[petsc-dev] MATNEST MATSUBMATRIX MATLOCALREF MATBLOCKMAT and friends

[Barry Smith]

    Since Mike McCourt is in dire need of flexible multiphysics infrastructure I've started to look at some of the newly implement Mat stuff in PETSc (and my head is spinning).

   Could the author's of these things take a look at my comments and respond (hopefully with changes, but I can change things also).

   Stream of consciousness writing because I can't get my head around it all.

----

   MatCreateNest() (and VecCreateNexs()).  Does not have the the proper construction process that always goes through MatCreate_Nest() (and VecCreate_Nest()). This needs to be fixed. Should add something like MatNestAddBlocks() or similar name for putting the blocks in after the call to MatCreate_Nest().

  MatCreateNest() seems to assume that all Mat's live on all processes in the outer Mat communicator. How hard would it be to change the code so that each block could live on some subcommunicator?

  MatCreateSubMatrix() will it eventually have MatSetValues() that directs requests down to the appropriate block?

   
----

   MatCreateSubMatrix(). Again not the proper construction process built on MatCreate_SubMatrix() (when the heck was this written 10 years ago before we enforced that rule.) This needs to be fixed. What's with the manual page description "Creates a composite matrix that acts as a submatrix" what the heck does 'composite' mean here and what does "acts as a submatrix" mean? 

  MatSubMatrixUpdate() what does the documentation "full matrix in the submatrix" mean? The full matrix is  bigger than the submatrix so how can it be in the submatrix. Terrible phrase.

   MatSubMatrixUpdate() why have this routine, why not model on MatGetSubMatrix() and have a single interface function MatCreateSubMatrix(Mat,IS,IS,MatRuse,Mat *)?

   MatCreateSubMatrix() terrible name, why not MatGetSubMatrixImplicit()? 

------

  MatCreateLocalRef() again with no proper construction with MatCreate_LocalRef() (come on guys get with the program)

  MatCreateLocalRef() 'Gets a logical reference to a local submatrix, for use in assembly' what does this mean? Assembly in PETSc is reserved for inside MatAssembly,VecAssembly it should not be used for the process of setting values into the matrix. Thus this sentence is meaningless gibberish.  Could possible be something like 'Gets a logical reference to a local submatrix, for use in putting values into that block with MatSetValues() or MatSetValuesLocal()?

MatCreateLocalRef()  How/why is this different than MatCreateSubMatrix()? Could they be merged into one construct? Why is this a 'local' ref, is it a block only only on this process?

-------
  How is MatCreateBlockMat() which fortunately has the proper constructor related to MatCreateNest()? Do they serve the same purpose? Different purposes? Can they be merged together? 
--------

   In the end we want to be able to run 'multphysics'  codes where it is a command line switch between 'block matrix' storage of the global Jacobian where each block is a submatrix of the entire matrix and stored itself for example in AIJ format; or one big honking AIJ matrix. In both cases the user calls MatSetValues() or MatSetValuesLocal() to put values in and has no switches in the code they write that depends on the matrix format, things like MatGetSubMatrix() and MatGetSubMatrices() work transparently on them. How do these four Matrix classes take us in that direction?

  Thanks


  Barry