high performance computing on graphics processing units: hgpu.org

In pursuit of obtaining high fidelity solutions to the fluid flow equations in a short span of time, Graphics Processing Units ( GPUs ) which were originally intended for gaming applications, are currently being used to accelerate Computational Fluid Dynamics codes. With a high peak throughput of about 1 TFLOPS on a PC, GPUs seem to be favorable for many high resolution computations. One such computation that involves a lot of number crunching, is computing time accurate flow solutions past apping wings at low speeds. The aim of the present paper is thus to discuss the development of an incompressible Navier-Stokes (INS) solver on unstructured and overset grids, and its implementation on GPUs. In its present form, the GPUINS solver solves the fluid flow equations on an unstructured/hybrid grid in the full Pressure-Poisson formulation with consistent treatment of the derivatives for the Pressure-Poisson equation (PPE) in three-dimensions. Since the equations are solved in a semi-implicit form (with viscous terms treated implicitly), the discretization results in a set of linear equations also for velocity. Hence the backbone of the GPU computation relies on developing efficient iterative linear solvers. The BiCGSTAB iterative algorithm is parallelized in a matrix free approach using several GPU kernels such as gradient, Laplacian and reduction. Some of the simple arithmetic vector calculations are implemented using the CU++ET approach where kernels are automatically generated at compile time. The solver is validated using standard test cases such as the (1) flow in a driven cavity, (2) flow past a cylinder, (3) flow past a plunging airfoil and (3) flow past a sphere (with and without overset grids).