[petsc-dev] field split/schur complement, incompressible flow

[Jed Brown]

On Wed, 12 May 2010 13:27:18 -0300, Lisandro Dalcin <dalcinl at gmail.com> wrote:
> Chris is going to use this from Python, and I have to add support in
> petsc4py. I think I'll have to provide a better API, something like:
> 
> pc.fieldSplitSetFields(0,[1,2,3],4) # (rho,[Ux,Uy,Uz],p)
> 
> or
> 
> pc.fieldSplitSetIS(is_rho, is_U, is_p)
> 
> and implement then by calling PCFieldSplitSet{Fields|IS} multiple times.
> 
> Does this API make sense?

Let's think about improving the C interface.  I would really like to
name the splits since otherwise options sitting in petscrc files don't
compose.  So what about

  PCFieldSplitSetFields(PC,const char name[],PetscInt nfields,PetscInt fields[]);
  PCFieldSplitSetFieldsByName(PC,const char name[],PetscInt nfields,const char *fieldnames[]);
  PCFieldSplitSetIS(PC,const char name[],IS);

and on the Python end, the direct wrappers could be called multiple
times to set individual splits (if you provided direct wrappers) or have
the higher level interface using a dict

  pc.FieldSplitSetSplits(splits)

with any of the definitions

  splits = dict(velocity=isvelocity, pressure=ispressure))
  splits = dict(velocity=(0,1,2), pressure=(3,))
  splits = dict(velocity=('u','v','w'), pressure='p')

The latter would only work if the user used DMSetFieldName to identify
fields u,v,w,p (recall that PC has access to the DM in -dev).  With any
of the definitions above, the runtime prefixes would be
-fieldsplit_velocity_* and -fieldsplit_pressure_* instead of just
numbers.

Does this seem sensible?

Jed