WIP: Panzer Owned and Ghosted GlobalEvaluationData Changes

[jmgate]

@eric-c-cyr and @rppawlo, we already did a code review of this refactor a while back, but changes pushed to Drekar required a refactor of the refactor (yay!).

The gist of it is Panzer's GlobalEvaluationData classes now have operator[]() and at() methods which take a local ID and give you the appropriate element of either the owned or ghosted vector. This way the user (in Panzer, Drekar, or elsewhere) doesn't need to know anything about the data being split out under the hood.

Should we ditch the at() routines? I included them for the sake of completeness and to mimmic the functionality of STL containers, but perhaps the bounds checking should happen in operator[]() surrounded by some #ifdef. Or maybe it already happens if -D Trilinos_ENABLE_DEBUG since they're Epetra_Vectors and Thyra::VectorBases behind the scenes—is that right?

In order to build and test this PR, you'll also need to pull the panzerGatherScatter branch from DrekarBase.

Note that I'll clean up the commit history of this branch before the merge to develop. This is just so we can talk it over.

[jmgate]

Including @ssmabuza on the discussion. He requests:

Is it possible to have the files with "Panzer_BasisIRLayout" modified to include this additional public data member:

Teuchos::RCP<PHX::DataLayout> local_mat_layout =
  Teuchos::rcp(new PHX::MDALayout<panzer::Cell, panzer::BASIS, panzer::BASIS>(
  this->numCells(), this->cardinality(), this->cardinality()));

This helps with creating local element matrices.

Sibu, can we talk this through in the meeting I set up for next Wednesday? I'm afraid I don't see right off the bat how this relates to this PR.

[eric-c-cyr]

I would say add it to PureBasis, though I also don't see what it has to do with the PR.

[eric-c-cyr]

@ssmabuza Do you have a patch for such a thing (a particularly design and set of members in mind?) This would not be hard to add.

[ssmabuza]

@eric-c-cyr This is to provide the layout for PHX::MDField local matrix-like objects in the evaluators. Currently, we only have functional, basis and a few other layouts. At the moment I am bypassing this issue by creating in the constructor a pointer to PHX::DataLayout. local_mat_layout can be added to BasisIRLayout and in panzer::BasisIRLayout::setup(const panzer::PointRule & point_rule) we can simply add the line I sent to you above.

I'll send you a patch shortly.

[jmgate]

Outcomes from yesterday's code review:

• Move forward declarations into the panzer namespace—check that they're even needed.
• Remove at() routines.
• Move operator[] routines and associated data from derived classes into base classes to avoid virtual function calls.

[jmgate]

@eric-c-cyr, @rppawlo: It looks like we have forward declarations outside the panzer namespace in more than just the code I've touched. Should I create an issue to deal with all those?

[jmgate]

The issue with forward declarations will be handled for all of Panzer through #1548.

[jmgate]

at() and associated routines removed.

[jmgate]

Moved operator[]() routines and ownedView_ and ghostedView_ data to base classes. Testing now.

[jmgate]

Tests pass. Rebasing on latest Trlinos/develop and DrekarBase/master.