Merge branch 'bugfix/fix-vbvectortest' into 'master'

ref:core/dune-istl Formerly, this would throw an exception if

DUNE_ISTL_WITH_CHECKING

is enabled since you must not call a create iterator on an initialized
VariableBlockVector.

That being said, maybe it's a good idea to enable these checks in the CI
system.

See merge request [core/dune-istl!214]

  [core/dune-istl!214]: gitlab.dune-project.org/core/dune-istl/merge_requests/214

Merge branch 'feature/suitesparse-cholmod' into 'master'

ref:core/dune-istl Take-over from patrick.jaap:feature/CHOLMOD

See merge request [core/dune-istl!233]

  [core/dune-istl!233]: gitlab.dune-project.org/core/dune-istl/merge_requests/233

Otherwise it cannot be found

Patrick Jaap authored 6 years ago and Christoph Grüninger committed 6 years ago

A simple implementation for the SuiteSparse CHOLMOD solver for real valued
(double) linear systems.
Ignore dof's are supported.
The user has to take care of the correct size of
vectors in the "apply" method.

Merge branch 'make-createiterator-an-output-iterator' into 'master'

ref:core/dune-istl The class VariableBlockVector::CreateIterator behaves
basically like an STL output iterator, but it didn't quite implement the
required interface. This patch adds the missing methods and extends the unit
test.

The main benefit of this (besides interface beauty) is that you can now use
std::fill etc. to set the block sizes of a VariableBlockVector.

See merge request [core/dune-istl!229]

  [core/dune-istl!229]: gitlab.dune-project.org/core/dune-istl/merge_requests/229

Merge branch 'feature/fix-quadratic-umfpacksetup' into 'master'

ref:core/dune-istl Looking up if a column index is contained in submatrix
could degenerate to accumulated O(n) complexity for each row in the old
implementation. By setting up a look-up before, this can be improved to O(1).
Fortunately such a vector is constructed anyway so we can simple use it after
minor reordering of code.

This should fix [#54].

See merge request [core/dune-istl!230]

  [#54]: gitlab.dune-project.org/NoneNone/issues/54
  [core/dune-istl!230]: gitlab.dune-project.org/core/dune-istl/merge_requests/230


Closes #54

Looking up if a column index is contained in submatrix could
degenerate to accumulated O(n) complexity for each row in the
old implementation. By setting up a look-up vector before, this
can be improved to O(1). Fortunately such a vector is constructed
anyway so we can simple use it after minor reordering of code.

Note that we keep the old O(n) implementation of addRowNnz()
because this is still used by some overlapping Schwarz code
that I do not intend to touch myself.

The class VariableBlockVector::CreateIterator behaves basically
like an STL output iterator, but it didn't quite implement the
required interface.  This patch adds the missing methods and
extends the unit test.

The main benefit of this (besides interface beauty) is that you can
now use std::fill etc. to set the block sizes of a VariableBlockVector.

Merge branch 'fix-typos' into 'master'

See merge request [core/dune-istl!228]

  [core/dune-istl!228]: Nonecore/dune-istl/merge_requests/228

Merge branch 'feature/disallow-multithreaded-suitesparse' into 'master'

Suitesparse, as installed with Debian, is thread-parallel using OpenMP. OpenMP
silently assumes, it can spawn as many threads as there are cores. In a worst
case scenario, this leads to a number of threads quadratic in the core count,
if you also do parallel test execution with the maximum number of cores. We
solve the issue by disallowing OpenMP to allocate more than one thread.

See merge request [core/dune-istl!227]

  [core/dune-istl!227]: gitlab.dune-project.org/core/dune-istl/merge_requests/227

Suitesparse, as installed with Debian, is thread-parallel using OpenMP.
OpenMP silently assumes, it can spawn as many threads as there are cores.
In a worst case scenario, this leads to a number of threads quadratic in
the core count, if you also do parallel test execution with the maximum
number of cores. We solve the issue by disallowing OpenMP to allocate more
than one thread.

Merge branch 'bugfix/fcg-initial-direction' into 'master'

This adds a missing reset of the vector for the initial direction in the first
iteration of the apply method.

In the current state the vector for the search direction contains the initial
solution in the first iteration. Most of the standard preconditioners will add
their result to the vector instead of replacing it. If the initial solution is
not zero, this results in an incorrect first search direction.

See merge request [core/dune-istl!226]

  [core/dune-istl!226]: gitlab.dune-project.org/core/dune-istl/merge_requests/226

Merge branch 'no-needless-simd' into 'master'

This simplifies io.

See merge request [core/dune-istl!224]

  [core/dune-istl!224]: gitlab.dune-project.org/core/dune-istl/merge_requests/224

This simplifies io.

Merge branch 'matrixtest-1x1-interface' into 'master'

Found by clang 6 -Wunused-variable.

No idea why the FieldMatrix is tested in ISTL, but whatever.

See merge request [core/dune-istl!222]

  [core/dune-istl!222]: gitlab.dune-project.org/core/dune-istl/merge_requests/222

Merge branch 'simd-io' into 'master'

Depends: [core/dune-common!551]

Addresses: [#50]

See merge request [core/dune-istl!223]

  [core/dune-common!551]: gitlab.dune-project.org/core/dune-common/merge_requests/551
  [#50]: gitlab.dune-project.org/core/dune-istl/issues/50
  [core/dune-istl!223]: gitlab.dune-project.org/core/dune-istl/merge_requests/223

Depends: core/dune-common!551

Addresses: #50

Found by clang 6 -Wunused-variable.

No idea why the FieldMatrix is tested in ISTL, but whatever.

Merge branch 'fix-tlime-residual-limit' into 'master'

The previous convergence limit was originally determined experimentally as
1e-11. This worked for many blas implementations and architectures. However,
when used with openblas on skylake, apparently the residual norm would not go
below ~1e-10, so convergence was never achieved. In fact, even on non-skylake
the residual norm would go above 1e-11 again after briefly dipping below, if
iterating further.

We believe that this is due to openblas selecting -- at runtime -- some
skylake specific algorithm leading to a different ordering of operations, in
turn leading to differences in numerical cancellation. We have however not
verified this conclusively, nor have we identified precisely which blas
algorithm is causing this.

This patch raises the convergence limit to
`sqrt(numeric_limits<field_type>::epsilon())`. This limit has no theoretical
justification -- it was selected because it usually works as a convergence
limit for other (completely unrelated) algorithms, and because it works for
both Skylake and other architectures (AMD Epyc) in this particular case.

Developed together with Sebastian Westerheide.

Fixes: \#48.


See merge request core/dune-istl!221

The previous convergence limit was originally determined experimentally
as 1e-11.  This worked for many blas implementations and architectures.
However, when used with openblas on skylake, apparently the residual norm
would not go below ~1e-10, so convergence was never achieved.  In fact, even
on non-skylake the residual norm would go above 1e-11 again after briefly
dipping below, if iterating further.

We believe that this is due to openblas selecting -- at runtime -- some
skylake specific algorithm leading to a different ordering of operations, in
turn leading to differences in numerical cancellation.  We have however not
verified this conclusively, nor have we identified precisely which blas
algorithm is causing this.

This patch raises the convergence limit to
`sqrt(numeric_limits<field_type>::epsilon())`.  This limit has no theoretical
justification -- it was selected because it usually works as a convergence
limit for other (completely unrelated) algorithms, and because it works for
both Skylake and other architectures (AMD Epyc) in this particular case.

Developed together with Sebastian Westerheide.

Fixes: #48.

[multirhstest] allow access to measured timestamps in multirhstest

See merge request core/dune-istl!219

[Clang] Fix warning about an unchecked enumerator in switch().

See merge request core/dune-istl!218

```
/home/joe/Projekte/dune-simd/dune-istl/dune/istl/matrixmarket.hh:706:18: warning: enumeration values 'general' and 'unknown_structure' not handled in switch [-Wswitch]
          switch(mmHeader.structure)
```

- Move the `switch(mmHeader.structure)` out of the loop over the entries and
  and check all structure types there.  This will throw immediately for
  unknown structure types, rather then reading all entries and then throwing.
- Write the `for()` loop in a less unusual format; i.e. don't try and save on
  local integer variables.

Mark scope guard as DUNE_UNUSED

See merge request core/dune-istl!217

This helps to avoid a clang warning

Allow autoCopy() for blocks of VariableBlockVector

Closes #49

See merge request core/dune-istl!215

fix compile error and several warnings with gcc 8

See merge request core/dune-istl!216

This addresses several compiler warnings from GCC 8 of the form

    dune/istl/repartition.hh:968:22: warning:
    unused variable ‘wgtflag’ [-Wunused-variable]

and

    dune/istl/repartition.hh:976:13: warning:
    variable ‘options’ set but not used [-Wunused-but-set-variable]