Eigen Matrix and Numpy Array

[petrasvestartas]

Hi,

Is there any example that shows numpy to eigen matrix transfer of nested lists?
For example [[0,0,0], [1,5,8]].

All examples in nanonbind tests are flat arrays: [0,0,0,1,5,8]

[wjakob]

There isn't a good solution for taking non-array input at the moment. Basically nanobind expects something that already implements either DLPack or the buffer protocol. I am open to external contributions in this regard but am not working on such a feature myself.

[petrasvestartas]

Thank you for a reply. I have something that compiles and works on python and c++. Took me a while to understand. Here how I transfer an eigen matrix with rows of 3 numbers (for mesh faces and vertices transfer).
The matrices below comes from pybind11 project and can be used in nanobind as well.

I think there is no really documentation about the dynamic and rowmajor types of eigen matrix.

I attach the full project as a zip.

compas_cgal_dev.zip

On C++:

I have types:

    using RowMatrixXd = Eigen::Matrix<double, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
    using RowMatrixXi = Eigen::Matrix<int, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;

  /**
   * @brief Remesh a triangle mesh with target edge length
   * 
   * @param V Matrix of vertex coordinates (row-major)
   * @param F Matrix of face indices (row-major)
   * @param target_edge_length Desired length for mesh edges
   * @param number_of_iterations Number of remeshing iterations
   * @param do_project Whether to project vertices onto the input surface
   * @return std::tuple<RowMatrixXd, RowMatrixXi> Tuple containing new vertices and faces (both row-major)
   */
  nb::tuple
  pmp_remesh(
      Eigen::Ref<compas::RowMatrixXd> V,
      Eigen::Ref<compas::RowMatrixXi> F,
      double target_edge_length,
      unsigned int number_of_iterations,
      bool do_project);

On Python:

def remesh():
    # Create a test mesh
    mesh = Mesh.from_polyhedron(20)
    vertices, faces = mesh.to_vertices_and_faces()
    
    # Convert to numpy arrays with row-major order (C-style)
    vertices = np.asarray(vertices, dtype=np.float64, order='C')
    faces = np.asarray(faces, dtype=np.int32, order='C')
    
    # Test remeshing
    target_edge_length = 0.5
    num_iterations = 10

    remeshed_vertices, remeshed_faces = remesh(
        vertices, faces, target_edge_length, num_iterations
    )

    return remeshed_vertices, remeshed_faces

[petrasvestartas]

One thing that I do not understand if I need to define the return policy to move to avoid copying?

Should I include nb::rv_policy::move below?

void init_meshing(nb::module_& m) {
    auto submodule = m.def_submodule("meshing");

    submodule.def(
        "remesh",
        &pmp_remesh,
        "Remesh a triangular mesh to achieve a target edge length",
        "V"_a,
        "F"_a,
        "target_edge_length"_a,
        "number_of_iterations"_a = 10,
        "do_project"_a = true,
        nb::rv_policy::move // the returned object is moved rather than copied
    );
}

[wjakob]

It is already the default when returning by value.

[petrasvestartas]

Thank you