Re: [eigen] Merge Tensor module

[Vincent Vanhoucke <vanhoucke@xxxxxxxxxx>]

Fantastic! Thanks a lot for making it happen.

On Thu, Feb 12, 2015 at 12:51 PM, Gael Guennebaud <gael.guennebaud@xxxxxxxxx> wrote:

The big merge has been bone! Have fun with tensors, and thanks again to Benoit!

gael

On Wed, Feb 11, 2015 at 1:05 PM, Wood, Tobias <tobias.wood@xxxxxxxxx> wrote:

A quick +1 to say I have been looking forward to seeing this in /unsupported so I can try it. Thanks to Benoit for working on it.

Toby

---

From: Gael Guennebaud <gael.guennebaud@xxxxxxxxx>
Sent: 09 February 2015 08:43
To: eigen
Subject: [eigen] Merge Tensor module

 

Hi,

I'm willing to merge Benoit Steiner's Tensor module in unsupported/. See pull-request #96 (https://bitbucket.org/eigen/eigen/pull-request/96).

Feel free to object with arguments.

Basically, it supports the following features:

 - arbitrary tensor rank (up to 5 in C++03)

 - fully-dynamic or fully-fixed dimensions

 - forward or backward storage order (like row-major/column-major)

 - slicing, chipping (pick a R-1 rank sub tensor), reduction,

 - contractions using Eigen's fast matrix-matrix and matrix-vector kernels,

 - reshape, shuffling of the dimensions,

 - convolutions

- etc.

There are also a few features aiming to be integrated in Eigen core. In particular, the evaluation of expressions on CUDA and a fast CUDA matrix-matrix kernel.

Before it can be integrated within Eigen's official modules, we will have to unify everything that can be unified:

 - the evaluator mechanisms (more on that later)

 - unify Tensor and FixedSizeTensor while enabling mixed fixed-dynamic sizes

 - TensorMap -> Map<Tensor>

 - CUDA stuffs (more on that later)

 - move the documentation to Doxygen

 - clean Eigen::array, Eigen::DSizes if possible

 - fix Eigen::IndexList (it currently requires C++14)

 - etc.

and also stabilize the naming convensions (more on that later).

Additional welcome features could be:

 - Ref<Tensor>

 - Iterator based evaluation for large-rank _expression_ with complex indexing

 - cost model (could allow to switch between index-based and iterator-based evaluation

 - unify 1D/2D tensors with MatrixBase?

Nonetheless, I think this is a good timing for the merge as the current implementation is already pretty mature and this will provide the visibility this module deserves.

Cheers,

Gael.