python-tensorflow-cuda (1.4.1-1) unstable; urgency=medium

  * New upstream release.
  * Merge Matan Goldman's patch to fix build on CUDA 7.5.
    It does not means it's fully functional, but could be.
  * Relax build dependencies to allow CUDA 7.5.
  * Define available CUDA 9.0 compute levels (Debian Buster).

 -- Adam Cecile <acecile@le-vert.net>  Fri, 09 Feb 2018 21:35:46 +0100

python-tensorflow-cuda (1.4.0-1) unstable; urgency=medium

  * New upstream release:
    - Refresh patches,
    - Drop Dont_use_locate_to_find_libdl.so.2 patch, not relevant anymore,
    - Now requires bazel 0.6.1 to build (minimum required is 0.5.4),
    - Modify configure-expect script to enable Amazon S3 (new), GDR (new),
      and XLA just-in-time compiler (not experimental anymore),
    - Add librdmacm-dev build dependency for GDR,
    - Re-order completely configure-expect script because TF upstream changed
      all options order AGAIN,
    - Export Intel MKL DNN folder as environment variable, it's not part
      of configure script anymore and create symlink for license.txt,
    - Make $HOME into a /tmp/temp.stg folder so bazel stops whining,
    - Disable MPI and VERBS on non-CUDA build, it seems to be unsupported.
  * Refactor debian/rules and debian/control to handle nocuda and nomkldnn
    build profiles.
  * Remove useless python wheel runtime dependencies.
  * Tensorflow now ships shared libs instead of monolithic Python/C module,
    these ones should not be renamed by dh_python* as they are not Python
    version dependent.

 -- Adam Cecile <acecile@le-vert.net>  Tue, 21 Nov 2017 19:44:00 +0100

python-tensorflow-cuda (1.3.1-1) unstable; urgency=medium

  * New upstream release:
    - Drop Remove-github-mirror-of-protobuf-due-to-mismatched-sha256-sum
      patch.
  * Drop the weird "external" folder.
  * Disable OpenMPI in < 1.10 (Debian Jessie, FTBFS).

 -- Adam Cecile <acecile@nsc-srve.service.emergency.lu>  Thu, 28 Sep 2017 21:19:07 +0000

python-tensorflow-cuda (1.3.0-1) unstable; urgency=medium

  * New upstream release:
    - Drop Fix_CUDA_7.x_build, merged upstream,
    - Build depends on OpenMPI and build conflicts on mpich and lam,
    - Build with MPI support by passing -DOMPI_SKIP_MPICXX=1,
    - Merge upstream 82bb560cf437ad47c8acdd010f143a7c8810f803 to avoid
      build failure due to github archives md5 mismatch,
    - Add -Wno-error=stringop-overflow to fix FTBFS on Debian Buster when
      building without CUDA (otherwise using GCC 4.9).
    - Add severals new workarounds because configure script is getting
      worse (mostly related to CUDA and sub-components detection),
    - Fix dependencies: tensorboard is now an external package,
    - Now rename Python C extentions (.so files) myself to support both
      Python 3.5 and 3.6 builds for Debian Buster \o/,
    - Make debian/rules simplier, using eval to define intermediate
      variables in targets,
    - Adapt list of Python dependencies, according to new setup.py.

 -- Adam Cecile <acecile@le-vert.net>  Thu, 21 Sep 2017 06:33:35 +0000

python-tensorflow-cuda (1.2.1-2) unstable; urgency=medium

  * Provide additional builds using Intel(R) Math Kernel Library
    for Deep Neural Networks.
    
 -- Adam Cecile <acecile@le-vert.net>  Fri, 15 Sep 2017 00:20:59 +0200

python-tensorflow-cuda (1.2.1-1) unstable; urgency=medium

  * Initial release.

 -- Adam Cecile <acecile@le-vert.net>  Thu, 03 Aug 2017 15:01:36 +0200
