格子波尔兹曼方法的GPU并行计算.docVIP

PAGE 57 摘 要 本文基于图形处理单元（GPU）的格子玻尔兹曼方法（LBM）实现了一些经典流场的计算。程序由两个函数迭代实现，碰撞函数和迁移函数。碰撞函数在共享内存内执行，迁移函数在设备内存内执行。为了优化程序性能，将边界条件在碰撞函数内实现，加长碰撞流程，减少设备内存的存取开销。 通过对同一程序设定不同迭代步数，发现GPU格子玻尔兹曼（LB）程序随迭代步数增大，执行时间在前N步随迭代步数先保持线性增长，随后会加速增长，最后回归线性增长，而相应的CPU程序则一直保持线性增长，这导致加速效果会从某一部开始急剧下降。 通过对对同一流场设定不同的格子数，发现GPU格子玻尔兹曼（LB）程序相对同样格子数的CPU程序加速比保持不变。 通过对同一程序设定不同的Block结构，发现随着Block中Thread数的增大，迭代时间线性缩短，在Thread达到某一值时，迭代时间会剧烈减小。 关键词：格子波尔兹曼 并行计算 GPU Abstract Based on graphics processing unit (GPU) Lattice Boltzmann method (LBM), the paper implements some classic flow field calculation . Procedure consists of two iterative function , collision function and transfer function. Collision functions performed within the shared memory, while transfer function within the device memory . In order to optimize the application performance, boundary conditions will be performed within the collision function , which lengthen the collision stream process,thus reduce device memory access . By setting up different iteration steps, we found that for the same GPU lattice boltzmann (LB) program ,with the increase of iteration steps, the program execution time will keep linear growth in N steps, then the growth accelerate, and return linear growth at last, while the corresponding CPU program kept always linear growth, resulting in accelerate ratio will fell sharply from some step. By setting different lattice for the same flow ,we found the GPU lattice boltzmann (LB) procedures remains unchanged speed up ratio relative to the CPU’s of the same lattice. Through imposing different Block structure on the same program, we found that with Threads in Blocks increasing , iteration time shortened linearly first, when the Threads reaches a certain value, the iteration time will drastically reduce. Key words: lattice Boltzmann parallel computing GPU  目录 TOC \o 1-3 \h \z \u 一 绪论 1 1.引言 1 2.格子Boltzmann方法概述 2 3.GPU概述 5 4.本文的主要工作 6 二 格子Boltzmann方法的基本原理 7 1. Boltzmann方程 7 2.格子Boltzmann方程 9 3.格子Bolt