Split forms on mixed function spaces by kynan · Pull Request #194 · firedrakeproject/firedrake

kynan · 2014-03-13T15:46:06Z

When passing a form to the ffc_interface, the form's arguments are
split and a form per block in the mixed space is returned.

dham · 2014-03-14T10:38:28Z

Do we really want to unconditionally silence FFC?

This is what we used to do in PyOP2, but maybe not.

wence- · 2014-03-14T11:28:06Z

list of kernels?

In fact, it's more complicated than that right. It's a list of block indices and kernels, yes?

True, that docstring needs updating.

wence- · 2014-03-17T09:57:48Z

This doesn't work for problems cast in nonlinear form:

from firedrake import *

mesh = UnitSquareMesh(2, 2)
V1 = FunctionSpace(mesh, "BDM", 1, name="V")
V2 = FunctionSpace(mesh, "DG", 0, name="P")
W = V1 * V2

# Define variational problem
lmbda = 1
x = Function(W)
u, p = TrialFunctions(W)
v, q = TestFunctions(W)
f = Function(V2)

f.interpolate(Expression("(1+8*pi*pi)*sin(x[0]*pi*2)*sin(x[1]*pi*2)"))
F = (p*q - q*div(u) + lmbda*inner(v, u) + div(v)*p) * dx - f*q*dx


a = lhs(F)
L = rhs(F)
F = action(a, x) - L

solve(F == 0, x)

Results in:

   solve(F == 0, x)
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 756, in solve
    _solve_varproblem(*args, **kwargs)
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 798, in _solve_varproblem
    nullspace=nullspace)
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 111, in __init__
    self._jac_tensor = assemble(self._problem.J_ufl, bcs=self._problem.bcs)
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 365, in assemble
    return _assemble(f, tensor=tensor, bcs=_extract_bcs(bcs))
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 384, in _assemble
    kernels = compile_form(f, "form")
  File "/data/lmitche1/src/firedrake/python/firedrake/ffc_interface.py", line 150, in compile_form
    for forms in FormSplitter().split(form):
  File "/data/lmitche1/src/firedrake/python/firedrake/ffc_interface.py", line 73, in split
    integrand = self.visit(it.integrand())
  File "/data/lmitche1/src/fenics/ufl/ufl/algorithms/transformer.py", line 101, in visit
    r = h(o, *map(self.visit, o.operands()))
  File "/data/lmitche1/src/fenics/ufl/ufl/algorithms/transformer.py", line 101, in visit
    r = h(o, *map(self.visit, o.operands()))
  File "/data/lmitche1/src/fenics/ufl/ufl/algorithms/transformer.py", line 101, in visit
    r = h(o, *map(self.visit, o.operands()))
  File "/data/lmitche1/src/fenics/ufl/ufl/algorithms/transformer.py", line 105, in visit
    r = h(o)
  File "/data/lmitche1/src/firedrake/python/firedrake/ffc_interface.py", line 86, in argument
    if self._idx[o.count()] == i:
KeyError: 0

kynan · 2014-03-17T17:54:27Z

@wence- I have fixed that case and added the test you were running. The form splitter wasn't correctly dealing with the case of renumbered arguments.

wence- · 2014-03-18T09:38:08Z

Thanks, here's another one:

from firedrake import *

mesh = UnitSquareMesh(2, 2)
V1 = FunctionSpace(mesh, "BDM", 1, name="V")
V2 = FunctionSpace(mesh, "DG", 0, name="P")
W = V1 * V2

# Define variational problem
lmbda = 1
x = Function(W)
u, p = split(x)
v, q = TestFunctions(W)

F = (inner(u, v) + v[1]*p)*dx

assemble(derivative(F, x))

   assemble(derivative(F, x))
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 364, in assemble
    return _assemble(f, tensor=tensor, bcs=_extract_bcs(bcs))
  File "/data/lmitche1/src/firedrake/python/firedrake/solving.py", line 383, in _assemble
    kernels = compile_form(f, "form")
  File "/data/lmitche1/src/firedrake/python/firedrake/ffc_interface.py", line 166, in compile_form
    kernel, = FFCKernel(form, name + str(i) + str(j)).kernels
  File "/data/lmitche1/src/PyOP2/pyop2/caching.py", line 69, in __new__
    obj.__init__(*args, **kwargs)
  File "/data/lmitche1/src/firedrake/python/firedrake/ffc_interface.py", line 132, in __init__
    ffc_tree = ffc_compile_form(form, prefix=name, parameters=ffc_parameters)
  File "/data/lmitche1/src/fenics/ffc/ffc/compiler.py", line 172, in compile_form
    pyop2_ir = [generate_pyop2_ir(ir, prefix, parameters) for ir in oir[2]]
  File "/data/lmitche1/src/fenics/ffc/ffc/quadrature/quadraturepyop2ir.py", line 54, in generate_pyop2_ir
    body_ir = _tabulate_tensor(ir, parameters)
  File "/data/lmitche1/src/fenics/ffc/ffc/quadrature/quadraturepyop2ir.py", line 148, in _tabulate_tensor
    opt_par, parameters)
  File "/data/lmitche1/src/fenics/ffc/ffc/quadrature/quadraturepyop2ir.py", line 406, in _generate_element_tensor
    nest_ir, ops = _generate_integral_ir(points, terms, sets, optimise_parameters, parameters)
  File "/data/lmitche1/src/fenics/ffc/ffc/quadrature/quadraturepyop2ir.py", line 599, in _generate_integral_ir
    if len(loop) == 0:
UnboundLocalError: local variable 'loop' referenced before assignment

wence- · 2014-03-18T09:38:28Z

This is simplified from Colin's Skamarock-Klemp model.

kynan · 2014-03-18T11:48:40Z

@wence- You need this UFL branch (which still waits on review): https://bitbucket.org/mapdes/ufl/pull-request/8/simplification-rules-required-for-form/

What is happening is that UFL doesn't detect that the first block of this form is effectively 0. This form is then fed to FFC, which does some more analysis and ends up with no terms in its IR. @FabioLuporini's PyOP2 IR currently chokes on this with this rather unhelpful error.

I've added some simplification rules to make UFL detect common cases of "empty" blocks so that we're not even passing those on to FFC.

kynan · 2014-03-21T17:20:57Z

Is this now good to go?

dorugeber · 2014-03-21T17:51:23Z

Why did this pass tests? It's not rebased on master, so it doesn't have Lawrence's updated facet tests in. The old test should be failing, now that the FFC fix is in.

kynan · 2014-03-21T18:04:07Z

@dorugeber Travis tests the result of the pull request i.e. this branch merged with master.

dorugeber · 2014-03-23T14:09:04Z

Thanks for clarifying that.

I assume this branch is fine, but Doru and I had some strange behaviour when we rebased our facet branch on this. (Rebasing on this was our solution to the integral ordering mismatch, and we found an unexpected extra argument was being passed into the kernel)

It shouldn't block this PR from going ahead, assuming everything works at the moment, but we'll probably be asking you for advice in the next couple of days.

Does not work with disk caching for now since Coefficients can't be serialised.

Adapt to changes in compile_form, which now returns a four-tuple of index, domain type, coefficients and kernel. We now need to pull apart the tensor and test and trial functions if it is a mixed type and extract the component for the block the kernel contributes to and filter the boundary conditions to only include those defined on the current block.

FFCKernel goes back to handling only a single kernel on a single integral.

Instead of making one pass over the form and getting back a list of subforms on the blocks of the mixed space, make as many passes as there are blocks in the mixed space and only enabling one component of each Argument at a time. When splitting an Argument, create a UFL ListTensor where only the component of the MixedFunctionSpace selected by the current block index is included and all other components are set to Zero. These Zeros are subsequently eliminated as they are used in expressions and not propagated up the form tree.

When building integral_data for the form_data object, the domain_data which Firedrake uses to carry the coordinate field is stripped. We therefore need to get the integrals with intact integral_data via the preprocessed form.

kynan · 2014-04-03T16:14:42Z

Can this finally land? The Travis failure is unrelated as you can tell and it passes on buildbot.

Split forms on mixed function spaces

L2pullbacks

dham reviewed Mar 14, 2014
View reviewed changes

wence- mentioned this pull request Mar 14, 2014

Migrate FFC interface to Firedrake OP2/PyOP2#305

Merged

wence- reviewed Mar 14, 2014
View reviewed changes

kynan added 15 commits April 2, 2014 11:55

Migrate FFC interface from PyOP2

e5f0bbf

Add compatible FFC version

91beb3d

First stab at a FormSplitter

6324bd8

Sum integrands on the same measure before splitting a form

14f4109

Reconstruct Indexed and IndexSum for non fixed indices

11211f0

Check for MixedFunctionSpace when splitting Argument/Coefficient

77ea556

Don't return a tuple if a Sum only has one block

ccb0155

Fix condition for reconstructing index_sum

4f077f6

Split forms when building FFCKernel

6fdea5c

Does not work with disk caching for now since Coefficients can't be serialised.

Split the form in compile_form

d1fb01d

FFCKernel goes back to handling only a single kernel on a single integral.

Handle grad in FormSplitter

daa105b

FormSplitter: don't split coefficients

7d5f3e8

Only sum integrands if they are on the same domain

5dee7a4

kynan added 12 commits April 2, 2014 18:15

Only call FormSplitter if we have a mixed form

67a39f6

Access integrals via preprocessed form

4278503

When building integral_data for the form_data object, the domain_data which Firedrake uses to carry the coordinate field is stripped. We therefore need to get the integrals with intact integral_data via the preprocessed form.

Extract coord field from measure in assemble thunk

b0d25e6

Add firedrake-clean script to remove cached FFC kernels

6ad536f

Form splitter: correctly deal with renumbered form arguments

6de929b

Add mixed helmholtz test cast in nonlinear form

92d066f

Add test assembling derivative of form with 0 block

7956122

Cache split arguments to keep them unique

9272831

Allow extracting subspace 0 from (Vector)FunctionSpace

e16e4dd

Add helper functions to build mat/vec in assemble thunk

e1ffde0

Eliminate duplicate code for horizontal interior/exterior facet

b202a7b

Allow overriding FFC kernel cache directory from environment

a158704

kynan added a commit that referenced this pull request Apr 3, 2014

Merge pull request #194 from firedrakeproject/split-forms

70749be

Split forms on mixed function spaces

kynan merged commit 70749be into master Apr 3, 2014

kynan deleted the split-forms branch April 3, 2014 16:18

This was referenced Apr 4, 2014

"Petsc has generated inconsistent data" error when running in parallel #233

Closed

Rank of terms in a form #237

Closed

connorjward pushed a commit that referenced this pull request Dec 4, 2024

Merge pull request #194 from firedrakeproject/l2pullbacks

3db8f06

L2pullbacks

Uh oh!

Conversation

kynan commented Mar 13, 2014

Uh oh!

dham Mar 14, 2014

Choose a reason for hiding this comment

Uh oh!

kynan Mar 14, 2014

Choose a reason for hiding this comment

Uh oh!

wence- Mar 14, 2014

Choose a reason for hiding this comment

Uh oh!

wence- Mar 14, 2014

Choose a reason for hiding this comment

Uh oh!

kynan Mar 14, 2014

Choose a reason for hiding this comment

Uh oh!

wence- commented Mar 17, 2014

Uh oh!

kynan commented Mar 17, 2014

Uh oh!

wence- commented Mar 18, 2014

Uh oh!

wence- commented Mar 18, 2014

Uh oh!

kynan commented Mar 18, 2014

Uh oh!

kynan commented Mar 21, 2014

Uh oh!

dorugeber commented Mar 21, 2014

Uh oh!

kynan commented Mar 21, 2014

Uh oh!

dorugeber commented Mar 23, 2014

Uh oh!

kynan commented Apr 3, 2014

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants