xt::linalg::dot() is not working with xarray_adaptor<> when shape is 2x2

[potpath]

This is shown in the following code example.

#include "xtensor/xadapt.hpp"
#include "xtensor-blas/xlinalg.hpp"

int main()
{
        std::vector<double> vec{0, 1, 2, 3};
        std::vector<size_t> shape{2, 2};

        auto arr = xt::adapt(
                vec,
                shape
        );

        auto arr2 = xt::adapt(
                vec.data(),
                vec.size(),
                xt::no_ownership(),
                shape,
                std::vector<int>{2, 1}
        );

        std::cout << arr << std::endl;
        std::cout << xt::linalg::dot(arr, arr) << std::endl;
        std::cout << arr2 << std::endl;
        std::cout << xt::linalg::dot(arr2, arr2) << std::endl;

        return 0;
}

Output:

{{ 0.,  1.},
 { 2.,  3.}}
{{  2.,   3.},
 {  6.,  11.}}
{{ 0.,  1.},
 { 2.,  3.}}
terminate called after throwing an instance of 'std::runtime_error'
  what():  No valid layout chosen.

This is caused by the dynamic layout_type and get_leading_stride() in https://github.com/QuantStack/xtensor-blas/blob/master/include/xtensor-blas/xblas_utils.hpp#L129.

[wolfv]

Hi @potpath indeed. We currently don't have support for dynamic layout in xtensor-blas. Also, it would be quite hard (potentially)...

We could check if the dynamic layout is in fact row- or column major (we actually do this in the view code in xtensor).
If it is not, we could allocate a new container, and store the data with the default layout (however, this will have bad performance characteristics).

I think we should also make it possible to pass a layout parameter to the adapt function you're using (as last argument), I'll open a bug report on xtensor.