[BUG]: In-place modification of ndarray does not work

[ivan-tkatchev]

Problem description

A C++ function accepts an ndarray and modifies it in place, but changes aren't reflected on the Python side.

>>> import numethods
>>> import numpy as np
>>> arr = np.array([[1,1],[1,2],[1,3],[2,4],[2,5],[2,6]])
>>> numethods.emva(arr)
>>> arr
array([[1, 1],
       [1, 2],
       [1, 3],
       [2, 4],
       [2, 5],
       [2, 6]])

If the C++ code is modified to return the input array then the function works as expected:

array([[1.        , 0.3       ],
       [1.        , 0.81000006],
       [1.        , 1.467     ],
       [2.        , 1.2       ],
       [2.        , 2.3400002 ],
       [2.        , 3.4380002 ]], dtype=float32)

Reproducible example code

#include <cmath>

#include <nanobind/nanobind.h>
#include <nanobind/ndarray.h>

using array_of_pairs_t = nanobind::ndarray<
    float,
    nanobind::shape<-1, 2>,
    nanobind::c_contig,
    nanobind::device::cpu>;

void emva(array_of_pairs_t&& data_) {
    constexpr float a = 0.3;

    auto data = data_.view();
    float prev_key = NAN;
    float emva = 0;

    for (size_t i = 0; i < data.shape(0); ++i) {
        float key = data(i, 0);
        float val = data(i, 1);

        if (key != prev_key) {
            emva = 0;
            prev_key = key;
        }

        emva = emva * (1 - a)  + val * a;
        data(i, 1) = emva;
    }
}

NB_MODULE(numethods, m) {
    m.def("emva", &emva);
}

[ivan-tkatchev]

Obviously the array is being copied when passed into the C++ function, but there is no indication of why or how.
Documentation says nothing about possible hidden magic copies of ndarray on the Python/C++ boundary.

[ivan-tkatchev]

If dtype is specified explicitly during array creation then the function works as expected:

>>> import numethods
>>> import numpy as np
>>> arr = np.array([[1,1],[1,2],[1,3],[2,4],[2,5],[2,6]], dtype='float32')
>>> arr
array([[1., 1.],
       [1., 2.],
       [1., 3.],
       [2., 4.],
       [2., 5.],
       [2., 6.]], dtype=float32)
>>> numethods.emva(arr)
>>> arr
array([[1.        , 0.3       ],
       [1.        , 0.81000006],
       [1.        , 1.467     ],
       [2.        , 1.2       ],
       [2.        , 2.3400002 ],
       [2.        , 3.4380002 ]], dtype=float32)

[wjakob]

If your input array on the Python side has a different dtype, then nanobind must make an implicit conversion to be able to do the call. You can disable implicit conversions via nb::arg("arg_name").noconvert(), in which case you will get an error message instead.

[wjakob]

This is documented, by the way: please review the ndarray part of the documentation here.