matmul and dot about xtensor HOT 9 CLOSED

We have a plan to support

• dot
• tensordot
• einstein sums.

The latter with a template-based syntax, instead of strings à la numpy.

The use of SIMD-accelerated versions of blas for certain memory contiguous cases is something we have experience with, but we don't have a general approach for tensor / einstein products.

from xtensor.

@SylvainCorlay any update about this feature?

from xtensor.

@edgarriba The dot and tensordot will probably be added in the version 0.9. We have urgent features to release in the next version (0.8) before we can focus on dots.

SIMD-accelerated support will be added after version 1.0 has been released.

from xtensor.

Also, we make release about every other week. Some of the foundations for how we want to make tensor dots were introduced in 0.7.

from xtensor.

@edgarriba I saw the discussion on tiny-dnn about adopting xtensor. Is this a blocking point for you guys?

from xtensor.

@SylvainCorlay not by the moment since we are still evaluating xtensor as our tensor backend. However, we have some codes were we have implemented dot by hand, so we could use them in the meantime.

from xtensor.

Also about batched interfaces and http://www.netlib.org/utk/people/JackDongarra/WEB-PAGES/Batched-BLAS-2016/Day1/05_Ahmad_vbatched.pdf

from xtensor.

Ongoing work in xtensor-blas for the curios.

from xtensor.

xtensor-blas has been released.

from xtensor.