Introduce PCFactorSetDefaultOrdering_Factor() to allow each factorization package
to indicate its preferered ordering, which may be external (its own) or one of PETSc.

This also affects factorizatio

Introduce PCFactorSetDefaultOrdering_Factor() to allow each factorization package
to indicate its preferered ordering, which may be external (its own) or one of PETSc.

This also affects factorization like dense PETsc where the ordering passed in by PETSc is ignored and not needed.

Also fixed options for MUMPS and SuiteSparse where users needed to pass a particular options database option
to the package (MUMP or KLU) indicating they should use the PETSc provided ordering. Now if the PETSc provided ordering is provided
these packages automatically use them without need a special option. This simplifies ordering usage for external packages

Commit-type: optimization, bug-fix, testing-fix, style-fix
/spend 6h

show more ...

Change MatFactorGetUseOrdering() to MatFactorGetCanUseOrdering()

We need to distinquish whether the package can use an external ordering and if that is desirable (the internal ordering may be faster

Change MatFactorGetUseOrdering() to MatFactorGetCanUseOrdering()

We need to distinquish whether the package can use an external ordering and if that is desirable (the internal ordering may be faster).

Commit-type: style-fix
/spend 15m

show more ...

Add cuSparse Band LU factorization

Rename -mat_cusparse_transgen to -mat_form_explicit_transpose

MATSEQAIJCUSPARSE: if MatSetFromOptions is not called these defaults are never set!

Add non-null PetscDefault{Cuda,Hip}Stream

MATSEQAIJCUSPARSE: better handling of explicit transposition

keep track when the transposed matrix needs to be updated
call csr2csc only once, use thrust calls otherwise

MatAXPY_SeqAIJCUSPARSE: CUDA routines are buggy with 1 column!

MATSEQAIJCUSPARSE: add MatScale support

Report resource errors from CUDA, cuBLAS, and cuSparse as PETSC_ERR_GPU_RESOURCE so test harness can retry them.

Convert the error code to a GPU resource error code so that the test harness will rer

Report resource errors from CUDA, cuBLAS, and cuSparse as PETSC_ERR_GPU_RESOURCE so test harness can retry them.

Convert the error code to a GPU resource error code so that the test harness will rerun the example multiple times.

This should decrease the number of false job failures for CUDA runs

/spend 1.5h

show more ...

MATSEQAIJCUSPARSE: add PetscInfo diagnostics for buffer size in MatMat operations

MatMatBackend: use device memory when populating the COO buffers

MatAXPY: add support for CUSPARSE

MatMatMPIAIJBACKEND: add few more options

do not copy CPU values down to GPU if not present!

MatGetLocalMatMerge: fix issues with thrust on SUMMIT

I suspect partition_copy is buggy

Mat: Move COO events out of CUSPARSE class

Fix a few typos in the code

MatSeqAIJCUSPARSEMergeMats: fix for CUDA < 10

Apply 7 suggestion(s) to 3 file(s)

MATSEQAIJCUSPARSE: remove copy to the CPU after GPU operations

MatAIJCUSPARSESetGenerateTranspose: convenience function for seq and mpi

PetscStackCallThrust

MATSEQAIJCUSPARSE: factors need to update the CPU memory

MatSetValuesCOO: use cuda memory if possible

MatMPIAIJGetLocalMatMerge: support for merging two matrices

MatSetValuesCOO: perform addition of repeated entries when INSERT_VALUES is specified

use temporary data instead of persistent storage in cusparse implementation