GPUのための前処理付き共役勾配法(<特集>GPGPUコンピューティングの数理)

概要

論文の詳細を見る
Many Preconditioning algorithms for Krylov subspace methods are developed and well tested on parallel computing environment like PC clusters, but performance of those methods on GPU differs from those on clusters. Although incomplete LU decomposition is the most popular preconditioner for Krylov subspace methods on CPU, its implementation on GPU was never successful because of the sequential nature of the algorithm. By using C for CUDA environment, we have efficiently implemented ILU(0) on GPU. We have also implemented and evaluated various kinds of preconditioning methods for Krylov subspace solvers on GPU, including Jacobi, Red-Black Gauss-Seidel, Line-by-Line and ILU(0). ILU(0) preconditioner turned out to be most efficient and fastest among these preconditioners when solving Poisson equation.
2010-06-25