This content discusses the development of code generation techniques for multigrid solvers on octree-based frameworks. It explores communication schemes, interpolation methods, and benchmark results to showcase the benefits of the proposed approach.
The work involves bridging code generation technology with mesh refinement concepts to enhance error convergence in numerical solvers. Various aspects such as data structures, operators, loops, and communication are addressed to optimize multigrid solvers for local refinement.
Key highlights include the implementation of specialized kernels, automatic re-discretization of operators, loop invariant motion mechanisms, and modular designs for efficient communication routines. Benchmark results on SuperMUC-NG CPU cluster demonstrate significant speedups with generated communication schemes.
The study also delves into weak scaling results and performance comparisons between manual baseline implementations and generated communication routines. Future work includes extending support for GPU platforms and optimizing communication protocols further.
לשפה אחרת
מתוכן המקור
arxiv.org
שאלות מעמיקות