Cast to float of Function(R) fails with 64-bit PETSc

[Prof Garth Wells]

Tests in test/unit/python/function/test_function.py of the form

def test_real_function_float_conversion1(R):
    c = Function(R)
    assert float(c) == 0.0

seg fault when PETSc is compiled with 64-bit integers. Last line is the problem.

[Prof Garth Wells]

@johanhake Can you tell me where the code that implements the cast is?

[Martin Sandve Alnæs]

site-packages/dolfin/functions/function.py

line 493

the dtype='intc' is probably the cause.

[Prof Garth Wells]

OK. How should we fix it? The NumPy wrapping is really confusing to me. Would it be better to just use copies for integer arrays?

[Chris Richardson]

@johanhake - can you take a look at this, or bump the milestone to 1.6 otherwise.

[Johan Hake]

Yes, I will suggest a fix for this in my fix for issue 54.

[Johan Hake]

This should be fixed now. Can someone with 64bit int check that it is?

[Prof Garth Wells]

@johanhake Doesn't seem to work with 64-bit PETSc. I get

function/test_function.py:270: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ >       values = vec.gather(indices)
E       TypeError: (1) numpy array of 'intc' expected. Make sure that the numpy array use dtype=intc.
        # Gather value directly from vector in a parallell safe way
        vec = self.vector()
        indices = numpy.zeros(1, dtype=cpp.la_index_dtype())
>       values = vec.gather(indices)

Buildbot also reports this error (http://fenicsproject.org:8010/builders/dolfin-master-full-mpich/builds/1345/steps/make%20run_unittests_py/logs/stdio)

[Johan Hake]

There wasn't any sizeof(dolfin::la_index)==8 typemap for const std::vector<dolfin::la_index>&. I have pushed a fix for this to next. Let see how that works out.

[Johan Hake]

Fixed in: fdb4d29

[Prof Garth Wells]

Removing milestone: 1.5 (automated comment)