assembling over subspace fails - create a better error message

[Nico Schlömer]

When assembling a form over a subspace, e.g., W.sub(0), the assembler will cowardly bail out.

MWE:

from dolfin import *

mesh = UnitSquareMesh(20, 20)

V = FunctionSpace(mesh, 'CG', 1)
W = V * V

print V
print W.sub(0)

u = TrialFunction(W.sub(0))
v = TestFunction(W.sub(0))
assemble(u * v * dx)

fails with

[0]PETSC ERROR: Arguments are incompatible!
[0]PETSC ERROR: Local column size 882 cannot be larger than global column size 441!

[Prof Garth Wells]

W.sub(0) provides a view into the FunctionSpace W, which is the cause the problem. We'll need to dig into the code to find a resolution.

We have had other issues with the distinction between FunctionSpaces and views into FunctionSpaces. Maybe we should make a FunctionSpace view a different type.

[Jan Blechta]

@nschloe What would you like to assemble? A matrix whose rows and columns are indexed by a subset of {0, 1, 2, ..., W.dim()-1}?

[Nico Schlömer]

@blechta I would have expected to get the same matrix as for assembling over V since it appears to be really the same as W.sub(0).

[Prof Garth Wells]

It could be that this threw an error in the past, but that since Johan added the dof ownership range to sub-maps it allocates a vector of the wrong length. Needs some digging.

[Jan Blechta]

@nschloe I think that Garth pointed out many times that V is not the same as W.sub(0).

[Prof Garth Wells]

The fix for this is a more informative error message.

[Anders Logg (Chalmers)]

Removing milestone: 1.7 (automated comment)