More speculative build fixes

Fix some Jenkins errors with careful compilers

Merge branch 'master' into knepley/feature-snes-patch

Small fix

DM: Reorganized DM fields
- A field is defined by a discretization (PetscFE or PetscFV) and a region (DMLabel or NULL)
- This will feed into the creation of PetscDS for every different cell system
-

DM: Reorganized DM fields
- A field is defined by a discretization (PetscFE or PetscFV) and a region (DMLabel or NULL)
- This will feed into the creation of PetscDS for every different cell system
- Changed interface for Get/SetField to add the region

show more ...

Merge branch 'barry/improve-error-propagation-ksp-pc'

Improve the error propagation between and within KSPs and PCs

- Added KSPCheckSolve() to be called consistently by PCs to set failures if the inner solvers fail.
- Added use of KSPCheckNorm() and KS

Improve the error propagation between and within KSPs and PCs

- Added KSPCheckSolve() to be called consistently by PCs to set failures if the inner solvers fail.
- Added use of KSPCheckNorm() and KSPCheckDot() to all the Krylov methods to propagate errors from PCs to the KSPs.

Commit-type: feature

show more ...

KSPs that are created and used by PCs or other KSPs should inherit
the error if not converged flag

Commit-type: bug-fix

Fix a subtle bug with multiplicative and user patches.

Restore the closure and star in the right place in the loop.

Line relaxation for advection-diffusion doesn't work without this.

Fix uninitialised memory error in valgrid.

Create the IS _after_ we set the contents of the indices.
(Since the IS owns the pointer, this shouldn't have any
functional effect, but just keeps valgrin

Fix uninitialised memory error in valgrid.

Create the IS _after_ we set the contents of the indices.
(Since the IS owns the pointer, this shouldn't have any
functional effect, but just keeps valgrind happy.)

show more ...

Add a pc_patch_pardecomp_overlap option. Currently applies closure of star
the specified number of times.

Rename PC_PATCH_OWNED to PC_PATCH_PARDECOMP.

Still not entirely settled on the name. However, "owned" is misleading
as it will soon incorporate overlap, so it's not just "owned". Since
it's construc

Rename PC_PATCH_OWNED to PC_PATCH_PARDECOMP.

Still not entirely settled on the name. However, "owned" is misleading
as it will soon incorporate overlap, so it's not just "owned". Since
it's constructing the domain decomposition from the solver from the
parallel decomposition, I want to get "parallel" in there somewhere.
-pc_patch_construct_type paralleldecomposition seems a mouthful,
so I thought let's go for -pc_patch_construct_type pardecomp for now.

show more ...

If the patch size is too large, use a MATPREALLOCATOR instead of a bitarray.

Sketch an "owned" patch constructor.

This has several uses. First of all, in serial it allows us to verify
the implementation of SNESPATCH (by comparing it to Newton/LU or
whatever). Second, this be

Sketch an "owned" patch constructor.

This has several uses. First of all, in serial it allows us to verify
the implementation of SNESPATCH (by comparing it to Newton/LU or
whatever). Second, this begins to address the functionality offered by
PCASM/SNESNASM, but in a more general way.

At the moment, in the patch code we compute the sparsity pattern of the
matrices using a PetscBT. This doesn't work for moderately sized
patches, as the comment in PCPatchCreateMatrix_Private indicates. This
should be fixed.

The convergence of this might be greatly improved by keeping some
degrees of freedom in the overlap; we should add options to allow the
user to control this, also. (This might entail changing the name.)

show more ...

When solving a nonlinear problem, we need _all_ the information
about the state on the patch, including the values of dofs that
feature in global (as well as artificial) boundary conditions.
So we mu

When solving a nonlinear problem, we need _all_ the information
about the state on the patch, including the values of dofs that
feature in global (as well as artificial) boundary conditions.
So we must build this function space on each patch, too, and
the scatters from the smallest patch to this one.

show more ...

Store a Boolean for whether we're PCPATCH or SNESPATCH

SNESPATCH: Do the same thing for the residual evaluation

We need to pass the state _with_ boundary data to the patch residual
evaluator, but compute a residual _without_ boundary data. We therefore
need all of the *withArtificial datastructures in PCPATCH

We need to pass the state _with_ boundary data to the patch residual
evaluator, but compute a residual _without_ boundary data. We therefore
need all of the *withArtificial datastructures in PCPATCH if we're
running in nonlinear mode. This requires passing two dofmaps to
the residual code, also.

show more ...

Start refactoring for multiplicative updates. Multiplicative + SNESPATCH
doesn't actually work yet because we need to pass the state _with BC values_
to the residual evaluation, but at the moment we'

Start refactoring for multiplicative updates. Multiplicative + SNESPATCH
doesn't actually work yet because we need to pass the state _with BC values_
to the residual evaluation, but at the moment we're only passing the
state _for the dofs we're to solve for_.

show more ...

Look for "pc_patch_construct_type" in PCPATCH, and
"snes_patch_construct_type" in SNESPATCH.

A PC only deals with two vectors, an RHS and an update.
A SNES deals with 4: residual calculation workspace, state vector, state update vector, optional RHS vector.

Rename things in PCPATCH to make

A PC only deals with two vectors, an RHS and an update.
A SNES deals with 4: residual calculation workspace, state vector, state update vector, optional RHS vector.

Rename things in PCPATCH to make clear what is RHS and what is update.
Scatter the global state onto the patch solve before calling SNESSolve.
Give each SNES its own separate residual calculation workspace.

show more ...

Small merge fix

Merge branch 'master' into knepley/feature-snes-patch

Not sure if this has done the right thing because there are no tests
yet. Also, the fortran stubs break (it doesn't know what to do with
SNESCo

Merge branch 'master' into knepley/feature-snes-patch

Not sure if this has done the right thing because there are no tests
yet. Also, the fortran stubs break (it doesn't know what to do with
SNESCompositeType), and I don't know how to fix it.

show more ...

Merge branch 'pr1235/florianwechsung/pcpatch-partition-of-unity/master' [PR #1235]

* pr1235/florianwechsung/pcpatch-partition-of-unity/master:
PCPATCH: Parallel partition of unity and multiplicative

Merge branch 'pr1235/florianwechsung/pcpatch-partition-of-unity/master' [PR #1235]

* pr1235/florianwechsung/pcpatch-partition-of-unity/master:
PCPATCH: Parallel partition of unity and multiplicative
This merge adds:
- partition of unity in parallel: we just have to add a scatter of the vector that counts for each dof in how many patches it appears in
- multiplicative: whenever a patch solve has been done, we can figure out how this changes the residual by multiplying with the patch matrix without artificial bcs applied.
Since we use static condensation for the bcs, this means that we now assemble a second matrix on each patch which has only got the global bcs applied to it.

show more ...