Add inter-grid transfer operators #10417

peterrum · 2020-05-29T12:03:44Z

No description provided.

doc/news/changes/minor/20200529MunchPrieto

tjhei

Would you mind describing the new classes on a high level? I got lost when looking at the review.

I left some initial comments.

include/deal.II/multigrid/mg_transfer_interface.h

tjhei · 2020-06-12T12:45:38Z

include/deal.II/multigrid/mg_transfer_interface.h

+
+
+/**
+ * Class for repartitioning data living on the same triangulation, which


triangulation or dofhandler? Or is this more generic? Are these vectors the same size?

This is probably unrelated to multigrid and could go somewhere else (vectortools?).

This is probably unrelated to multigrid and could go somewhere else (vectortools?).

But for now I would like to keep in this PR since it uses the same infrastructure as the global-coarsening algorithm is using.

include/deal.II/multigrid/mg_transfer_interface.h

include/deal.II/multigrid/mg_transfer_interface.templates.h

peterrum · 2020-08-24T19:53:05Z

@tjhei Sorry for the late reply! The idea is following.

Let's consider for now two levels (the level with more dofs I will call fine level; the other one is the coarse level). The class MGTwoLevelTransfer is responsible for transfer between these two levels. However, within levels we might need to distinguish how to perform element-wise prolongate/restrict:

hp-context: restriction/prolongation between element of different polynomial degrees
global coarsening: restriction/prolongation to children or application of identity matrix (if coarse cell is not refined)

This information is encrypted in the class MGTransferScheme. MGTwoLevelTransfer practically loops over all MGTransferSchemes.

The class MGTransferMatrixFreeNew (or however it will be called) is a collection of MGTwoLevelTransfers.

In the namespace MGTransferUtilities, I have collected functions that set up MGTwoLevelTransfer for hp-context and global coarsening.

include/deal.II/multigrid/mg_transfer_interface.templates.h

tests/multigrid-transfer/mg_transfer_a_01.cc

tjhei · 2020-09-18T15:06:57Z

include/deal.II/multigrid/mg_transfer_interface.h

+ * elements are currently not implemented.
+ */
+template <typename MatrixType>
+class MGTransferMatrixFreeNew


maybe MGTransferGlobalCoarsening?

tjhei · 2020-09-18T15:07:59Z

Maybe move VectorRepartitioner into separate PR?
What about custom constraints?

peterrum · 2020-09-19T11:55:20Z

maybe MGTransferGlobalCoarsening?

Done!

Maybe move VectorRepartitioner into separate PR?

I have removed it. Once this PR is merged, I will open the PR for that.

What about custom constraints?

Could you have a look at the utility functions. There users have the possibility to pass constraint matrices. In what way are these different from custom constraints?

peterrum · 2020-09-19T12:06:21Z

/rebuild

include/deal.II/multigrid/mg_transfer_interface.h

tjhei · 2020-09-20T16:04:31Z

include/deal.II/multigrid/mg_transfer_interface.h

+ *   MGTransferUtilities to setup the Transfer operators.
+ */
+template <typename Number>
+struct MGTransferScheme


If I understand it correctly, then this is an implementation detail of TwoLevelTransfer, right? I would make this a member of it or move it into an implementation namespace.

include/deal.II/multigrid/mg_transfer_interface.h

tjhei · 2020-09-20T16:06:12Z

include/deal.II/multigrid/mg_transfer_interface.h

+ * Class for transfer between two multigrid levels.
+ */
+template <int dim, typename Number>
+class MGTwoLevelTransfer


I would suggest templating on VectorType and only implementing d:Vector for now.

tjhei · 2020-09-20T16:06:33Z

include/deal.II/multigrid/mg_transfer_interface.h

+
+
+/**
+ * Class for transfer between two multigrid levels.


can you extend to say that this can be p or global coarsening?

tjhei · 2020-09-20T16:07:19Z

include/deal.II/multigrid/mg_transfer_interface.h

+ * elements are currently not implemented.
+ */
+template <typename MatrixType>
+class MGTransferMatrixFreeNew


consider renaming to GlobalCoarsening

tjhei · 2020-09-20T16:09:11Z

include/deal.II/multigrid/mg_transfer_interface.h

+  : matrices(matrices)
+  , transfer(transfer)
+{
+  AssertDimension(matrices.max_level() - matrices.min_level(),


the documentation of this function talks about l and l-1 entries, but here you assert that they are equal. Fix, please.

tjhei · 2020-09-20T17:26:39Z

include/deal.II/multigrid/mg_transfer_interface.templates.h

+    bool
+    run(Fu &fu)
+    {
+      return do_run_degree<1>(fu) || do_run_degree<2>(fu) ||


what about degree=0?

We can think about degree=0. But let's leave that for a follow-up PR.

tjhei · 2020-09-20T17:29:14Z

include/deal.II/multigrid/mg_transfer_interface_util.h

+   *   2) bisect, and 3) decrease-by-one is supported.
+   */
+  bool
+  polynomial_transfer_supported(const unsigned int fe_degree_fine,


is this just FE_Q or more generic?

Currently FE_Q and FE_DGQ (and I guess FE_DGQArbitraryNodes). I guess it should be the same as in the case MGTransferMatrixFree.

peterrum · 2020-09-20T20:45:05Z

@tjhei I have made the changes.

tjhei · 2020-09-21T00:35:17Z

include/deal.II/multigrid/mg_transfer_global_coarsening.h

+
+
+/**
+ * Implementation of the MGTransferBase interface for which the transfer


this documentation is outdated. Please mention p and h global coarsening.

tjhei · 2020-09-21T00:38:00Z

include/deal.II/multigrid/mg_transfer_global_coarsening_util.h

+
+DEAL_II_NAMESPACE_OPEN
+
+namespace MGTransferUtilities


Please document the namespace. Also, the name is not ideal, is it?

tjhei · 2020-09-21T00:38:42Z

include/deal.II/multigrid/mg_transfer_global_coarsening_util.h

+   *   2) bisect, and 3) decrease-by-one is supported.
+   */
+  bool
+  polynomial_transfer_supported(const unsigned int fe_degree_fine,


Based on the documentation, I have no idea what this function does. Would this be more appropriate as a static member of MGTransferTwoLevel? I assume that this is what the function reports on.

tjhei · 2020-09-21T00:38:59Z

include/deal.II/multigrid/mg_transfer_global_coarsening_util.h

+   */
+  template <int dim, typename Number, typename MeshType>
+  void
+  setup_global_coarsening_transfer(


Make this a constructor of MGTwoLevelTransfer?

tjhei · 2020-09-21T00:39:05Z

include/deal.II/multigrid/mg_transfer_global_coarsening_util.h

+   */
+  template <int dim, typename Number, typename MeshType>
+  void
+  setup_polynomial_transfer(


tjhei · 2020-09-21T00:39:24Z

include/deal.II/multigrid/mg_transfer_global_coarsening_util.h

+    MGTwoLevelTransfer<dim, LinearAlgebra::distributed::Vector<Number>>
+      &transfer);
+
+} // namespace MGTransferUtilities


With all three functions gone, we could get rid of this namespace and file.

peterrum · 2020-09-21T06:32:37Z

@tjhei I have following changes:

remove utility namspace and add the functions to MGTwoLevelTransfer (with new names reinit_geometric_transfer() and reinit_polynomial_transfer())
remove old files
remove the template argument MeshType (since hp::DoFHandler is now part of DoFHandler)

Le's not over-think the structure of the internal functions. My guess is once we want to support additional vector types, we can reuse many functions.

tjhei

Thanks for the quicj edits. Things look much better now. I have more complaints though, sorry.

tjhei · 2020-09-21T12:28:02Z

include/deal.II/multigrid/mg_transfer_global_coarsening.h

+   * scheme for transfer between children and parent cells, as well as, one
+   * transfer scheme for cells that are not refined).
+   */
+  struct MGTransferScheme


does this need to be public?

tjhei · 2020-09-21T12:31:09Z

tests/multigrid-transfer/global_id_01.cc

+// ---------------------------------------------------------------------
+
+
+// Test ConsensusAlgorithms::AnonymousProcess.


tjhei · 2020-09-21T12:32:46Z

tests/multigrid-transfer/mg_transfer_a_01.cc

+
+
+/**
+ *


"Test MGTwoLevelTransfer with mesh coarsening"

tjhei · 2020-09-21T12:33:55Z

tests/multigrid-transfer/mg_transfer_a_01.cc

+  deallog.precision(8);
+
+  for (unsigned int i = 1; i < 5; i++)
+    test<2, double>(i);


The test output seems to be really large (github doesn't even show it). Can you see if this can be reduced somewhat?

tjhei · 2020-09-21T12:34:26Z

tests/multigrid-transfer/mg_transfer_p_01.cc

+
+
+/**
+ *


Test MGTwoLevelTransfer with polynomial coarsening

tjhei · 2020-09-21T12:34:45Z

tests/multigrid-transfer/mg_transfer_p_01.cc

+         fe_degree_coarse++)
+      if (MGTwoLevelTransfer<2, LinearAlgebra::distributed::Vector<double>>::
+            polynomial_transfer_supported(fe_degree_fine, fe_degree_coarse))
+        test<2, double>(fe_degree_fine, fe_degree_coarse);


same here, the files seem to be quite big?

tjhei · 2020-09-21T12:35:01Z

tests/multigrid-transfer/mg_transfer_p_02.cc

+
+
+/**
+ *


same here, add a one line description

tjhei · 2020-09-21T12:36:39Z

tests/multigrid-transfer/mg_transfer_support_01.output

+DEAL::1   1   1 1                                     
+DEAL::1   1     1 1                                   
+DEAL::1     1     1 1                                 
+DEAL::1     1       1 1                               


tjhei · 2020-09-21T12:39:03Z

tests/multigrid-transfer/test.h

@@ -0,0 +1,167 @@
+// ---------------------------------------------------------------------


I just realized that this is a new folder. Don't you need to add a CMakeList.txt? Also, maybe we should name this folder multigrid-global-coarsening ?

peterrum · 2020-09-21T12:59:17Z

@tjhei I have addressed your comments. Once you are happy, I will squash.

tjhei · 2020-09-21T13:18:48Z

Looks good, please squash.

peterrum · 2020-09-21T13:33:36Z

@tjhei Done!

tjhei

This looks good to me, but I will let somebody else to take a look. @kronbichler ?

kronbichler

Apart from a few comments, this looks good to me.

include/deal.II/multigrid/mg_transfer_global_coarsening.h

kronbichler · 2020-09-22T07:45:33Z

include/deal.II/multigrid/mg_transfer_global_coarsening.templates.h

+      return do_run_degree<1>(fu) || do_run_degree<2>(fu) ||
+             do_run_degree<3>(fu) || do_run_degree<4>(fu) ||
+             do_run_degree<5>(fu) || do_run_degree<6>(fu) ||
+             do_run_degree<7>(fu) || do_run_degree<8>(fu) ||
+             do_run_degree<9>(fu);


I know that other cases are supposed to happen later; my question would be if that follow-up code would then go to a generic version with -1 template argument (or 0 in the tensor product evaluators) so we do not bail out.

I see your point. Let's leave it to a follow-up PR.

include/deal.II/multigrid/mg_transfer_global_coarsening.templates.h

kronbichler · 2020-09-22T07:48:12Z

include/deal.II/multigrid/mg_transfer_global_coarsening.templates.h

+        AssertDimension(dof_handler_fine.get_fe(0).n_base_elements(), 1);
+        std::string fe_name =
+          dof_handler_fine.get_fe(0).base_element(0).get_name();
+        {
+          const std::size_t template_starts = fe_name.find_first_of('<');
+          Assert(fe_name[template_starts + 1] ==
+                   (dim == 1 ? '1' : (dim == 2 ? '2' : '3')),
+                 ExcInternalError());
+          fe_name[template_starts + 1] = '1';
+        }
+        const std::unique_ptr<FiniteElement<1>> fe(
+          FETools::get_fe_by_name<1, 1>(fe_name));
+
+        transfer.schemes[0].prolongation_matrix_1d.resize(fe->dofs_per_cell *
+                                                          fe->dofs_per_cell);
+
+        for (unsigned int i = 0; i < fe->dofs_per_cell; i++)
+          transfer.schemes[0]
+            .prolongation_matrix_1d[i + i * fe->dofs_per_cell] = Number(1.0);


This is a candidate for #9544.

kronbichler · 2020-09-22T07:49:34Z

include/deal.II/multigrid/mg_transfer_global_coarsening.templates.h

+          const auto ierr = cell_transfer.run(cell_prolongator);
+          (void)ierr;
+          Assert(ierr,
+                 ExcMessage("Prolongation " +
+                            std::to_string(scheme.degree_coarse) + " -> " +
+                            std::to_string(scheme.degree_fine) +
+                            " not instantiated!"));


This is the place I was referring to in my comment above about the default case degree=-1.

peterrum · 2020-09-22T08:49:00Z

@kronbichler I have addressed your comments!

peterrum force-pushed the mg-transfer branch 2 times, most recently from 451e863 to e16d7a8 Compare May 29, 2020 12:28

peterrum added Multigrid Parallel distributed hp adaptivity labels May 31, 2020

tjhei self-requested a review June 1, 2020 00:21

kronbichler reviewed Jun 3, 2020

View reviewed changes

doc/news/changes/minor/20200529MunchPrieto Outdated Show resolved Hide resolved

tjhei reviewed Jun 12, 2020

View reviewed changes

peterrum mentioned this pull request Aug 24, 2020

Add some brackets #10841

Merged

peterrum mentioned this pull request Sep 16, 2020

Refactor access to FEEvaluationImpl and friends #10922

Merged

tjhei reviewed Sep 18, 2020

View reviewed changes

peterrum force-pushed the mg-transfer branch from 4f7b7e9 to 974ad4f Compare September 19, 2020 11:52

peterrum added the ready to test label Sep 19, 2020

peterrum force-pushed the mg-transfer branch from 974ad4f to c9eacda Compare September 19, 2020 12:45

tjhei reviewed Sep 20, 2020

View reviewed changes

tjhei reviewed Sep 21, 2020

View reviewed changes

peterrum force-pushed the mg-transfer branch from e92ef3d to bf26fad Compare September 21, 2020 13:33

tjhei approved these changes Sep 21, 2020

View reviewed changes

kronbichler approved these changes Sep 22, 2020

View reviewed changes

peterrum force-pushed the mg-transfer branch from bf26fad to 7e4f149 Compare September 22, 2020 08:44

Add inter-grid transfer operators

5118bbc

peterrum force-pushed the mg-transfer branch from 7e4f149 to 5118bbc Compare September 22, 2020 08:46

kronbichler approved these changes Sep 22, 2020

View reviewed changes

kronbichler merged commit 704a858 into dealii:master Sep 22, 2020

peterrum mentioned this pull request Sep 24, 2020

Enable MGTwoLevelTransfer for non-precompiled p-coarsening sequences #10966

Merged



		/**
		* Class for repartitioning data living on the same triangulation, which



		/**
		* Implementation of the MGTransferBase interface for which the transfer

		// ---------------------------------------------------------------------


		// Test ConsensusAlgorithms::AnonymousProcess.

		@@ -0,0 +1,167 @@
		// ---------------------------------------------------------------------

Add inter-grid transfer operators #10417

Add inter-grid transfer operators #10417

Conversation

peterrum commented May 29, 2020

tjhei left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

peterrum commented Aug 24, 2020

Choose a reason for hiding this comment

tjhei commented Sep 18, 2020

peterrum commented Sep 19, 2020

peterrum commented Sep 19, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

peterrum commented Sep 20, 2020

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

peterrum commented Sep 21, 2020

tjhei left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

peterrum commented Sep 21, 2020

tjhei commented Sep 21, 2020

peterrum commented Sep 21, 2020

tjhei left a comment

Choose a reason for hiding this comment

kronbichler left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

peterrum commented Sep 22, 2020