Reference to a Vector of Eigen Matrix

[petrasvestartas]

How do you make a reference to a Vector of Eigen Matrix?

For vectors I used this, but this does not work for vector of eigen matrices, values cannot be modified:

from {{cookiecutter.project_slug}}._vectors_reference import DoubleVector
from {{cookiecutter.project_slug}}._vectors_reference import subtract_inplace

a = DoubleVector([1, 2, 3])
b = DoubleVector([4, 6, 8])

subtract_inplace(a, b)

for i in range(len(a)):
    print(a[i])

#include "compas.h"
#include <nanobind/stl/bind_vector.h>


using DoubleVector = std::vector<double>;
void subtract_inplace(DoubleVector& a, const DoubleVector& b) {
    for (size_t i = 0; i < a.size(); i++) {
        a[i] = a[i] - b[i];
    }
}


NB_MODULE(_vectors_reference, m) {
    m.doc() = "Vectors by reference example.";
    nb::bind_vector<DoubleVector>(m, "DoubleVector");
    m.def("subtract_inplace", &subtract_inplace, "a"_a, "b"_a, "Subtract two vectors in place");
}

[wjakob]

Please look at the caveats of nb::bind_vector. The problem is that vectors can resize. If Python objects reference the existing Eigen vectors, then you can run into situations where a resize operation causes dangling Python objects. So they must make a copy.

[petrasvestartas]

Thank you.

An without using vectors. Is there any example to use an array of arrays std::array<std::array, int>> ?

[wjakob]

You would need to make your own binding, there is no bind_vector version of std::array. Other than that, this combination should work.

[petrasvestartas]

When I do the following, it works, but am I making a copy?

typedef std::vector<nb::ndarray<double>> VectorNumpyArrayDouble;

NB_MAKE_OPAQUE(VectorNumpyArrayDouble);

void lmgc90_solve_numpy(
    VectorNumpyArrayDouble& vertices_arrays,  // Each array has shape (n_vertices, 3)
) {

    if (!vertices_arrays.empty()) {
        nb::ndarray<double> vertices = vertices_arrays[0]; // Python based assignment no copy
        double* vertex_data = vertices.data();
        vertex_data[0] = 10;  // Set x coordinate of first vertex to a large value
    }
}



NB_MODULE(_lmgc90, m) {
    
    nb::bind_vector<VectorNumpyArrayDouble>(m, "VectorNumpyArrayDouble");

    m.def("lmgc90_solve_numpy", 
        &lmgc90_solve_numpy, 
        "vertices_arrays"_a
    );
}

[wjakob]

Right, this is allowed. You can access a bound std::vector<> in C++ bindings and change the contents. But that same operation copies elements when done in Python. The rationale and a workaround (dangerous) is described in the reference documentation of nb::bind_vector.

[petrasvestartas]

Thank you, there are two options then
a) unique pointers
b) nb::rv_policy::reference_internal - more dangerous