作者: D. Göddeke , D. Ribbrock , M. Geveler , S. Turek , P. Zajac
DOI:
关键词:
摘要: We consider geometric multigrid solvers for linear systems stemming from the finite element discretisation of partial differential equations on unstructured grids. Our implementation technique reduces complete solver to sequences sparse matrixvector multiplications and is thus well-suited both GPUs multicore CPUs. In particular, our can handle several low- high-order spaces in 2D 3D, while only matrix-vector kernel needs receive significant tuning. For benchmark problems, we achieve close an order magnitude speedup a single GPU over multithreaded CPU code.