Behaviour of Function(f), where f is a Function is not consistent with DOLFIN about firedrake

[funsim]

The branch with the suggested fix is called "issue14".

[dham]

You probably want to implement the copy using fnew.assign(f). This will
cause the copy to be implemented in a PyOP2 kernel with the effect that, in
the GPU case, you avoid a spurious copy back to host.

[dham]

I don't think that's correct. By just grabbing dat.val and using it as val,
you'll get two Functions which point to the same data space. I think you
want to create the new dat without Val and do the copy after.

[funsim]

Good point.
I extended the test case to check that indeed a deep copy is created:

from firedrake import *

mesh = UnitIntervalMesh(2)
V = FunctionSpace(mesh, "CG", 1)
f = Function(V)
f.interpolate(Expression("1"))
assert (f.dat.vec.array == [1, 1, 1]).all() # Passes

g = Function(f)
print g.dat.vec.array
assert (g.dat.vec.array == [1, 1, 1]).all() # Fails

# Check that g is indeed a deep copy
f.dat.vec[:] = np.array([2, 2, 2])

assert (f.dat.vec.array == [2, 2, 2]).all()
assert (g.dat.vec.array == [1, 1, 1]).all()

Will work on a fix.

[wence-]

FWIW, you should look at data in dats with:

f.dat.data_ro

The vec property is the petsc vector which won't play nicely with GPU implementations at the moment (CUDA doesn't currently use petsc).

Oh, and if you're going to write to things use f.dat.data.

So I'd write the test like:

m = UnitIntervalMesh(2)
V = FunctionSpace(m, 'CG', 1)
f = Function(V)
f.interpolate(Expression("1"))
assert (f.dat.data_ro == 1.0).all()
g = Function(f)
assert (g.dat.data_ro == 1.0).all()
g.interpolate(Expression("2"))
assert (g.dat.data_ro == 2.0).all()
assert (f.dat.data_ro == 1.0).all()

[dham]

In general, that's true. But in this case you shouldn't directly access the
data at all, but instead employ a par loop to do the copy.

[wence-]

I meant in terms of the testing part.

[funsim]

ok, will change that.

[wence-]

Closed by 3af7d4b