Wolf Vollprecht

Hi @conjam, can you give me some more context on the slowdown, and especially your matrix / vector sizes? If you have small matrices, it's very possible that hand-written code...

I think there are two differences with eigen: - we always call BLAS, where Eigen has it's own implementation that's optimized better for small matrices - when you do a...

As I mentioned before, the dot product is more general than Eigen's matrix / matrix-vector multiplication as it performs broadcasting. Without broadcasting checks you'd probably already be 2x faster for...

MKL won't help much I am afraid :)

calling into BLAS functions has some intrinsic overhead (since even a function call vs. inlined cost has some cost attached to it).

comparing `MatrixXd` and `xt::xarray` is never a fair comparison. Quite different containers (since xarray is dynamically ndimensional, MatrixXd is statically 2d).

you can shave off some more ns by doing `xt::noalias(r) = a + b;` There might be some similar trick for eigen.

One problem is that we probably *don't* want to specify which BLAS library to link against from the xtensor-blas target. So that the user can easily override it.

yes, we plan to support this functionality by taking advantage of views. we want (and already have prototyped) a view-iterator. The iterator will work something like this: auto [it_start, it_end]...

This might require some discussion. Should we specialize the norms for certain types so that they are handled by BLAS vs. xtensor? Or should blas norms just live in the...