基于CUDA的离散元法DEM实现方法的改进

发布时间：2018-04-30 02:20

本文选题：GPU + OpenCL　；参考：《吉林大学》2017年硕士论文

【摘要】：DEM(Discrete Element Method)离散元法是一种数值模拟仿真方法,在研究并分析大量颗粒物体在运动过程中产生相互作用方面颇有成效,是目前为止最为流行的一种分析方法,最先由Cundall提出。它的理论核心是离散化运动过程中的非连续物体,从而通过公式计算出非连续体的整体运动状态。快速的计算、较小存储空间的需求是DEM离散元法的特点。本课题组持续多年的研究,对离散元法的理论架构和体系结构有了深入的了解。但是当进行大规模数量的颗粒的运算时仍然会出现由于计算时间过长导致的计算结果不及预期的现象。本课题组成员针对上述问题提出了一种解决方法,使用GPU来替代CPU,从而加快计算速度。GPU在单核性能方面远不及CPU,如处理复杂逻辑算法方面GPU相对CPU稍逊一些;但是在多个核心的并发方面,GPU的并发能力要远远超过CPU,该比例可达1~2个数量级。为了达到GPU替代CPU的目的,该成员使用了并行计算架构Open CL。Open CL是较为普遍的并行计算架构,不局限于硬件条件,因此经常受到大多数科研人员的青睐。但是,在性能与精确度方面,远不如CUDA并行计算架构,如在处理双精度数值时,由于Open CL只能处理单精度,在计算结果中会出现较大的误差。CUDA是由NVIDIA公司提出的一种并行计算架构。该计算架构提供了双精度运算,同时还提供了大量丰富的API,在提高开发效率的同时降低了实现设备端程序的成本。本论文在已有的理论基础与算法之上,使用CUDA,对以Open CL颗粒系统进行改进,从而满足了运算速度与精确度的要求。通过大量数据的测试,CUDA颗粒系统的计算结果与Open CL颗粒系统的计算结果相似,在使用相同计算参数时程序的运行时间差别不大,但是在颗粒细微运动下的仿真效果更为精确。同时,根据算法和CUDA并行计算架构的特点,对CUDA颗粒系统进行优化,从而进一步加快运行速度。最后,本论文对CUDA颗粒系统进行了全面测试,通过颗粒与颗粒、颗粒与多种图元边界碰撞的测试来分析CUDA颗粒系统的正确性以及运行效率。
[Abstract]:DEM(Discrete Element method discrete element method is a numerical simulation method, which is very effective in studying and analyzing the interaction of a large number of particles in the process of motion. It is the most popular analysis method by far, first proposed by Cundall. The core of its theory is the discontinuous object in the process of discrete motion, and the whole motion state of the discontinuous body is calculated by the formula. Fast calculation and small storage space are the characteristics of DEM discrete element method. After many years of research, we have a deep understanding of the theoretical framework and architecture of the discrete element method (DEM). However, when a large number of particles are calculated, the calculation results are not as expected due to the long calculation time. In view of the above problems, the members of our group put forward a solution, using GPU to replace CPU, so as to speed up the calculation speed. GPU in the performance of single core is far less than CPU, such as handling complex logic algorithm GPU is a bit inferior to CPU; However, the concurrency capability of GPUs in multiple core concurrency is much higher than that of CPU, and the ratio can reach 1 ~ 2 orders of magnitude. In order to replace CPU with GPU, Open CL.Open CL is a common parallel computing architecture, which is not limited to hardware, so it is often favored by most researchers. However, in terms of performance and accuracy, it is far less than the CUDA parallel computing architecture, for example, when dealing with double-precision values, because Open CLs can only handle single-precision, Large errors occur in the calculation results. CUDA is a parallel computing architecture proposed by NVIDIA. The architecture provides double precision operation and a lot of rich API, which improves the development efficiency and reduces the cost of implementing the device side program. Based on the existing theories and algorithms, this paper improves the particle system with Open CL by using CUDA, which meets the requirements of operation speed and accuracy. The calculation results of CUDA particle system are similar to those of Open CL particle system, and the running time of the program is not different when the same calculation parameters are used, but the simulation results are more accurate under the fine motion of particles. At the same time, according to the characteristics of the algorithm and CUDA parallel computing architecture, the CUDA particle system is optimized to further accelerate the running speed. Finally, the CUDA particle system is tested in this paper, and the correctness and operation efficiency of CUDA particle system are analyzed by the collision test of particle and particle, particle and many kinds of graph element boundary.
【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TB115

【相似文献】