筑科学—结构工程]