The Numerical Study Of Simple Based On GPU

With the improvement of the computational accuracy and the continuous expansion of the computational domain,the required period for the simulation of flow and heat transfer is becoming longer.The parallel algorithm with higher computational efficiency and less cycle for the simulation,has received widespread attention in the actual computation.As a common and cheaper parallel computing device,GPU(Graphics Processing Units)has become a research focus to study heterogeneous parallel algorithms in the simulation of flow and heat transfer.The numerical simulation of flow and heat transfer operated on CPU(Central Processing Unit)and GPU was developed firstly in this thesis with the improved GPU parallel implementation.A fine-grained,completely parallelized Red-Black SOR solver,called pure floating-point method was developed to solve the problem of low utilization rate of threads of the traditional Red-Black SOR solver on GPU,where the adjacent points' remapping process was completed without logical operation.The proposed GPU algorithm' technique for floatingpoint Red-Black iteration is 70%-80% faster than traditional Red-Black algorithm.Compared with the direct implementation of monochrome mapping algorithm,the computation efficiency of floating-point Red-Black solver is improved by 20%-30%.2D and 3D heat conduction cases showed that the heterogeneous parallel solver could save 80%-90% computational time compared with serial program on CPU.Based on above work,the parallel model of SIMPLE(Semi-Implicit Method for Pressure Linked Equations)was established in OpenCL+OpenMP mode,including the pressure correction equation GPU acceleration pattern and full variables GPU acceleration pattern.Furthermore,the local storage on GPU was used to optimize the algorithm togather with the iterative error solving model.2D and 3D cases showed that the heterogeneous parallel SIMPLE solver could save 90%-93% run time compared with CPU serial code.The computational mesh size has a great influence on the acceleration performance of the algorithm,where the acceleration effect is poor on the small mesh size.Other factors,such as the flow heat transfer condition,the discrete scheme of convective terms,and the type of problem,have little effect on the performance of the SIMPLE parallel GPU solver.