Thanks @yoshipon, we really like this solution.

The main remark is about code formatting. Our convention is the following

Regarding formatting, here is the current convention:

namespace bar
{                                               // indent in namespace
    class Foo
    {                                           // new line before {
    public:
                                                // empty line after public /  private / protected
         Foo();
    };

    Foo::Foo()
    {
    }
}                                               // no use of PYBIND macros for NAMESPACE_END NAMESPACE_BEGIN

@SylvainCorlay Thank you for the comments.
I formatted the files. Please confirm the revised codes.

@yoshipon Thanks for this PR.

Not sure about that, but the errors may be due to the fact that we do not use the same versions of pybind. The CI (travis and appveyor) install pybind11 2.1.1, which is not the last version of pybind.

@JohanMabille Exactly, my code could not be compiled with pybind11 2.1.1 due to the following two reasons:

1. move_cast_op_type does not exist
2. strides and shape have to be std::vector<size_t> (std::array was not supported yet)

I fixed these bugs.

However, even now there are several problems on Travis CI, I'm still working for debugging.

The Travis CI errors seem to be coming from the new miniconda that was released yesterday. (Hence the build flag which seem to come from their python/distutils build).

Downloading Miniconda2-4.3.21-Linux-x86_64.sh instead of latest might solve the issue.

@SylvainCorlay Thank you for the information. I re-pushed with a modified .travis.yml.

Success!

For your information, I have reported the issue upstream the new miniconda build (in their gitter channel).

Merging 🎉 congrats on your first PR into xtensor!

@yoshipon regardin the xtensor/xarray to ndarray cast, one thing that we will probably try to do before the next release is not make use of pybind11::array, and use xt::pyarray instead.

It should probably be a matter of adding a constructor to xt::pyarray that is optimized in that it would reuse the buffer of xtensor/xarray.

@SylvainCorlay Thank you very much! Actually, this is my first PR. It was very impressive.
btw, it is a good idea. The current type_caster specifically for xarray/xtensor is not the best way.
With this extension, it will be more simplified.