dolfin team mailing list archive

[Johan Hake <783857@xxxxxxxxxxxxxxxxxx>]

This is inheritly difficult as I do not know of any way to "attach" the 
original Matrix to the generated data arrays. We do this other places, like 
when we down_cast a Matrix. Then we store a reference on the down_casted 
matrix so the latter will not go out of scope when no references are pointing 
to the parent matrix.

But it is a bit trickier with this case, as the NumPy arrays are generated in 
a C++ layer outside the C++ Matrix class. In this method we do not know 
anything about the host Python object. This means that we cannot store a 
reference to the host Python object to the generated NumPy arrays. One could 
hypothetize that we could attach the host object after the NumPy arrays were 
generated, but a NumPy array does not support setting arbitrary attributes.

I do not see any straight forward solution to this.

Johan

-- 
You received this bug notification because you are a member of DOLFIN
Team, which is subscribed to DOLFIN.
https://bugs.launchpad.net/bugs/783857

Title:
  uBLASSparseMatrix data deleted when object goes out of scope

Status in DOLFIN:
  New

Bug description:
  This bug is similar to https://bugs.launchpad.net/dolfin/+bug/747273

  The following code produces the output:

  0.0333333333333
  1.8218374251e-316

  I.e. after deleting the uBLASSparseMatrix object the data members are
  freed even though there are other references to the data.

  
  from dolfin import *

  mesh = UnitTetrahedron()
  V = FunctionSpace(mesh, "Nedelec 1st kind H(curl)", 1)
  u = TrialFunction(V)
  v = TestFunction(V)
  m = inner(v,u)*dx
  M = uBLASSparseMatrix()
  assemble(m, tensor=M)
  row,col,data = M.data()
  print data[0]
  del M
  import gc
  gc.collect()
  print data[0]