Rename ELL to SELL

SELL is more accurate. It stands for sliced ELLPACK.

Funded-by:
Project: PETSc for Theta
Time: 2.5 hours
Reported-by:
Thanks-to:

Merge branch 'maint'

Merge remote-tracking branch 'origin/balay/fix-mkl-sparse/maint' into maint

mkl: add mkl_sparse.py as a separate package [as the interface code is organized this way]
and use PETSC_HAVE_MKL_SPARSE to enable/disable this interface code [so that this flag
represents functional

mkl: add mkl_sparse.py as a separate package [as the interface code is organized this way]
and use PETSC_HAVE_MKL_SPARSE to enable/disable this interface code [so that this flag
represents functionality that is more than PETSC_HAVE_MKL flag].

Also disable mkl_sparse, mkl_sparse_optimize for 64-bit-incides

Reported-by: Victor Eijkhout <eijkhout@tacc.utexas.edu>

show more ...

MatInvertBlockDiagonalMat: error checking and typo fix

Merge remote-tracking branch 'origin/hongzh/add-ell-format'

MatResetPreallocation: Reset mat to use the original nonzero pattern provided by users.

MPIELL seems to work

Funded-by:
Project: PETSc for Theta
Time: 14 hours
Reported-by:
Thanks-to:

Add seqell format and its matmul kernel

-Make memkind error out when HBW is not usable

Funded-by:
Project: PETSc for Theta
Time: 52 hours
Reported-by:
Thanks-to:

Merge remote-tracking branch 'origin/balay/vs6-warning-fixes'

build: fix compile warnings with VS6

Mat: MatInvertBlockDiagonalMat. KSP: Block diagonal Schur inverse.

Merge branch 'pr746/irinasok/feature-baijmkl/master'

* pr746/irinasok/feature-baijmkl/master:
Implementation of BAIJMKL matrix type similar to AIJMKL.
Supports MKL calls for MatMult (add and transpo

Merge branch 'pr746/irinasok/feature-baijmkl/master'

* pr746/irinasok/feature-baijmkl/master:
Implementation of BAIJMKL matrix type similar to AIJMKL.
Supports MKL calls for MatMult (add and transposed).
In case if used MKL version does not support mkl_sparse_optimize() (MKL 11.2 and earlier) PETSc default kernels are used.
Test added on configuration step to check if used version of MKL supports column major layout with zero based indexing.
All released versions of MKL do not support such cross case and MatSeqBAIJMKL_create_mkl_handle() converts matrix arrays to one based indexing before passing them to mkl routines.

show more ...

Bug fixes
Fix for test 107_2. To prevent MKL routine returning error in case when MatSetOption(mat,MAT_STRUCTURE_ONLY,PETSC_TRUE) is called. In this case matrix will be optimized before call to mkl_s

Bug fixes
Fix for test 107_2. To prevent MKL routine returning error in case when MatSetOption(mat,MAT_STRUCTURE_ONLY,PETSC_TRUE) is called. In this case matrix will be optimized before call to mkl_sparse_x_mv.
The same problem will appear in AIJMKl case.
Also added MatCreateBAIJMKL and MatCreateSeqBAIJMKL to external include to call them from examples

show more ...

Added new matrix classes to include files
Added define MKL_SUPPORTS_BAIJ_ZERO_BASED to suuport different versions
of MKL (older verions support only one based indexing for column major
blocks in baij

Added new matrix classes to include files
Added define MKL_SUPPORTS_BAIJ_ZERO_BASED to suuport different versions
of MKL (older verions support only one based indexing for column major
blocks in baij format)

show more ...

Merge branch 'rmills/add-aijmkl'

* rmills/add-aijmkl:
Adds AIJMKL matrix class to facilitate use of MKL sparse BLAS routines
BitBucket Pull Request 737
This is the initial version of the AIJMKL matr

Merge branch 'rmills/add-aijmkl'

* rmills/add-aijmkl:
Adds AIJMKL matrix class to facilitate use of MKL sparse BLAS routines
BitBucket Pull Request 737
This is the initial version of the AIJMKL matrix class,
a subclass of AIJ that has the goal of using MKL sparse BLAS routines whenever practicable.
If the MKL version is new enough, the sparse inspector-executor
("SpMV2" -- terrible name, but that's what the MKL team called it) routines are used by default.
Currently only MatMult, MatMultAdd, MatMultTranspose, and MatMultTransposeAdd operations are supported,
but more will be added once this initial attempt is at a point
that it provides an acceptable framework for adding additional further functionality.

show more ...

Fix MATOP_GET_VECS and MATOP_GET_SUB_MATRICES_PARALLE to match function names

Commit-type: style-fix
Reported-by: Greg Meyer <gregory.meyer@gmail.com>

Development Tools: Vim, Emacs, Eclipse

Merged in dalcinl/fix-gcc-warning (pull request #738)

Fix GCC warning from -Wsign-conversion

Approved-by: BarryFSmith <bsmith@mcs.anl.gov>

Fix GCC warning from -Wsign-conversion

Clarified documentation for MatDuplicate() and MatDuplicateOption.

Merge remote-tracking branch 'origin/hzhang/reorg-matdestroysubmats'

add MATOP_DESTROYSUBMATRICES to cleanup MatDestroySubMatrices()
Reported-by: Stefano Zampini <stefano.zampini@gmail.com>

Initial skeleton for MATSEQAIJMKL matrix type that uses MKL sparse BLAS
routines. Currently only implements mult and multadd operations, but
ready to be expanded.

Merge remote-tracking branch 'origin/barry/feature-aij-subtypes'

Add subtype concept to SeqAIJ matrices

Introduced -mat_seqaij_type <seqaijperm, ...> MatSeqAIJSetType() to allow running code written for AIJ to work
with any subtype. See MatSeqAIJRegister() and Ma

Add subtype concept to SeqAIJ matrices

Introduced -mat_seqaij_type <seqaijperm, ...> MatSeqAIJSetType() to allow running code written for AIJ to work
with any subtype. See MatSeqAIJRegister() and MatSeqAIJRegisterAll() for registering new subtypes

Use

export PETSC_OPTIONS="-mat_seqaij_type seqaijperm -malloc_test"
make alltests

to run the entire test suite using an alternative subclass, in this case seqaijperm. This will produce output regarding the different name for
the matrix class but otherwise should produce no errors.

This was implemented to 1) allow easy testing of an MKL subtype and 2) to allow users to easily use the MKL subtype
without changing their code.

Also fixes several bugs in seqaijperm, if it isn't tested its broken.

Commit-type: testing-fix,style-fix, feature
Funded-by: ECP
Project: MKL
Time: 32 hours
Reported-by: Irina Sokolova <irina.sokolova@intel.com>

show more ...