Designing language bindings with xtensor

xtensor and its Related projects make it easy to implement a feature once in C++ and expose it to the main languages of data science, such as Python, Julia and R with little extra work. Although, if that sounds simple in principle, difficulties may appear when it comes to define the API of the C++ library. The following illustrates the different options we have with the case of a single function compute that must be callable from all the languages.

Generic API

Since the xtensor bindings provide different container types for holding tensors (pytensor, rtensor and jltensor), if we want our function to be callable from all the languages, it must accept a generic argument:

template <class E>
void compute(E&& e);

However, this is a bit too generic and we may want to enforce that this function only accepts xtensor arguments. Since all xtensor containers inherit from the “xexpression” CRTP base class, we can easily express that constraint with the following signature:

template <class E>
void compute(const xexpression<E>& e)
{
    // Now the implementation must use e() instead of e
}

Notice that with this change, we lose the ability to call the function with non-constant references or rvalue references. If we want them back, we need to add the following overloads:

template <class E>
void compute(xexpression<E>& e);

template <class E>
void compute(xexpression<E>&& e);

In the following we assume that the constant reference overload is enough. We can now expose the compute function to the other languages, let’s illustrate this with Python bindings:

PYBIND11_MODULE(pymod, m)
{
    xt::import_numpy();

    m.def("compute", &compute<pytensor<double, 2>>);
}

Full qualified API

Accepting any kind of expression can still be too permissive; assume we want to restrict this function to 2-dimensional tensor containers only. In that case, a solution is to provide an API function that forwards the call to a common generic implementation:

namespace detail
{
    template <class E>
    void compute_impl(E&&);
}

template <class T>
void compute(const xtensor<T, 2>& t)
{
    detail::compute_impl(t);
}

Exposing it to the Python is just as simple:

template <class T>
void compute(const pytensor<T, 2>& t)
{
    detail::compute_impl(t);
}

PYBIND11_MODULE(pymod, m)
{
    xt::import_numpy();

    m.def("compute", &compute<double>);
}

Although this solution is really simple, it requires writing four additional functions for the API. Besides, if later, you decide to support array containers, you need to add four more functions. Therefore this solution should be considered for libraries with a small number of functions to expose, and whose APIs are unlikely to change in the future.

Container selection

A way to keep the restriction on the parameter type while limiting the required amount of typing in the bindings is to rely on additional structures that will “select” the right type for us.

The idea is to define a structure for selecting the type of containers (tensor, array) and a structure to select the library implementation of that container (xtensor, pytensor in the case of a tensor container):

// library container selector
struct xtensor_c
{
};

// container selector, must be specialized for each
// library container selector
template <class C, class T, std::size_t N>
struct tensor_container;

// Specialization for xtensor library (or C++)
template <class T, std::size_t N>
struct tensor_container<xtensor_c, T, N>
{
    using type = xt::xtensor<T, N>;
};

template <class C, class T, std::size_t N>
using tensor_container_t = typename tensor_container<C, T, N>::type;

The function signature then becomes

template <class T, class C = xtensor_c>
void compute(const tensor_container_t<C, T, 2>& t);

The Python bindings only require that we specialize the tensor_container structure

struct pytensor_c
{
};

template <class T, std::size_t N>
struct tensor_container<pytensor_c, T, N>
{
    using type = pytensor<T, N>;
};

PYBIND11_MODULE(pymod, m)
{
    xt::import_numpy();

    m.def("compute", &compute<double, pytensor_c>);
}

Even if we need to specialize the “tensor_container” structure for each language, the specialization can be reused for other functions and thus reduce the amount of typing required. This comes at a cost though: we’ve lost type inference on the C++ side.

xt::xtensor<double, 2> t {{1., 2., 3.}, {4., 5., 6.}};

compute<double>(t);  // works
compute(t);          // error (couldn't infer template argument 'T')

Besides, if later we want to support arrays, we need to add an “array_container” structure and its specializations, and an overload of the compute function:

template <class C, class T>
struct array_container;

template <class C, class T>
struct array_container<xtensor_c, T>
{
    using type = xt::xarray<T>;
};

template <class C, class T>
using array_container_t = typename array_container<C, T>::type;

template <class T, class C = xtensor_c>
void compute(const array_container_t<C, T>& t);

Type restriction with SFINAE

The major drawback of the previous option is the loss of type inference in C++. The only means to get it back is to reintroduce a generic parameter type. However, we can make the compiler generate an invalid type so the function is removed from the overload resolution set when the actual type of the argument does not satisfy some constraint. This principle is known as SFINAE (Substitution Failure Is Not An Error). Modern C++ provide metafunctions to help us make use of SFINAE:

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

template <class T, std::size_t N, layout_type L, class Tag>
struct is_tensor<xtensor<T, N, L, Tag>> : std::true_type
{
};

template <class T, template <class> class C = is_tensor,
          std::enable_if_t<C<T>::value, bool> = true>
void compute(const T& t);

Here when C<T>::value is true, the enable_if_t invocation generates the bool type. Otherwise, it does not generate anything, leading to an invalid function declaration. The compiler removes this declaration from the overload resolution set and no error happens if another “compute” overload is a good match for the call. Otherwise, the compiler emits an error.

The default value is here to avoid the need to pass a boolean value when invoking the compute function; this value is of no use, we only rely on the SFINAE trick.

This declaration has a slight problem: adding enable_if_t to the signature of each function we want to expose is cumbersome. Let’s make this part more expressive:

template <template<class> class C, class T>
using check_constraints = std::enable_if_t<C<T>::value, bool>;
template <class T, template <class> class C = is_tensor,
          check_constraints<C, T> = true>
void compute(const T& t);

All good, we have type inference and an expressive syntax for declaring our function. Besides, if we want to relax the constraint so the function can accept both tensors and arrays, all we have to do is to replace the default value for C:

// Equivalent to is_tensor<T>::value || is_array<T>::value
template <class T>
struct is_container : xtl::disjunction<is_tensor<T>, is_array<T>>
{
};

template <class T, template <class> class C = is_container,
          check_constraints<C, T> = true>
void compute(const T& t);

This is far more flexible than the previous option. This flexibility comes at a minor cost: exposing the function to the Python is slightly more verbose:

template <class T, std::size_t N, layout_type L>
struct is_tensor<pytensor<T, N, L>> : std::true_type
{
};

PYBIND11_MODULE(pymod, m)
{
    xt::import_numpy();

    m.def("compute", &compute<pytensor<double, 2>>);
}

Conclusion

Each solution has its pros and cons and choosing one of them should be done according to the flexibility you want to impose on your API and the constraints you are imposed by the implementation. For instance, a method that requires a lot of typing in the bindings might not suit for libraries with a huge number of functions to expose, while a full generic API might be problematic if the implementation expects containers only. Below is a summary of the advantages and drawbacks of the different options:

  • Generic API: full genericity, no additional typing required in the bindings, but maybe too permissive.

  • Full qualified API: simple, accepts only the specified parameter type, but requires a lot of typing for the bindings.

  • Container selection: quite simple, requires less typing than the previous method, but loses type inference on the C++ side and lacks some flexibility.

  • Type restriction with SFINAE: more flexible than the previous option, gets type inference back, but slightly more complex to implement.