Accelerating Stencil Computation on GPGPU by Novel Mapping Method Between the Global Memory and the Shared Memory

Abstract

Acceleration of stencil computation can be effectively improved by utilizing the memory resource. In this paper, in order to reduce the branch divergence of traditional mapping method between the global memory and the shared memory, we devise a new mapping mechanism in which the conditional statements loading the boundary stencil computation points in every XY-tile are removed by aligning ghost zone to reduce the synchronization overhead. In addition, we make full use of single XY-tile loaded into registers in every stencil computation point, common sub-expression elimination and software prefetching to reduce overhead. At last detailed performance evaluation demonstrates our optimized policies are close to optimal in terms of memory bandwidth utilization and achieve higher performance of stencil computation.