A Two-Level Direct Solver for the Hierarchical Poincaré-Steklov Method

Abstract: The Hierarchical Poincar\'e-Steklov method (HPS) is a solver for linear elliptic PDEs that combines a multidomain spectral collocation discretization scheme with a highly efficient direct solver. It is particularly well-suited for solving problems with highly oscillatory solutions. However, its hierarchical structure requires a heterogeneous and recursive approach that limits the potential parallelism. Our work introduces a new two-level solver for HPS which divides the method into dense linear algebra operations that can be implemented on GPUs, and a sparse system encoding the hierarchical aspects of the method which interfaces with highly optimized multilevel sparse solvers. This methodology improves the numerical stability of HPS as well as its computational performance. We present a description and justification of our method, as well as an array of tests on three-dimensional problems to showcase its accuracy and performance.