Quoting Harish Narayanan <email address hidden>:

> I noticed one more thing since this fix to get it compiling with FFC.
>
> The .h file that is generated with and without this
> strip_variables(expand_derivatives()) hack is different. Furthermore, in
> my test cases, the version with the hack results in code that converges
> quadratically whereas the version without does not converge. I have
> attached the test case I am considering.
>
> ** Attachment added: "test.tbz2"
>    http://launchpadlibrarian.net/31646352/test.tbz2

Sure it converges, if I increase the maximum number of iterations in the Newton
solver... :)

First of all, variable names are not guaranteed to be the same on consecutive
runs so the output might differ if you compare using e.g., kdiff3 or meld. The
convergence problems, however, clearly indicates that something is wrong.

I use the same code to handle the Variable object in the quadrature transformer
as is used in the strip_variables function. The reason for the different
results is that expand_indices is called in quadrature transformer BEFORE
handling variables. In your hack strip_variables is called first and surely, if
I change
strip_variables(expand_derivatives())
to
strip_variables(expand_indices(expand_derivatives()))
the convergence problems appears again. So we can conclude that expand_indices
is indeed evil as Martin pointed out.

I'll disable the handling of Variables in the transformer and instead call
strip_variables before expand_indices, then your form should work without any
hacks such that we can close this bug.

I'll add a blueprint about removing the expand_indices call, which should also
speed up compilation. However, this will not be very high on my todo list.

Kristian

> --
> Problems parsing form file with diff and derivative
> https://bugs.launchpad.net/bugs/426512
> You received this bug notification because you are a bug assignee.
>
> Status in FEniCS Form Compiler: Confirmed
>
> Bug description:
> FFC has problems parsing the following form file. As I have pointed out in
> the code, I need to use the following:
>
> from ufl.algorithms import expand_derivatives, strip_variables
> L = strip_variables(expand_derivatives(L_temp))
> a = strip_variables(expand_derivatives(a_temp))
>
> to get it to compile via FFC. This is not needed for SFC, and while SFC takes
> a few seconds (without these lines), FFC takes 6m 25s)
>
> Also, FFC doesn't like fractional exponents which is pointed out in bug
> #423398.
>
> # Finite strain elasticity
>
> # Function space
> vector = VectorElement("Lagrange", "tetrahedron", 1)
>
> # Functions
> v  = TestFunction(vector)   # Test function
> du = TrialFunction(vector)  # Incremental displacement
> B  = Function(vector)       # Body force
> u  = Function(vector)       # Displacement at previous iteration
>
> # Kinematics
> I = Identity(3)
> F = I + grad(u).T
> C = F.T*F
> J = sqrt(det(C))
> Fbar = J**(-1./3.)*F  # Does not like fractional exponents
> Cbar = J**(-2./3.)*C  # Does not like fractional exponents
>
> E = (J*C - I)/2       # Want to try Cbar here, but using J*C instead
> E = variable(E)
>
> # Material law
> mu    = Constant("tetrahedron")
> lmbda = Constant("tetrahedron")
>
> psi = lmbda/2*(tr(E)**2) + mu*tr(E*E)
> S = diff(psi, E)
> P = F*S
>
> # Balance of momentum in the reference configuration
> L_temp = inner(P, grad(v).T)*dx - inner(B, v)*dx
> a_temp = derivative(L_temp, u, du)
>
> # FIXME: Shouldn't need to do the following
> from ufl.algorithms import expand_derivatives, strip_variables
> L = strip_variables(expand_derivatives(L_temp))
> a = strip_variables(expand_derivatives(a_temp))
>