We introduced two runtime parameters for this: ilu_redblack and
ilu_reorder_spheres. If the last one is false, we try to preserve
the ordering within in the colors. Otherwise we try to achieve a D2
(alternative diagonal) ordering.
These versions are inspired by the ones used in SuperLU and the enums
to choose them have simuilar names, but without leading S (MILU_1-MILU_3).
The following variants are supported (chosen by the enum MILU_VARIANT):
ILU: plain ILU
MILU_1: lump diagonal with dropped row entries.
MILU_2: lump diagonal with the sum of the absolute values of the dropped row
entries.
MILU_3: if diagonal is positive add sum of dropped row entrires. Otherwise substract them.
MILU_4: if diagonal is positive add sum of dropped row entrires. Otherwise do nothing
These versions are inspired by the ones used in SuperLU and the enums
to choose them have simuilar names, but without leading S (MILU_1-MILU_3).
The following variants are supported (chosen by the enum MILU_VARIANT):
ILU: plain ILU
MILU_1: lump diagonal with dropped row entries.
MILU_2: lump diagonal with the sum of the absolute values of the dropped row
entries.
MILU_3: if diagonal is positive add sum of dropped row entrires. Otherwise substract them.
MILU_4: if diagonal is positive add sum of dropped row entrires. Otherwise do nothing
We did add the dropped matrix blocks to the diagonal blocks,
but this is not MILU where oone only modifies the diagonal.
With this patch we fix this behaviour and now only modify the
diagonal of the diagonal block.
Using the parameter ilu_milu=true|false (default=false) the user can now choose
to use the modified ILU0 decomposition. If selected values will
not be dropped for nonzero entries but added to the diagonal of U.
This approach will result in A*e = L*U*e for vector e with all entries
beging 1.
Currently we run into issues with the parallel AMG if
redistribution happened on the coarsest level. That is
not detected by AMG and it will construct smoothers on
all processors present before the redistribution. Some of
them will get OwnerOverlapCommunication objects that have
an invalid MPI communicator in them (MPI_COMM_NULL) and an
MPI_ERROR will be raised as we communicate in the constructor.
With this patch we detect this situaton and set the pointer to
OwnerOverlapCommunication to null to prevent communication. A
sanity check that the matrix has zero rows has been added.
This seems to have been forgotten previously. Now the code int CPRPreconditioner.hpp
uses ParallelOverlappingILU0 instead of SeqILU[0n]/BlockPreconditioner which
makes the code more slim.
The versions are missing the specialized code for inverting
a 3x3 matrix that makes the algorithms quite a bit more stable.
With this patch we fall back to using our own MatrixBlock that does
not suffer from this deficiency.
