When (re)assembling a matrix, we can restrict the input for
efficiency/debugging purposes. These options include
|MAT_NEW_NONZERO_LOCATIONS ||- additional insertions will be
allowed if they generate a new nonzero
|MAT_NEW_DIAGONALS ||- new diagonals will be allowed (for block diagonal format only)
|MAT_IGNORE_OFF_PROC_ENTRIES ||- drops off-processor entries
|MAT_NEW_NONZERO_LOCATION_ERR ||- generates an error for new matrix entry
|MAT_USE_HASH_TABLE ||- uses a hash table to speed up matrix assembly
+ MAT_NO_OFF_PROC_ENTRIES - you know each process will only set values for its own rows, will generate an error if
any process sets values for another process. This avoids all reductions in the MatAssembly routines and thus improves
performance for very large process counts.
Some options are relevant only for particular matrix types and
are thus ignored by others. Other options are not supported by
certain matrix types and will generate an error message if set.
If using a Fortran 77 module to compute a matrix, one may need to
use the column-oriented option (or convert to the row-oriented
MAT_NEW_NONZERO_LOCATIONS set to PETSC_FALSE indicates that any add or insertion
that would generate a new entry in the nonzero structure is instead
ignored. Thus, if memory has not alredy been allocated for this particular
data, then the insertion is ignored. For dense matrices, in which
the entire array is allocated, no entries are ever ignored.
Set after the first MatAssemblyEnd()
MAT_NEW_NONZERO_LOCATION_ERR indicates that any add or insertion
that would generate a new entry in the nonzero structure instead produces
an error. (Currently supported for AIJ and BAIJ formats only.)
This is a useful flag when using SAME_NONZERO_PATTERN in calling
KSPSetOperators() to ensure that the nonzero pattern truely does
remain unchanged. Set after the first MatAssemblyEnd()
MAT_NEW_NONZERO_ALLOCATION_ERR indicates that any add or insertion
that would generate a new entry that has not been preallocated will
instead produce an error. (Currently supported for AIJ and BAIJ formats
only.) This is a useful flag when debugging matrix memory preallocation.
MAT_IGNORE_OFF_PROC_ENTRIES indicates entries destined for
other processors should be dropped, rather than stashed.
This is useful if you know that the "owning" processor is also
always generating the correct matrix entries, so that PETSc need
not transfer duplicate entries generated on another processor.
MAT_USE_HASH_TABLE indicates that a hash table be used to improve the
searches during matrix assembly. When this flag is set, the hash table
is created during the first Matrix Assembly. This hash table is
used the next time through, during MatSetVaules()/MatSetVaulesBlocked()
to improve the searching of indices. MAT_NEW_NONZERO_LOCATIONS flag
should be used with MAT_USE_HASH_TABLE flag. This option is currently
supported by MATMPIBAIJ format only.
MAT_KEEP_NONZERO_PATTERN indicates when MatZeroRows() is called the zeroed entries
are kept in the nonzero structure
MAT_IGNORE_ZERO_ENTRIES - for AIJ/IS matrices this will stop zero values from creating
a zero location in the matrix
MAT_USE_INODES - indicates using inode version of the code - works with AIJ and
ROWBS matrix types
MAT_NO_OFF_PROC_ZERO_ROWS - you know each process will only zero its own rows. This avoids all reductions in the
zero row routines and thus improves performance for very large process counts.
MAT_IGNORE_LOWER_TRIANGULAR - For SBAIJ matrices will ignore any insertions you make in the lower triangular
part of the matrix (since they should match the upper triangular part).
Notes: Can only be called after MatSetSizes() and MatSetType() have been set.
Index of all Mat routines
Table of Contents for all manual pages
Index of all manual pages