XDMFFile.write second argument type

[Marco Morandini]

This test

from dolfin import *
mesh = UnitCubeMesh(1,1,1)
element_eps_p_eq_project = FiniteElement("CG", mesh.ufl_cell(), degree=1)
V_eps_p_eq_project = FunctionSpace(mesh, element_eps_p_eq_project)
eps_p_eq_project = Function(V_eps_p_eq_project, name="eps_p_eq")
filex = XDMFFile('pippo.xdmf')
filex.write(eps_p_eq_project, 0)
#filex.write(eps_p_eq_project, 0.) # this works
#filex.write(eps_p_eq_project, t=0) # this works
#filex.write(u=eps_p_eq_project, t=0) # this works

fails for me with

Traceback (most recent call last):
  File "pippo3.py", line 7, in <module>
    filex.write(eps_p_eq_project, 0)
ValueError: vector::reserve

[Michal Habera]

Thanks for reporting this. I have spent several hours trying to debug... The issue is related to pybind11 bindings.

Binding of XDMFFile::write in https://bitbucket.org/fenics-project/dolfin/src/925edd1586195242d27aa1b438464f2ed83372be/python/src/io.cpp?at=master&fileviewer=file-view-default#io.cpp-374 is for time variable as double. The first call run according to pybind11 overload resolution order (section 7.8 in https://media.readthedocs.org/pdf/pybind11/latest/pybind11.pdf ) tries the binding without int -> double conversion, so the function binding is not matched.

For some unknown reason, the dolfin.function.Function matches std::vector<dolfin::Point> in "write" method binding for points (https://bitbucket.org/fenics-project/dolfin/src/925edd1586195242d27aa1b438464f2ed83372be/python/src/io.cpp?at=master&fileviewer=file-view-default#io.cpp-387) that follows in the first run. This then fails with vector::reserve strange error.

I have a workaround where "write" method matches general py::object and then does the logic to distinguish Point from Function, but the problem is still strange dolfin.function.Function matching std::vector<dolfin::Point>. Any ideas why this happens, @chris_richardson @garth-wells ?

[Michal Habera]

After several days of fighting I have finally found the snag.

Any Python object that has __getitem__ magic method is matched as std::vector<> for pybind11 bindings python -> cpp conversion. Since dolfin.function.function.Function (terrible naming, btw...) has __getitem__ (this is from UFL's exproperators.py) the problem is born.

@chris_richardson I will prepare PR also for the other related issue.

Consider the following simple code

import dolfin as d

base_code = '''
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
namespace py = pybind11;

#include <vector>
#include <stdio.h>
#include <dolfin/function/Function.h>
#include <dolfin/geometry/Point.h>

void write(std::vector<dolfin::Point>& points)
{
  std::cout << "Write called for dolfin::Point" << std::endl;
}

void write(dolfin::Function& fction)
{
  std::cout << "Write called for dolfin::Function" << std::endl;
}


PYBIND11_MODULE(SIGNATURE, m)
{
  m.def("write", (void (*)(std::vector<dolfin::Point>&)) &write);
  m.def("write", (void (*)(dolfin::Function&)) &write);
  m.def("write", [](py::object obj){
    std::cout << "Py obj" << std::endl;
  });
}
'''

compiled = d.compile_cpp_code(base_code)

mesh = d.UnitSquareMesh(2, 2)
V = d.FunctionSpace(mesh, "CG", 1)

u = d.Function(V)

points = [d.Point(1), d.Point(1)]

class BadObject:
    def __getitem__(self, pos):
        return pos

bo = BadObject()

# Fails when converting to std::vector
# but should be caught by the last binding
#compiled.write(u)

# Fails when converting to std::vector
# but should be caught by the last binding
#compiled.write(bo)

# OK
compiled.write(points)

# OK
compiled.write(u.cpp_object())

[Nicholas Musolino]

This report was the top Google result for “pybind11 ValueError: vector::reserve”.

The underlying bug in pybind11 has been reported as issue #2076: https://github.com/pybind/pybind11/issues/2076