The Study Of Acceleration Technology Of GPU-based3D Smoke Simulation

Fluid simulation, particularly simulation of natural phenomena, such as churning smoke,turbulent streams, always has been a hot and difficult topic in computer graphics researchdomain. In recent years, researchers have achieved a lot of progress in the physically-basedfluid simulation, which has been widely used in move industry, video games, advertising andother areas. However, due to the computational complexity of this method, it can’t meetreal-time requirement, which limits its applications seriously. To improve the efficiency offluid simulation, this paper has done an in-depth research in the following aspects:1. This paper introduced three physically-based fluid simulation methods, Euler method,Lagrange method and Lattice Boltzmann method, and chooses the grid-based Euler method tosolve the Navier-Stokes equations. The Semi-Lagrangian method is used to compute theadvection term and the implicit iteration method used to compute the diffusion term andpressure Possion equation. Then we paraellizes the separated step (such as advection,diffusion, pressure Possion and external force term) with CUDA technology by mapping theseto GPU architecture, while those sequential step are executed by CPU. With such a hybridGPU+CPU way, the performance can be improved.2. This paper introduced multigrid method to solve the pressure Possion equation in NSequations to reduce the computation cost of fluid simulation. We constructed different scalegrids, while the fine grids is used to attenuate swing component, the coarse grids used tosmooth the swing component, here restriction operator and prolongation operator are used toeliminate errors among the different scale grids. In this paper, we implemented multigridmethod on GPU. The experimental results show it can speed up the iterative convergence rate,improve computing performance for fluid simulation.3. This paper introduced sharing system Multi-GPU approach to accelerate thesimulation of large-scale fluid phenomena. The large-scale fluid region is divided into severalsub-regions, and these sub-regions are mapped to different GPUs for cooperative computing. In the course of simulation, every GPU corresponding to the host thread communicates witheach other and synchronizes. In this paper, we simulated smoke movement on Multi-GPU.The results show that sharing system Multi-GPU approach is more efficient for the simulationof hyper-scale fluid phenomenon.