CG+FFT Poisson solver for CFD problems

LISANDRO D. DALCIN; MARIO A. STORTI; RODRIGO R. PAZ; SANTIAGO D. COSTARELLI

Buenos Aires

Congreso; MECOM del Bicentenario; 2010

Asociación Argentina de Mecánica Computacional

Graphic Processing Units (GPU's) have received much attention in lastyears. Compute-intensive algorithms operating on multidimensionalarrays that have nearest neighbor dependency and/or exploit datalocality can achieve massive speedups.  Simulation of problems modeledby time-dependent Partial Differential Equations (PDE's) by usingexplicit time-stepping methods on structured grids is an instance ofsuch GPU-friendly algorithms.Solvers for transient incompressible fluid flow cannot be developed ina fully explicit manner due to the incompressibilityconstraint. Fractional step methods require the solution of a Poissonproblem for the pressure field at each time level. This stage isusually the most time-consuming one.This work discuss a Poisson solver for the pressure problem onstructured grids and fluid+solid domains. This solver is based onstandard Conjugate Gradients (CG) iterations and Fast FourierTransform (FFT) preconditioning. The implementation targets NVIDIACUDA architecture and takes advantage of freely available softwarecomponents like *Thrust*, *Cusp*, and *CUFFT* libraries.