Are you thinking of using PCSetModifyMatrices()? Actually, this brings up another issue (now on petsc-dev). Should this callback be attached to a Mat, instead of a PC? Then MatGetSubMatrices() can use this callback on each of the extracted submatrices, if necessary. In the case of a KKT system, it would then set the nontrivial block size on the primal submatrix. Dmitry. On Wed, Apr 11, 2012 at 3:33 PM, Jed Brown <[email protected]> wrote:

The original matrix might not have blocks, but that does not mean that a submatrix does not have blocks. On Apr 11, 2012 3:30 PM, "Mark F. Adams" <[email protected]> wrote:

...

On Apr 11, 2012, at 11:24 AM, Jed Brown wrote:

On Wed, Apr 11, 2012 at 12:01, Mark F. Adams <[email protected]>wrote:

...

eed to know the block size (eg, 4). I get this info by having the user set the block size of the matrix. This is what you should do for the non-contact test. This will not work with contact equations ... or at least it does not make sense. How can I get this info?

1) hard wire it: if I see 7 vectors then bs=4.

Hell no.

...

2) have setNearNullSpace take a block size, and set the vector block size, which I can get like I do from the matrix now.

No.

When you call MatGetSubMatrix(), you will get back a matrix that has a near null space (and a block size if that happens to be relevant; there will not be a block size if deformation and thermodynamics are in different spaces). Don't worry about how that is done, just call MatGetSubMatrix().

OK, I don't see how you know that the block size is 4 ... do you call MatSetBlockSize even if there are constraints that are not blocked?