Typemap problem for std::vector<Point>

[August Johansson]

There is some problem with the typemap for std::vector<Point>. For example, unit testing of ConvexTriangulation::triangulate yields

TypeError: in method 'ConvexTriangulation_triangulate', argument 1 of type 'std::vector< dolfin::Point >'

I've disabled these tests for now.

[August Johansson]

The ConvexTriangulation test is in

geometry/test_intersection_construction.py

The error also causes four XDMFFile write errors in

io/test_XDMF.py.

I will disable these as well.

[Anders Logg (Chalmers)]

Is this issue still relevant? @augustjohansson @benjamik

[August Johansson]

Yes, I think it's still a problem. At least I haven't gotten a solution into my code yet :)

[Anders Logg (Chalmers)]

I think @magneano did some work on typemaps for Point? Perhaps that could be used.

[Magne Nordaas]

I did some work on "out" typemaps for multimesh quadrature rule types. Those use std::vector<double> instead of std::vector<Point> to represent sets of points.

Currently there is no typemap for Point and creating a Point in the Python interface results in a swig proxy object, so I don´t think it makes sense to add an "in" typemap for std::vector<Point>. Instead, we could modify the Python interface for the relevant functions to handle Python list inputs.

[Jan Blechta]

It should be pretty easy to make the typemaps work. Check also https://bitbucket.org/fenics-project/dolfin/src/6440f6114c3e20f6188f003bbce3b1b283fbcab5/dolfin/swig/typemaps/std_vector.i?at=master&fileviewer=file-view-default#std_vector.i-745.

Tweaking typemaps might be preferable than changing the interface.

[Jan Blechta]

The typemap indeed works. Try

from dolfin import *

cpp = """
#include <vector>

namespace dolfin {

  void foo(const std::vector<Point>& p)
  {
    info("%f", p[0].x());
  };

}
"""

foo = compile_extension_module(cpp).foo
foo([Point(666.0)])

[August Johansson]

@blechta is of course absolutely correct. @benjamik fixed this some time ago. I think we can mark this as resolved.

[Martin Sandve Alnæs]

If a large number of points is wanted, Magne is right that swig proxy object can be an issue. There's both a severe memory penalty and takes time to allocate and wrap/unwrap these objects.

[Jan Blechta]

Well, should be there std::vector<Point> interface on C++ side if the number is really large?

[August Johansson]

I'd guess that the size is typically less than 10. More importantly, the functions are typically not used from python at all, but internally in the C++ MultiMesh class.

[Anders Logg (Chalmers)]

Correct. This is mainly useful for development and unit testing of functionality implemented in Python (with the convenience of Python for doing testing and experiments).

[Magne Nordaas]

I was not aware that this was fixed. Thanks Jan!

I suggested changing the interface instead of defining a typemap because it seems a little convoluted to create a list of swig proxy objects for passing to a typemap, even more so if it is typemap that is not used anywhere else. I would like to discuss when typemaps are preferable to modifying the interface, but this is probably not the right place for it.

[Anders Logg (Chalmers)]

Is this still an issue? @magneano @augustjohansson

[Anders Logg (Chalmers)]

No longer relevant when soon moving to pybind11