Discretization is a method for transforming continuous variables, equations, functions, and models into their discrete equivalents. A multigrid technique uses a hierarchy of discretization to solve elliptic partial differential equations (PDEs). Computing numerical solutions to the Poisson equation on a cube geometry or a nonlinear minimal surface equation requires fast computers and algorithms. So, how should efficient algorithms be designed for a matrix-free geometric multigrid solver on graphics processing units (GPUs)?

Kronbichler and Ljungkvist concisely review the proposed implementations of matrix-free finite element and multigrid algorithms on GPUs in the literature. Matrix-free finite element algorithms on GPUs have been used to solve specific “time stepping of hyperbolic problems,” for example, seismic wave propagation and industrial generic assembly tools. Capitalizing on the well-established algorithms for computing the matrix-free higher-order finite elements on GPUs in the literature, the authors offer new insights into, and alternative algorithms for, the sum-factorization methods for solving the many matrix-free set up problems of alternative multigrid numerical problems.

The authors offer unique research contributions to the design and implementation of effective algorithms for Laplace operator evaluation, multigrid transfer operations, and vector operations, as well as for handling constraints.

The authors perform several experiments to investigate the effectiveness of multigrid computational ideas on GPUs for solving numerical problems. The experimental results examine impacts related to hardware, setup costs, and the nature of the required numerical solutions for solving alternative multigrid numerical problems. Compared to the existing (and reputable) matrix-free higher-order algorithms, results for the proposed algorithms appear promising, such as the higher-order polynomials on hexahedral meshes. The authors also clearly raise the unresolved issue of efficiency: “Thus, application codes relying on complex solvers need to put focus not solely on the matrix-vector product but also on the whole solver stack including the vector operations, where it is beneficial to optimize for data access.”

I strongly encourage all numerical analysts to read the paper’s insightful ideas and recommend alternative solutions.