Dynamically typed xarray

[davidbrochart]

In xtensor-stack/xtensor-zarr#1 we will need to "discover" an array from e.g. a file, and the data type will be encoded in the metadata. This means that the array's element type will only be known at run-time. Thus, we won't be able to do something like the following if the array is an xarray:

auto a = get_array(from_file);

because xarray's dtype is a template parameter and must be known at compile-time.
Type erasure techniques involve dealing with pointers, so I guess we don't want to go in that direction because we don't want the user to use array pointers and all the complexity that comes with it.
Instead, it would be interesting to have a dynamically typed version of xarray, maybe xdtyped_array?

[ankane]

I was wondering about this as well to create a Ruby wrapper around the C++ library. ATen (library that powers PyTorch/LibTorch) has a at::Tensor class that can be any type that makes it really easy to work with.

at::Tensor t1 = at::ones({2, 2}, at::kInt);
auto t2 = at::abs(t1);

https://pytorch.org/cppdocs/

[davidbrochart]

Please have a look at zarray, although it is still a work in progress.

[ankane]

Thanks @davidbrochart, will try it out.