Well, if the tensorflow deviation of eigen3 is too large, I think it is allowed to use the embedded eigen3 for the tensorflow.
Anton 2018-08-18 14:53 GMT+02:00 Lumin <cdlumin...@gmail.com>: > On Sat, Aug 18, 2018 at 01:28:18PM +0200, Andreas Tille wrote: >> Hi, >> >> On Fri, Aug 10, 2018 at 01:24:16PM +0000, Lumin wrote: >> > >> > So I'd suggest that upload the 3.3.5 version with the above ppc64el fix. >> >> I commited new upstream source to Git but there are quilt patches that >> do not apply cleanly. I have no time to investigate this more deeply >> and left a note in d/changelog about this work item to be done. Sorry >> for not beeing more helpful - feel free to take over > > Thanks Andreas, Anton is working on newer eigen3, see #906126 > > importing 3.3.5 is not enough for building tensorflow. I need not > only the following function signature in > unsupported/Eigen/CXX11/src/ThreadPool/NonBlockingThreadPool.h > > 25 ThreadPoolTempl(int num_threads, bool allow_spinning, > 26 Environment env = Environment()) > > But bitbucket is very hard to use and I struggled for a long while > with it but I still cannot find out which commit introduced this change. > > > @Anton: > > I think I can make a differential between eigen 3.3.5 and the eigen > snapshot that tensorflow used, and patch the eigen library when building > tensorflow. Since eigen3 is a header only library, this is not hard to > achive. > > I'm able to build libtensorflow.so and the python interface library > with the patch series https://github.com/tensorflow/tensorflow/pull/21699 , > embedded eigen3 and embedded double-precision. > However the python package still misses some python dependencies. > >> >> Andreas. >> >> -- >> http://fam-tille.de