Hi William, Eugene Zhulenev's work on the new BlockV2 evaluation framework, should make it a lot easier to address some of the performance issues you are seeing, including adding block evaluation for concat. So far, TensorFlow as been using a custom concat operation that does not use Eigen, so we never thought to address this. But if we can make it fast in Eigen and get rid of custom code for this op, we will.
Rasmus On Sat, Oct 19, 2019 at 6:52 PM William Tambellini <[email protected]> wrote: > Ok for me. Specially if that could ease to resolve the speed issues of the > Tensor module. > Kind > WT. > > <http://bit.ly/2LB1qRN> > *Share yourfeedback with us* <https://www.surveymonkey.com/r/PYF190816> > > ------------------------------ > *From:* Christoph Hertzberg <[email protected]> > *Sent:* Friday, October 18, 2019 10:03 AM > *To:* [email protected] <[email protected]> > *Subject:* Re: [eigen] Should we require c++11 for Eigen tensor > > Hi! > > I think some time ago there was some opposition against that step -- in > fact, originally C++03 was only introduced after the first versions of > the Tensor module. > > I can't find a reference to that discussion at the moment, I think it > concerned supporting Tensor for some GPU architecture. But this was a > few years ago, and compilers likely evolved since then. > > So unless there is an outcry on this list, I'm not against dropping > C++03 support of the Tensor module. > > > For transitioning it would of course be nice to have a clear "latest > working" version which still supports C++03, one possibility would be to > make that with the 3.4 release. OTOH, this will just complicate > maintenance and Tensors are still "unsupported" for a reason (e.g., not > having to spend unnecessary time to support them ...) > > So just making a clear cut before 3.4 sounds actually better -- this > will also simplify to maintain the upcoming 3.4 and master versions in > parallel (I guess the 3.3.x version of the Tensor module has barely been > touched in recent years). > > > Cheers, > Christoph > > On 18/10/2019 18.46, Rasmus Munk Larsen wrote: > > Dear Eigen tensor users, > > > > Today, various c++11 features are in use in many parts of Eigen tensor, > > while other parts compile with c++03 without warnings. This division is > > rather arbitrary and primarily dictated by what parts of the code gets > > compiled for the subset of tests that are enabled in c++03 mode: > > > > > https://bitbucket.org/eigen/eigen/src/default/unsupported/test/CMakeLists.txt#lines-112 > > > > A larger set of tests covering the majority of the code base are only > > compiled in c++11 mode: > > > https://bitbucket.org/eigen/eigen/src/c4368f48fef3737ef5a48604cfc63ce946b68616/unsupported/test/CMakeLists.txt#lines-179 > > > > I don't think a user can realistically use Eigen tensor and avoid the > c++11 > > parts of the code. Therefore I would propose that we stop pretending that > > Eigen tensor supports c++03 and simply guard all the code with > > EIGEN_HAS_CXX11 or similar. This will then allow us to start cleaning up > > the code. > > > > Please let me know if you have any objections to this. > > > > Best, > > Rasmus > > > > -- > Dr.-Ing. Christoph Hertzberg > > Besuchsadresse der Nebengeschäftsstelle: > DFKI GmbH > Robotics Innovation Center > Robert-Hooke-Straße 5 > 28359 Bremen, Germany > > Postadresse der Hauptgeschäftsstelle Standort Bremen: > DFKI GmbH > Robotics Innovation Center > Robert-Hooke-Straße 1 > 28359 Bremen, Germany > > Tel.: +49 421 178 45-4021 <+49%20421%20178454021> > Zentrale: +49 421 178 45-0 > E-Mail: [email protected] > > Weitere Informationen: http://www.dfki.de/robotik > ------------------------------------------------------------- > Deutsches Forschungszentrum für Künstliche Intelligenz GmbH > Trippstadter Strasse 122, D-67663 Kaiserslautern, Germany > > Geschäftsführung: > Prof. Dr. Jana Koehler (Vorsitzende) > Dr. Walter Olthoff > > Vorsitzender des Aufsichtsrats: > Prof. Dr. h.c. Hans A. Aukes > Amtsgericht Kaiserslautern, HRB 2313 > ------------------------------------------------------------- > > > >
