Expand PetscMatrix interface #2169
Conversation
Methods added: - local_m - local_n - get_local_size - reset_preallocation
|
If someone knows the other matrix types, then at least the size methods could be moved down into the base class and made virtual |
src/numerics/petsc_matrix.C
Outdated
|
|
||
| auto ierr = MatGetLocalSize (_mat, &m, PETSC_NULL); |
There was a problem hiding this comment.
Minor issue: we just simply use NULL in PETSc right now. We were using PETSC_NULL, but not any more.
There was a problem hiding this comment.
Haha, really? I guess we should actually be changing them all to nullptr now but I don't really care either way. Was PETSC_NULL ever not NULL? What issue was it used to address?
There was a problem hiding this comment.
I have no idea whey we were using PETSC_NULL. I still have a lot of PETSC_NULL in my PhD code:-)
In PETSc, we have this:
#define PETSC_NULL NULL
But most of PETSc code is using NULL
Cody forced me to use nullptr for PETSc code in MOOSE.
| { | ||
| libmesh_assert (this->initialized()); | ||
|
|
||
| auto ierr = MatResetPreallocation(_mat); |
There was a problem hiding this comment.
Ugh @fdkong you told me the wrong version number here. Hence the test failures on idaholab/moose#13682. Looks like it should be 3.9.0 based on git tag --contains. I'm guessing you just made a typo sigh.
There was a problem hiding this comment.
https://www.mcs.anl.gov/petsc/petsc-3.9/docs/manualpages/Mat/MatResetPreallocation.html
It is in 3.9. I am confused here.
There was a problem hiding this comment.
!PETSC_VERSION_LESS_THAN(3,8,0) means >= 3.8.0, i.e. including 3.8.0
|
Comments hopefully addressed |
| template <typename T> | ||
| void PetscMatrix<T>::reset_preallocation() | ||
| { | ||
| #if !PETSC_VERSION_LESS_THAN(3,8,0) |
There was a problem hiding this comment.
Oh I see.
#if !PETSC_VERSION_LESS_THAN(3,9,0)
8-->9
I was thinking about 9. Why I gave you 8
| /** | ||
| * Get the number of columns owned by this process | ||
| */ | ||
| numeric_index_type local_n () const; |
There was a problem hiding this comment.
I'm confused by this one. My first interpretation here is "the total number of columns with non-empty entries in the sparsity pattern in rows owned by this process", but I'm not actually sure how PETSc would determine that - we give them n_nz and n_oz rather than complete sparsity info, right? And without the latter there's no way to immediately determine which non-empty columns from which rows overlap? So is this method only well-defined for matrices that have already been assembled??
There was a problem hiding this comment.
It is not related to nnz, noz or matrix assembly. It is a layout thing. It is something we tell PETSc to use in libMesh.
In PetscMatrix::init(), we have something like
ierr = MatSetSizes(_mat, m_local, n_local, m_global, n_global);
LIBMESH_CHKERR(ierr);
There was a problem hiding this comment.
Thanks @fdkong. This is what I had determined through experimentation this morning 😄
There was a problem hiding this comment.
That certainly clears it up! Thanks!
| /** | ||
| * Get the number of rows owned by this process | ||
| */ | ||
| numeric_index_type local_m () const; |
There was a problem hiding this comment.
Can we bump this up to the parent class and make the default implementation row_stop()-row_start()?
Methods added: