On the GPU-accelerated Preconditioners for Pressure Poisson Equation
Mandhapati Raju, Jaber J. Hasbestan, Nitesh O. Attal, Anshul Mittal · 2023
View Video Presentation: https://doi.org/10.2514/6.2023-3429.vid In the context of the finite volume discretization of fluid dynamics equations, pressure equation is computationally expensive to solve. Ever-increasing performance of GPUs presents an opportunity to reduce the overall turn-around time for industrial scale simulations. GPUs excel at Matrix Vector multiplications and are a natural fit for Krylov type solvers. We use the flexible BiCGSTAB as our linear solver and study the performance of two preconditioners for solving pressure Poisson equation. The solver as well as the preconditioners are implemented using CUDA and benchmarked on NVDIA V100 GPUs. The first preconditioner is a reduced SOR preconditioner (RSOR) and the second is based on Chebyshev iterations. The performance of the preconditioned BiCGSTAB solver is compared with the NVIDIA’s open-source AMGX library in terms of the time taken for solving the pressure equation. The simplicity of the implementation of these preconditioners along with their satisfactory performance presents them as good candidates for solving pressure Poisson equation on GPUs.