Convert cuda ndarray to void**

[johnviljoen]

Hi - I have a 2D ndarray that I need to convert into a vector of pointers, which each point to a row of the ndarray. My problems lie in the fact that the ndarrays seem to only be metadata until the python runs, so don't seem to want to give me a memory address beyond the first one at ndarray.data(). I have tried using strides like this:

template <typename Scalar, typename... Other>
std::vector<void*> make_pointer_array_2d(const nb::ndarray<Scalar, Other...> &arr) {
    // 1) Check that we have a 2D array
    if (arr.ndim() != 2) {
        throw std::runtime_error("make_pointer_array_2d() only handles 2D arrays!");
    }

    // 2) Number of rows, columns
    int64_t nrows = arr.shape(0);

    // 3) stride in "elements" not bytes
    int64_t row_stride = arr.stride(0);

    // 4) The base pointer
    Scalar* base_ptr = (Scalar*) arr.data(); // device or host, depending on array domain

    // 5) Build a vector of row pointers
    std::vector<void*> pointers(nrows);
    for (int64_t i = 0; i < nrows; i++) {
        pointers[i] = static_cast<void*>(base_ptr + i * row_stride);
    }

    return pointers;
}

This seems to not work, however I am not sure why. I have also tried to use numpy indexing like ndarray(i); to get a row of a 2D matrix, but it seems I can only retrieve a single element with two indices: ndarray(i, i);

What should I do in this situation? I would rather not pass every row from python separately. Also I am very much a c++ beginner so sorry if I am making silly mistakes here!