A note on the vectorization of algebraic multigrid algorithms

摘要

The algebraic multigrid code AMGIRI is a flexible sparse linear equation solver which achieves very good rates of convergence. We discuss algorithms for vectorizing the solution phase of the code, including a modification of the current Gauss–Seidel iteration and some methods for reducing the overhead associated with these techniques. High degrees of vectorization are achieved, and experiments on supercomputers with hardware gather/scatter show satisfactory speedups. This approach can be applied to other sparse linear-algebra iterations to enhance vectorization.