Using Eigen or Xtensor

[fschlimb]

I wonder why pythran brings its own ndarray instead of using XTensor (which looks like obvious choice since it's already using xsimd) or Eigen's Tensor? Have these come around too late? Or do they have serious limitations/disadvantages?

[Chronum94]

From a conversation I vaguely remember from pythran's IRC: I believe (could be wrong) that preliminary tests showed that pythonic (Pythran's backend) is faster than xtensor.

Plus, a full switch of a backend would be (I believe) a fair amount of work.

[serge-sans-paille]

@fschlimb there's on going work to evaluate the capability of xtensor to sustain pythran's API, see https://github.com/serge-sans-paille/numpy-benchmarks/pull/11 . Unfortunately, as of now, the ownership model of xtensor is incompatible with pythran's, and the performance are not satisfying. That's still something we keep in mind, but you can't expect it in a near future.