TSFC representation is not compatible with DOLFIN no-cell hack

[Jan Blechta]

demo_poisson1D-in-2D.py with parameters.form_compiler.representation = 'tsfc' fails with the message

[...]
  File "/home/fenics/local/lib/python3.5/site-packages/ffc/tsfc/tsfcrepresentation.py", line 45, in compute_integral_ir
    interface=ufc_interface)
  File "/home/fenics/local/lib/python3.5/site-packages/tsfc/driver.py", line 108, in compile_integral
    builder.set_coefficients(integral_data, form_data)
  File "/home/fenics/local/lib/python3.5/site-packages/tsfc/kernel_interface/ufc.py", line 98, in set_coefficients
    expressions = prepare_coefficients(coeffs, n, name, self.interior_facet)
  File "/home/fenics/local/lib/python3.5/site-packages/tsfc/kernel_interface/ufc.py", line 174, in prepare_coefficients
    elements = [create_element(coeff.ufl_element()) for coeff in coefficients]
  File "/home/fenics/local/lib/python3.5/site-packages/tsfc/kernel_interface/ufc.py", line 174, in <listcomp>
    elements = [create_element(coeff.ufl_element()) for coeff in coefficients]
  File "/home/fenics/local/lib/python3.5/site-packages/tsfc/finatinterface.py", line 177, in create_element
    raise ValueError("Don't know how to build element when cell is not given")
ValueError: Don't know how to build element when cell is not given

[Jan Blechta]

It's not just "some DOLFIN demos" - the same problem exists even in poisson demo.

This is caused by DOLFIN no-cell hack in Expressions. Changing degree=2 to element=FiniteElement("P", mesh.ufl_cell(), 2) fixes the problem. Need to find a way how to make tsfc-repr compatible with the hack. One would expect that cell is already attached by UFL preprocessing. @martinal, any hints?

[Martin Sandve Alnæs]

This should already be worked around in UFL preprocessing.

[Jan Blechta]

Thanks! Maybe TSFC's compute_form_data circumvents this workaround.

[Jan Blechta]

@miklos1, any clue why this happens? Preprocessing goes through here which should take care of the problem but TSFC later complains. Any possibility that it works with unpreprocessed form?

[Miklós Homolya]

@blechta: I suspect this should be easy to solve if only I could inspect it in the debugger. But to do that, I first have to figure out how to install DOLFIN...

[Miklós Homolya]

OK, just reading the source code I suspect the problem is that the element map is never used in TSFC.

[Jan Blechta]

But to do that, I first have to figure out how to install DOLFIN...

Docker would be a good one-shot alternative.

[Jan Blechta]

UFLACS does it here

[Jan Blechta]

I can try fixing it. The question is

• whether use just element_replace_map or function_replace_map,
• whether to fix it just in UFC kernel interface and leave Firedrake pipeline intact,
• whether coefficient numbering is correct now.

Or should I leave it to you?

[Miklós Homolya]

I see UFLACS "just" (not considering that futzing about coefficient numbering) calls replace on the preprocessed integrand expression, and then everything just works later. That's why I didn't see anything related around FFC's fiatinterface.py.

whether to fix it just in UFC kernel interface and leave Firedrake pipeline intact

I don't think we have an API for kernel interface specific UFL munging.

I'm trying to understand what happens in the UFLACS example before the ufl.replace call. I first thought it was there to undo some UFL renumbering, but looking at assert i == g.count() it looks more like a sanity check, since coefficient_numbering is completely redundant since coefficient_numbering[g] == g.count().

If you have a patch that fixes this in DOLFIN, I can review it and test it in Firedrake. Meanwhile I'm trying to get a working DOLFIN installation...

[Miklós Homolya]

I can reproduce the issue...

[Miklós Homolya]

This patch seems to fix poisson and poisson1D-in-2D. @blechta: I don't know how to test all things in DOLFIN with parameters.form_compiler.representation = 'tsfc'.

[Jan Blechta]

Yeah, tried to do the same but got the replace on the wrong place.

test all things in DOLFIN

Me neither. I think @martinal removed FFC_FORCE_REPRESENTATION in 1fc650149442122e3d06fe2ee0e1bf9b5f19e17b. I'll bring it back now.

[Jan Blechta]

To be hones it is a mess. I found it difficult to get understand which members of FormData should be remapped. And UFLACS part demonstrates it too. Maybe there should be a clear mechanism in UFL to perform the mapping.

@martinal, why was an application of replace maps removed from UFL and put to representations?

[Jan Blechta]

@miklos1, do you want an access to optimize through env too?

[Miklós Homolya]

Not sure what you mean.

[Jan Blechta]

To enforce optimize=False through env to test without optimizations (default is True).

[Jan Blechta]

With branch https://bitbucket.org/fenics-project/ffc/branch/jan/repr-env-var one can do export FFC_FORCE_REPRESENTATION=tsfc. I would add FFC_FORCE_OPTIMIZE too if you want and merge.

[Miklós Homolya]

I'm happy to have optimisations on, that will test more things. I guess I can always disable that for debugging by editing some global default in a Python file, right?

[Jan Blechta]

Sure, you can do parameters.form_compiler.optimize = False.

I've taken a risk and merged the branch into master and next.

[Jan Blechta]

cd test/unit
DIJITSO_CACHE_DIR=$PWD FFC_FORCE_REPRESENTATION=tsfc py.test-3 python/multimesh/test_multimesh.py

fails as expected but more informative message would be nice

E         File "/vivid/home/jan/dev/fenics-master/src/ffc/ffc/representation.py", line 453, in _compute_integral_ir
E           parameters)
E         File "/vivid/home/jan/dev/fenics-master/src/ffc/ffc/tsfc/tsfcrepresentation.py", line 45, in compute_integral_ir
E           interface=ufc_interface)
E         File "/vivid/home/jan/dev/fenics-master/src/tsfc/tsfc/driver.py", line 93, in compile_integral
E           integration_dim, entity_ids = lower_integral_type(fiat_cell, integral_type)
E         File "/vivid/home/jan/dev/fenics-master/src/tsfc/tsfc/driver.py", line 352, in lower_integral_type
E           basedim, extrdim = dim
E       TypeError: 'int' object is not iterable

DIJITSO_CACHE_DIR=$PWD FFC_FORCE_REPRESENTATION=tsfc py.test-3 python/fem/test_form.py -k test_derivative

fails with

E         File "/vivid/home/jan/dev/fenics-master/src/finat/finat/fiat_elements.py", line 24, in degree
E           return self._element.degree()
E       AttributeError: 'MixedElement' object has no attribute 'degree'

This might be resolved with #143 (so far tested only with masters + this patch).

Parallel tests

DIJITSO_CACHE_DIR=$PWD FFC_FORCE_REPRESENTATION=tsfc mpirun -n 3 py.test-3

are fine.

[Jan Blechta]

I'd like to get a confidence that there's nothing wrong with coefficient numbering (after the map is applied). Do you understand the logic behind that, @miklos1?

I guess that if it was wrong we might not see a crash in DOLFIN regression tests but FFC regression test should catch it.

[Miklós Homolya]

I'm confident that the TSFC does not change the ordering of coefficients, so it was correct so far, it shall stay that way. Yes, I also believe that the FFC regression tests ought to have caught such an issue.

[Jan Blechta]

Fixed upstream in TSFC. Thanks @miklos1.

[Jan Blechta]

https://github.com/blechta/tsfc/commit/ee5670ab1ad90aaf8f550de01255c6752583255d