A generic type for all tensor (array) containers

[benbovy]

I haven't found anywhere in xtensor such a generic type, i.e., something like xtensor_t<T, N> that we could use for xt::xtensor<T, N>, xt::pytensor<T, N>, xt::rtensor<T, N> or any other container adaptator using the tensor interface (there could be a similar type for the array interface). Am I missing something?

This would be handy for writing specific functions (e.g., that only deals with arrays of known dimensions) with clear signatures and just reuse them for bindings (python, r...).

[wolfv]

I like this idea for interface composition.

I think we can achieve a nice interface if we use the std::placeholders. Because we can't overload the template when using a size_t in the template unfortunately.

So something like

x<double>
x<T, _3>
x<int, xshape<3, 4, 2>

could work quite nicely... we could inherit from xexpression and add static asserts for the shape type or something like that to get nice error messages.

[benbovy]

I haven't thought about adding static asserts but it would be nice indeed!

I'd like to implement this kind of interface, but I have to admit that I'm a bit lost in C++ templates right now. @wolfv could you give me some hints on how to achieve this?

To give more context, here is an example of what I'd like to be able to write:

template<class T>
void func(xtensor_t<T, 2>& arr1,
          xtensor_t<int, 1>& arr2,
          xtensor_t<T, 1>& arr3)
{
   auto arr1_flat = xt::flatten(arr1);
   // do something with arr1_flat, arr2, arr3
}

and call func with either xt::xtensor or xt::pytensor or any object having the tensor interface...

The same example with a super explicit template would be something like:

template<class T, class N_Rows, class N_Cols, class N_Nodes = N_Rows * N_Cols>
void func(xtensor_t<T, xshape<N_Rows, N_Cols>>& arr1,
          xtensor_t<int, xshape<N_Nodes>>& arr2,
          xtensor_t<T, xshape<N_Nodes>>& arr3)
{
   auto arr1_flat = xt::flatten(arr1);
   // do something ...
}

but I don't know if it would work with numpy arrays and I don't even know if this is valid C++14. I'd be already very happy with the simpler version above.

[wolfv]

I'm going to look into that right now. I think we need to add another "xexpression-like" type to the inheritance chain to enable this stuff... i'll be doing some quick experiments right now.