Matrix-Free Solvers for Nonlinear Problems - Performance Optimizations and Implementational Details

DI Daniel Jodlbauer

Jan. 21, 2020, 2:30 p.m. S2 054

This talk is intended to cover some implementational topics that arise when we want to solve nonlinear partial differential equations. In particular, we will discuss some features of a (preliminary) framework for the matrix-free solution of nonlinear PDEs. A single program could never implement a (reasonably good) solverfor all possible PDEs. Hence, the described framework does not attempt to do that, but rather to provide a variety of common building blocks necessary for tackling nonlinear PDEs.

Furthermore, we want to address some less frequently used performance optimizations with potentially huge impact on the computational time, like SIMD parallelization or static inheritance. All code details are intended for C++, but may carry over to other languages as well.