Adaptive Multigrid #4726
Adaptive Multigrid #4726
Conversation
pbrubeck
force-pushed
the
pbrubeck/adaptive-mg
branch
2 times, most recently
from
b1a3ac2 to
cc7d814
Compare
pbrubeck
force-pushed
the
pbrubeck/adaptive-mg
branch
from
cc7d814 to
ff94938
Compare
| """ | ||
|
|
||
| def __init__(self, mesh, refinements_per_level=1, nested=True): | ||
| self.meshes = tuple(mesh) |
There was a problem hiding this comment.
Why does it make two independent tuples of mesh?
| Then computes the coarse_to_fine and fine_to_coarse mappings. | ||
| Constructs intermediate submesh hierarchies with this. | ||
| """ | ||
| if mesh.topological_dimension <= 2: |
There was a problem hiding this comment.
Are these bounds empirically determined, or are they guaranteed by theory?
| (coarse_splits, fine_splits, num_children) = split_to_submesh( | ||
| mesh, coarse_mesh, c2f, f2c | ||
| ) | ||
| for i in range(1, max_children + 1): |
There was a problem hiding this comment.
Can we add a comment for what this is doing?
There was a problem hiding this comment.
More generally some comments for all the steps below would be useful, I don't have the big picture of what they are trying to do.
| markers = Function(W) | ||
|
|
||
| with eta.dat.vec_ro as eta_: | ||
| eta_max = eta_.max()[1] |
There was a problem hiding this comment.
Is this taking the max over all processes (global communication), or just over the local part of eta?
|
|
||
| def adapt(self, eta, theta): | ||
| """ | ||
| Implement Dorfler marking, refines mesh from error estimator |
There was a problem hiding this comment.
I don't think this is correct, Dörfler marking involves sorting all of the elements by decreasing error estimator and taking the minimal set that exceeds some fixed fraction of the total error. What this code implements is the simpler variant that doesn't have a proof of convergence (as far as I know) but works as well in practice.
|
|
||
| def recombine(self, split_funcs, f, child=True): | ||
| """ | ||
| Recombines functions on submeshes back full mesh |
| __all__ = ("AdaptiveTransferManager",) | ||
|
|
||
|
|
||
| class AdaptiveTransferManager(TransferManager): |
There was a problem hiding this comment.
If I wanted to use a custom transfer operator (for e.g. some complicated multigrid thing), does this class compose with that?
| @@ -0,0 +1,365 @@ | |||
| """ | |||
There was a problem hiding this comment.
The tests look persuasive.
Do we have any profiled runs of using AdaptiveMeshHierarchy (where all elements are refined) vs MeshHierarchy to see how much the overhead is, and where that overhead is spent?
dham
left a comment
dham
left a comment
There was a problem hiding this comment.
needs a manual section and docstrings need to document where netgen meshes are required.
Description