Periodic boundary conditions do not work with Epetra backend and nonlinear solver

[Patrick Farrell]

Hi,

(I know there's a bug about periodic BCs with Epetra not working before, but I think that this is a different case.)

Consider the following code:

from dolfin import *
import random
import sys

parameters["linear_algebra_backend"] = "Epetra"

mesh = UnitInterval(40)
V = FunctionSpace(mesh, "CG", 2)

class PeriodicBoundary(SubDomain):
  def inside(self, x, on_boundary):
    return abs(x[0]) < DOLFIN_EPS and on_boundary

  def map(self, x, y):
    y[0] = x[0] - 1

def main(ic):

  out = Function(V)
  u = TrialFunction(V)
  v = TestFunction(V)
  bc = PeriodicBC(V, PeriodicBoundary())

  solve(inner(u, v)*dx == inner(ic, v)*dx, out, bc)

  F = inner(out, v)*dx - inner(ic, v)*dx
  solve(F == 0, out, bc)

  return out

if __name__ == "__main__":
  ic = Function(V)
  vec = ic.vector()
  for i in range(len(vec)):
    vec[i] = random.random()

  out = main(ic)

The *first* solve, written in the form solve(a == L), works fine.
The *second* solve, written in the form solve(F == 0), fails:
*** -------------------------------------------------------------------------
*** Error: Unable to set block of values for Epetra matrix.
*** Reason: Did not manage to perform Epetra_FECrsMatrix::ReplaceGlobalValues.
*** Where: This error was encountered inside EpetraMatrix.cpp.
*** Process: 0
*** -------------------------------------------------------------------------

With the fenics-dev package 1.0.0-2+bzr6748~ppa1~precise1, if you comment out the Epetra backend, then both solves work fine.