A Data Structure for Optimizing Coalescence and Code Divergence in GPU based solutions for Finite Difference Methods

Computational finite difference methods for partial differential equations presents data stencil patterns, when calculating time steps in a discrete domain. Due memory access strategies, parallel implementation of these methods may benefit from data locality if the data is arranged accordingly. In this paper we propose a new data structure and memory access pattern that may reduce non-coalescent
memory access and code divergence when loading neighborhood data in a GPU architecture. In our strategy, we use subdomains and extra buffers in order to optimize global memory access by warps. Our method achieves up to 1.3 times faster results than a classical strategy for finite difference GPU implementation