Adapting std::array and using in expression results in xtensor, not xtensor_fixed

[vakokako]

Take a look at this example:

std::array shape{3, 4, 5};
auto shape_adapt = xt::adapt(shape);
auto tensor = xt::eval(2 * shape_adapt);

The resulting type of tensor is xtensor_container, which is vector underneath. So we went from array to vector.

This is a very common use case for working with tensor shapes and would optimize quite a bit. Currently we go for std::transform and xtl::make_sequence in these cases, which is not so nice. Adapting shapes would be much nicer))

[tdegeus]

I think this is a good idea to have. I do think though that using shape as an example is confusing. xt::adapt transforms data into an xtensor object (of the appropriate shape). It does not allocate a tensor of a certain shape (as the example is suggesting).

Furthermore, there is a question about defaults. In particular, I do think that xtensor_fixed should get a bit of attention before this could be the default: I have been having many problems on Windows, so personally I have temporarily stopped using xtensor_fixed altogether.

In that light, it may make sense to allow an option or template argument for xt::adapt?

[vakokako]

The example was about adapting the shape of the container (usually std::array or std::vector) to perform some arithmetic on it (get the shape that is twice as big).

Sorry for the confusion.

Also a nice example is working with multidimensional indices: e.g. getting the central element:

auto& center = tensor[xt::adapt(tensor.shape()) / 2];

[tdegeus]

@vakokako OK, that clarifies things. But then isn't the wanted extension more that some of xtensor's function should work on xtensor objects too (if they now only work on std::array or std::vector)?