assembly of mixed spaces gives wrong results
Here @dnarnold reported the following:
Here is a simple piece of code high-lighting a bug which results in a stiffness matrix being completely wrong. It computes a bilinear form acting on a test function u
from Lagrange P1 on the unit interval, and a pair of test functions (v, c)
from a mixed function space consisting of piecewise constants for v
and a real number for c
. The bilinear form has two parts:
a0 = u.dx(0) * v * dx # integral of u'*v over the interval
a1 = u * c * Expression("1.-x[0]") * ds # value of u*c at the point x=0
If the mesh has n
intervals, then the first bilinear form should assemble into the first n
rows of the (n+1) x (n+1)
stiffness matrix, and the second into the final row. The problem is that in fact the second bilinear form assembles into the wrong row. Therefore if the full matrix is assembled from a0 + a1
it has a row of all zeros at the end, so is of course singular.
I first reported this in FEniCS Q&A for version 1.3.0. It persists in version 1.4.0 and the latest development version.
from dolfin import *
mesh = UnitIntervalMesh(4)
V1 = FunctionSpace(mesh, 'CG', 1)
V2 = FunctionSpace(mesh, 'DG', 0)
R = FunctionSpace(mesh, 'R', 0)
u = TrialFunction(V1)
X = MixedFunctionSpace([V2, R])
(v, c) = TestFunctions(X)
# this bilinear form does not involve the real test function c,
# so one row should be identically zero
a0 = u.dx(0) * v * dx
print assemble(a0).array()
# this bilinear form only involves the real test function c
# so the corresponding row should have a nonzero
a1 = u * c * Expression("1 - x[0]") * ds
print assemble(a1).array()
Finally, here is the output.
[[ 0. 0. 0. 1. -1.]
[ 0. 1. -1. 0. 0.]
[ 0. 0. 1. -1. 0.]
[ 1. -1. 0. 0. 0.]
[ 0. 0. 0. 0. 0.]]
[[ 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 1.]
[ 0. 0. 0. 0. 0.]
[ 0. 0. 0. 0. 0.]]
Note that the zero row of the first array is the 5th row, but the nonzero row of the 2nd array is the third, not the fifth row.
As an answer to the posting in Q&A MiroK added some diagnostic output that indicates the problems is in DofMapBuilder::reorder_local, and found that setting
parameters['reorder_dofs_serial'] = False
works around the problem.
Comments (12)
-
-
Setting parameters['reorder_dofs_serial'] = False still makes the code work.
Changing the Real subspace to DG0 and keeping reordering on gives this message:
File "/home/martinal/opt/fenics/dorsal-dev-141022/lib/python2.7/site-packages/dolfin/fem/assembling.py", line 203, in assemble assembler.assemble(tensor, dolfin_form) RuntimeError: *** Error: All dofmaps must have same block size (for now)
-
@garth-wells is this something you know about and plan to fix (considering the "for now" comment in the error)?
-
-
- changed milestone to 1.5
-
- marked as critical
-
-
assigned issue to
Assigning to Garth because of the relation between this and https://bitbucket.org/fenics-project/dolfin/issue/352/assembly-of-rectangular-matrices-broken which Garth has said to fix.
-
assigned issue to
-
Looking into this. Seems to be an issue with DG + Real elements. Below code throws an error (at least an error is better than a wrong result).
from dolfin import * mesh = UnitIntervalMesh(4) V2 = FunctionSpace(mesh, 'DG', 0) R = FunctionSpace(mesh, 'R', 0) X = MixedFunctionSpace([V2, R])
-
@garth-wells, with your failing example I observe using debugger that block size is 2. So my suspicion is that
DofMapBuilder::compute_blocksize
does not work as expected in this case. -
@blechta , yes looks like the block size is not computed correctly. If I force
bs=1
, the resulting matrices appear correct. I'll figure out what's wrong with the block size. -
- changed status to resolved
Check for global dofs before computing dofmap block size.
Fixes
#331. The 'actual' bug is in FFC/UFC, see https://bitbucket.org/fenics-project/ffc/issue/61/.→ <<cset 3c0d0be1b571>>
-
- removed milestone
Removing milestone: 1.5 (automated comment)
- Log in to comment
This code now gives: