Ordering of nb::ndarray

[qnzhou]

I am having some trouble returning a column major matrix with nanobind's ndarray class. Here is a simple example code:

m.def("get_matrix", []() {
    float *data = new float[4]{1, 2, 3, 4};

    nb::capsule owner(data, [](void *p) noexcept { delete[] (float *)p; });

    return nb::ndarray<nb::numpy, const float, nb::f_contig,
                       nb::shape<2, 2>>(data, {2, 2}, owner);
});

Here is the python output:

>>> import nanobind_example
>>> nanobind_example.nanobind_example_ext.get_matrix()
array([[1., 2.],
       [3., 4.]], dtype=float32)
>>> nanobind_example.nanobind_example_ext.get_matrix().flags
  C_CONTIGUOUS : True
  F_CONTIGUOUS : False
  OWNDATA : False
  WRITEABLE : False
  ALIGNED : True
  WRITEBACKIFCOPY : False

The result shows the numpy matrix has the right (non-owning, read only) raw data, but its order is c-contiguous instead f-continuous. It seems nb::f_contig does not work as intended. Any hint on how to make the result f-continuous? Thanks!

[wjakob]

When returning an array, the annotations like f_contig or shape are just decorations to get a nice type annotation. It's your task to give all the relevant info to the nb::ndarray constructor (runtime values) so that the result indeed has these properties. In your case, I think that you need to specify the strides parameter.

It would be an interesting possibility to compute strides automatically if the return value is annotated, but this is not implemented right now. These decorations just play a role on the input end.

[qnzhou]

Thanks for the clarification!