A two-kernel based strategy for performing assembly in FEA on the graphics processing unit

This paper presents a strategy to perform assembly of system of equations arising in Finite Element Analysis (FEA) on Graphics Processing Units (GPU) based on the principle of dividing the workload into separate kernels. Three different sparse formats are analyzed for efficient storage along with two different implementations for the race condition arising in the traditional assembly (addto method). In the present study, ELL sparse storage format is found to be effective in terms of both storage space and performance on the GPU. To avoid race condition, the method of coloring is found to provide superior result in comparison to the method with atomics for the proposed assembly strategy.