Derek Gaston
Derek Gaston
> The C++ way to change function pointers on the fly is with inheritance. Inheritance doesn't work on the fly. Essentially, what the PETSc guys have is the ability to...
Oh - if you do your own pointer and memory pool... don't forget about threading.
Ok - so empty-DSNA is basically what we could get if we implement PotentiallyDualNumber? Or could we do even better in that case because we don't even need to allocate...
@roystgnr thanks for everything you've done! I've done some more monkeying around and I think we're in good shape to go ahead and use this and we can do more...
It's not "inefficient"... it does a LOT of work and gets called A LOT (every side of every element in the mesh). Basic breakdown of what it does: 1. Evaluate...
There is no getting rid of it... for multiple reasons. There is no other way to know where to evaluate the shape functions on the neighbor side. Also: Don't forget...
No - quadrature points on the side of the element are almost _never_ a subset of the quadrature on the element! With Gauss they _never_ are! With mesh adaptivity you...
Ah - I was responding from email - now that I come to Github I see that you've changed your text a bit. Sneaky! 😄
Getting the dofs is a small part of the work. The big chunk of the work is evaluating the FE shape functions on both sides... and there's no way around...
The real way to make an impact on the speed of DG is not by changing MOOSE, but by optimizing the FE stuff in libMesh: https://github.com/libMesh/libmesh/issues/890 That is currently ongoing...