Implicit pytensor to xtensor?

[colinfang]

From https://xtensor.readthedocs.io/en/latest/bindings.html#full-qualified-api we need to define 2 extra generic functions to bridge pytensor & xtensor

However, it seems the conversion is implicit. The following works.

double test_xtensor2(const xt::xtensor<double, 2>& x){
    return xt::sum(x)();
}

double test_xtensor(const xt::pytensor<double, 2>& x){
    return test_xtensor2(x);
}

m.def("test_xtensor", &test_xtensor, "sum");

Is there anything I need to concern? From my test it doesn't seem extra copy.

[JohanMabille]

xtensor and pytensor both inherit from xexpression and both have constructor and assign operators that accept an xexpression objetct. What happens here is that a temporary of type xtensor is created from the pytensor (data is copied) and then passed to test_xtensor2.

Removing the const in test_xtensor2 should reveal that copy, it will fail with something like "cannot bind non const reference to temporary"

[colinfang]

You are absolutely correct. Silly me :(

I do hope we can have a type xtensor_t<T, N> that like what #860 suggests

[colinfang]

Perhaps xtensor_t isn't easy to build. I saw an open xexpression_dimension pr which seems to have lost momentum. I will go with SFINAE route in this case but it feels cumbersome. Is this a C++ language problem, i.e. no template constraints. (unlike rust trait/generic)

[JohanMabille]

template constraints in C++ are really primitive for now (basically enable_if), you can add some syntactic sugar, like we did with XTL_REQUIRE, but that's still requires improvements. Concepts and constraints are supposed to land in the C++20.

You can also have a look at this blogpost for solutions about API designs.

[colinfang]

I ended up with

    template<class C>
    struct is_2d: std::false_type
    {
    };

    template<>
    struct is_2d<xt::xtensor<double, 2>>: std::true_type
    {
    };
    template<>
    struct is_2d<xt::pyensor<double, 2>>: std::true_type
    {
    };

    template<class T>
    using check_2d = std::enable_if_t<is_2d<T>::value, bool>;

    // Prepend the following to the function.
    template<class T, typename = check_2d<T>>
    void compute2(const T& t) {
    }

Maybe xtensor can ship a few traits like check_2d out of box?