We use template specializations to make sure we get the correct precision. The problem, for instance, with a non-template f(const xt::xarray<float> &input) is that it will also compile when you pass a xt::xarray<double>. Passing by const-reference in this sense behaves similarly to passing by value; it triggers the creation of a temporary variable -- input in this case -- though in the case of a reference the data is not actually copied.

Using = delete syntax on the default template, e.g.:

template<typename real_t>
xt::xarray< std::complex<real_t> > fft(const xt::xarray<real_t> &input) = delete;

we can make sure that calls to non-implemented specializations don't compile. If this is left out, the compilation will succeed, but the linker will fail, and this gives less informative error messages.

Unfortunately, clang has a bug that prevents us from using the nice = delete syntax. For the moment we're using a workaround:

template<typename real_t>
xt::xarray< std::complex<real_t> > fft(const xt::xarray<real_t> &input) {
   static_assert(sizeof(real_t) == 0, "Only specializations of fft can be used");
};

This is ugly. Once the bug in clang is fixed, we will want to replace this with the = delete syntax.