Memory leakage in python bindings

[bleyerj]

There seems to be memory leakage problems when running some mgis.fenics demo (phase-field example at least)

[thelfer]

The trouble is in the mgis_convert_to_span which relies on the PyArray_FROM_OTF function to retrieve the data associated with an numpy array. I misread the documentation of this function and though it only retrieves the pointer of the numpy object, but it does more:

• if its argument is a numpy array it increases the internal counter.
• if it is not, it tries to convert the object in a new numpy array.

In both cases, the counter of the returned object must be decreased at the end of the function, which I did not. Hence the memory leak.

The following implementation shall work:

    mgis::span<mgis::real> mgis_convert_to_span(
        const boost::python::object& o) {
      PyArrayObject* a = nullptr;
      if (!PyArg_ParseTuple(o.ptr(), "O!", &PyArray_Type, &a)) {
        mgis::raise(
            "convert_to_span: argument not convertible to PyArrayObject");
      }
      if (a->descr->type_num != NPY_DOUBLE) {
        mgis::raise("convert_to_span: invalid numpy object");
      }
      if (a == nullptr) {
        mgis::raise(
            "convert_to_span: argument not convertible to PyArrayObject");
      }
      // number of dimensions
      const auto nd = PyArray_NDIM(a);
      if (nd != 1) {
        mgis::raise("convert_to_span: expected one dimensional array");
      }
      auto* const shape = PyArray_DIMS(a);
      const auto n = shape[0];
      auto* const values = static_cast<double*>(PyArray_GETPTR1(a, 0));
      return {values, static_cast<mgis::span<mgis::real>::index_type>(n)};
    }  // end of mgis_convert_to_span

[thelfer]

Fixed in master and rliv-1.1