removedMatScaleSystem() and MatUnscaleSystem() since they are never used (and confusing).

Hg-commit: 68ac9ef26bce6c77fd63c90d793869c24ca4006c

updated and merged from petsc-dev

Hg-commit: 4bdbb24181fe1c8ce13e9b82c1d499b152c3d14f

reactivated inheritance of block sizes in MatGetSubMatrix methods with a fix of taking the block size from the IS.

Hg-commit: f497b269393bfde79ad711ca7edb9bdcec8ad83b

Lots of changes. Also addign new files for cusparse matrix type.

Hg-commit: d2f118b395b24fdc836f857e3f221f6c0c9eca58

Add basic Elemental interface

## Implemented
* MatMult and MatMultAdd
* MatSetValues for locally owner parts (no stash at this point)
* Use MatGetOwnershipIS() to determine which entries a process s

Add basic Elemental interface

## Implemented
* MatMult and MatMultAdd
* MatSetValues for locally owner parts (no stash at this point)
* Use MatGetOwnershipIS() to determine which entries a process should be responsible for assembling
* Viewing is currently in Elemental ordering, can use MatComputeExplicitOperator() for debugging

## Short-term priorities
* Grid needs to be shared between multiple Mats on the same communicator, implement by attribute caching on the MPI_Comm
* Stash for assembly of unowned entries, can cache locally and use Elemental's (collective) "axpy" interface
* MatMat operations, factorization, etc

Hg-commit: 9d1af7db82c9477605dc39b514d53db5a9855fc7

show more ...

Add MatGetOwnershipIS() in preparation for Elemental interface that assembles arbitrary blocks

Hg-commit: 6cd4929d272f8c56c38c3e0b3132f60cabd49474

Removed check on ncols from MatRestoreRow()

Hg-commit: 0e365b9f93745fdc883fbbb9fe0db062c63d8afd

missing E at end of fortran name MATNULLSPACEREMOVE

Hg-commit: c73774eebb74e2ff26b0801f485e8ede1eb835d5

removed old code for column block sizes in GAMG.

Hg-commit: 4ca60e8beaf3c684480d127e797f44d7b491ceb3

added MatSet[Get]BlockSizes.

Hg-commit: 2b256a2808b1d92f504a512b9f9dfc98e2a84aa4

PetscTypeCompare -> PetscObjectTypeCompare

Hg-commit: 3fc8f665421ddb862bf4d1ba95505378246340be

dead global variables whould have been removed when macro changed to in line function

Hg-commit: d06cfc99078494fb1858d3ea56075bc231da15a0

MatSetOption() no longer requires Mat preallocation/setup to have taken place

Hg-commit: 0df01447dc2284fb8bcb2dbaa74633dff06d0835

removed processing of -mat_block_size from MatXXXSetPreallocation() routines moved to MatSetFromOptions()
fixed various locations that reset blocksize after Mat was setup

Hg-commit: 253c882709a78d62

removed processing of -mat_block_size from MatXXXSetPreallocation() routines moved to MatSetFromOptions()
fixed various locations that reset blocksize after Mat was setup

Hg-commit: 253c882709a78d627d884888fbdb4a7528173c93

show more ...

MatSetSize() called after MatSetUp() is now forbidden even for Dense matrices
MatSetBlockSize() must be called beofre MatSetUp()/ MatXXXSetPreallocation() for all matrices
Fixed vec nest example that

MatSetSize() called after MatSetUp() is now forbidden even for Dense matrices
MatSetBlockSize() must be called beofre MatSetUp()/ MatXXXSetPreallocation() for all matrices
Fixed vec nest example that pushed format that prevented values from being printed

Hg-commit: 80ec81eb9db92735994e950eda2d2db6657bb6cb

show more ...

added -vec_view_info

Hg-commit: 374aff94c8c189edae42759446538d3b1eb8e792

help menu for Mat-Mat operations

Hg-commit: b76800020dccedc4257f1374e207b6eb3ffe15de

Add near null space operations

Hg-commit: d571e74a2045f867c385653e6c9505b53dfd847e

MatView() prints the block size if it is different from 1

Hg-commit: 43296097787c9d2a1ae33da82f7338a6d6a8d667

shorten event names, because 'long names messing up alignment in -log_summary' as Jed complains.

Hg-commit: 3912dfbfb18c6b2c4acb669d2596ea6fc2e6e2e4

Renamed MatMerge_SeqsToMPI() to MatCreateMPIAIJSumSeqAIJ(); Renamed MatMerge() to MatCreateMPIAIJConcatenateSeqAIJ()

Hg-commit: 24f46f7b98f555b1e552573a5bd3ac3737ff4f2b

Rename include/private/ to include/petsc-private/

Hg-commit: 9a1730720b0d9756437b40462617551504641ade

Move printing of nthreads to MatView_SeqAIJPThread()

Hg-commit: e81245fedddbc65416af67d8073d3e16e055344f

Undid this change since it breaks stuff. Note that the entire MatAssembly Infrastructure depends on these flags being set correctly. I just wasted hours tracking down a bug due to this change. Also n

Undid this change since it breaks stuff. Note that the entire MatAssembly Infrastructure depends on these flags being set correctly. I just wasted hours tracking down a bug due to this change. Also note that our test system sucks since the bug was introduced a month ago.

Note also that assembled flag is MUCH more than just "values need to be communicated". The business of garray and colmap depend on this flag being set correctly.

Note also that all the other MatSetXXX still changed the flags correctly.

If you want some functionality like below that allows changing numerical values (but not nonzero locations) then introduce MatGetRowWrite() and MatRestoreRowWrite() and change MatGetRow() to MatGetRowRead(). But please don't don't underestimate side affects when you make such dramatic changes.

author Matthew G Knepley <knepley@gmail.com>
Tue, 14 Feb 2012 00:45:05 -0600 (4 weeks ago)
changeset 22179 1fc53ac2b8ab
parent 22178 c1d3e1b5603f
child 22180 708aa4809612
Now MatSetValues() does not automatically mark the matrix as unassembled
- Individual types set the assembly flag if values need to be communicated
- This is done to allow programming in the following way:
for(r = rStart; r < rEnd; ++r) {
MatGetRow()
<transform values>
MatRestoreRow()
MatSetValues()
}

Hg-commit: 7ae47bfbd6e9aa96e390d0fe01d428bae6c9eac4

show more ...

Print number of threads in MatView() when using pthreadsclasses

Hg-commit: f10b5a825d7b8643fd2e00ed7d93a735cf21a468