In this case the parallel index set might represent N entries (this might be the number of
cells of grid). Nevertheless, there several (n) equations/unknowns attached to each index.
In this case we construct a larger index set representing N*n unknows, where each unknown
is attached to an index.
This change only affects parallel runs.
This commit updates the source code comment about using operator[] to
initialize the unordered map. Thanks to Bard's persistence we found
out that the cause is not the construction of the key value of type
std::string from const char* but the mapped type being a (mutable)
char* (due to C?).
This completes the PR #784.
g++-4.4 has problems converting const char* to char*
which it thinks is needed for constructing std::string.
Using operator[] circumvents this problem.
The compiler error fixed here was:
/usr/include/c++/4.4/bits/stl_pair.h: In constructor ‘std::pair<_T1, _T2>::pair(std::pair<_U1, _U2>&&) [with _U1 = const char*, _U2 = const char*, _T1 = const std::basic_string<char, std::char_traits<char>, std::allocator<char> >, _T2 = char*]’:
/home/mblatt/src/dune/opm/opm-core/opm/core/linalg/LinearSolverPetsc.cpp:40: instantiated from here
/usr/include/c++/4.4/bits/stl_pair.h:107: error: invalid conversion from ‘const char*’ to ‘char*’
make[2]: *** [CMakeFiles/opmcore.dir/opm/core/linalg/LinearSolverPetsc.cpp.o] Fehler 1
Well, you never know. There are containers that use a signed integer
for storing its size. This results in a warning about comparing signed with
unsigned integers. This commit prevents this by explicitly casting the size
to std::size_t.
Due to the size of the overlap layer and the discretization scheme
the rhs might not contain correct values for overlap cells. This
commit makes sure they are correct by an additional communication step.
One would think that such an assumption is safe in any case,
wouldn't one? But foen Eigen'S container this does not hold.
They do not provide STL compliant iterators, and access to them.
With this patch make the even stricter assumption that the containers
are random access and use operator[] instead of iterators.
As there are no functors for computing the minimum and maximum,
we convert the std::max and std::min function pointers to
functors (which is not really nice.) Previously we were somehow
tricked into using std::greater and std::less, which of course do
return true or false and not what we need. Additionally, do more
excessive testing with different ranges.
We need to compute quite a few global reductions in the
Newton method of opm-autodiff. This commit adds the functionality
to compute several reductions combined using only one global
communication. Compiles and test succeeds with one or more process.
The error checking macro makes it harder to read and harder to write, so
instead we now only check for functions that can contain errors. Bounds and
range checks are handled by PETSc and not OPM.
The previous implementation set plenty of values in the initialization list and
immediately overwrote these values with values looked up from the param group.
This patch makes it look up the parameteres from the param group argument,
making the constructor simpler.
Petsc only supports initialisation through the ParameterGroup constructor.
Calling the default, non-arg constructor is a static error, and not
implementing it makes using it break compiles.
call_petsc.c was really a thin C wrapper around the call to petsc itself and
turns out was mostly unnecessary C++ emulation. This removes the file entirely
and ports its functionality into LinearSolverPetsc.cpp.
All features from the file should now be more readable as well as properly
utilising modern C++ features.
The patch uses the CHKERRXX macro from petsc to handle errors reported by
petsc, and currently does not handle this and give the control back to OPM's
error/throw system.
The class OwnerOverlapCopyCommunication is not defined unless MPI is
avilable. Therefore, we cannot reference the type unless we know
that MPI is available in the current translation unit.
To support this the solveSystem methods of the LinearSolverInterface gets
an optional additional template parameter of type boost::any. It can hold any
copy constructable object. In our case it is used to pass the information about
the parallelization into the solvers of dune-istl without the compiler needing to know
their type. Inside of LinearSolverIstl::solveSystem we check whether the type stored inside of
boost::any is the new ParallelIstlInformation. If this is the case we extract the information
and use the parallel solvers if available, otherwise we solve serial/sequential.
The new ParallelIstlInformation is needed as the OwnerOverlapCopyCommunication is not copy
constructable. This is indeed a design flaw that should and will fixed upstream, but for the
time being we need ParallelIstlInformation to transfer the ParallelIndexSet and RemoteIndices
objects.
