Skip to content

Simplify implementation of local Laplace assembler

Carsten Gräser requested to merge feature/simplify-laplaceassembler into master

While the previous implementation tries hard to optimize several aspects (e.g. exploiting symmetry) I was surprised, that it performes significantly worse than the simple, non-optimized assembler from the dune-functions Poisson example.

In my tests it turned out that removing some optimizations (e.g. don't exploit symmetry) improves performance by 15%-28% (15% for PQ2 on a 2d mixed UGGrid and 28% for Q2 a 3d cube UGGrid).

The test I used is the benchmark example from https://git.imp.fu-berlin.de/agnumpde/dune-fufem-forms. BTW: Using integrate(dot(grad(u),grad(v))) from the latter performes still a bit (14%) better in 2d and significantly (40%) 3d.

Merge request reports