Building shared tensorflow_cc about tensorflow_cc

[FloopCZ]

Hi, it seems that the library is trying to be installed to the current working directory. Have you used only cmake .. or have you used something as cmake -DCMAKE_INSTALL_PREFIX=... ?

[rspezialetti]

Hi, i only used cmake -DTENSORFLOW_STATIC=OFF -DTENSORFLOW_SHARED=ON ..

[FloopCZ]

Strange, I am not really sure where this could happen. Can you tell me your operating system version (uname -a) and the exact sequence of commands you have done (i.e., the steps to reproduce the issue)?

[rspezialetti]

Output of uname -a : 4.4.0-96-generic #119-Ubuntu SMP Tue Sep 12 14:59:54 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux

Commande sequence: cmake -DTENSORFLOW_STATIC=OFF -DTENSORFLOW_SHARED=ON ..
make && sudo make install

[dmalkr]

Also have this bug

[FloopCZ]

I still do not know from where does the issue originate. :-( Can you try whether it also happens in the case of building the static library?
(just cmake .., please remove the whole build folder before doing it)
I will rebuild the Dockerfile tests to find out whether it is not because of a new version of some package.

[FloopCZ]

I have made some improvements to the library. Can you please make a clean clone of the project and try over? Beware that the list of requirements has slightly changed, now g++-6 is used instead of g++-5, see the README.md.

In the new version, all tests are green and TensorFlow 1.4.0-rc0 is used.

[sjdrc]

I'm encountering this error after a clean clone and working through the issues in #27. I actually get a build failure and need to manually move a line in cuda/cuda_dnn.cc in order to get tensorflow to compile through bazel, but perhaps completing the build with an existing build folder is the issue here? I'll try editing one of your scripts to get the change to happen automatically.

[sjdrc]

So the issue I have with cuda/cuda_dnn.cc is resolved by applying this fix tensorflow/tensorflow#13255.

I added the line

COMMAND printf %s\\n 235m247 w q | ed -s tensorflow/stream_executor/cuda/cuda_dnn.cc

After the other COMMAND lines in tensorflow_cc/cmake/TensorflowBase.cmake.

[zhenglaizhang]

Hi, have you solved the issue? currently I also met with the issue as below:

Target //tensorflow:libtensorflow_cc.so up-to-date:
  bazel-bin/tensorflow/libtensorflow_cc.so
INFO: Elapsed time: 14.921s, Critical Path: 5.78s
cp: cannot copy a directory, '/tmp/repo/tensorflow_cc/tensorflow_cc/build/tensorflow/tensorflow/contrib/makefile/gen/protobuf', into itself, './tensorflow/contrib/makefile/gen/protobuf/protobuf'
make[2]: *** [CMakeFiles/tensorflow_shared.dir/build.make:112: tensorflow-stamp/tensorflow_shared-configure] Error 1
make[1]: *** [CMakeFiles/Makefile2:68: CMakeFiles/tensorflow_shared.dir/all] Error 2

My Os info:

Linux workstation-p50 4.13.9-1-ARCH #1 SMP PREEMPT Sun Oct 22 09:07:32 CEST 2017 x86_64 GNU/Linux

And since tensorflow 1.4 is out, so I updated the version to 1.4

[FloopCZ]

Hi, I have not been able to reproduce the issue yet. It looks as if the CMake install prefix is set to the current working directory instead of /usr/local. Can you please try to configure the project using

cmake -DTENSORFLOW_STATIC=OFF -DTENSORFLOW_SHARED=ON -DCMAKE_INSTALL_PREFIX=/usr/local ..

That is, explicity setting the install prefix? Thanks.

[hyperdrive]

@FloopCZ Explicitly adding CMAKE_INSTALL_PREFIX solved the the "cannot copy a directory" on Ubuntu 16 for me.

[FloopCZ]

@hyperdrive Thank you for the information, I will have to put this information to the README.md. I still wonder how is it possible that for some installations, the default cmake install prefix is set to the current workdir, though.

[FloopCZ]

I have finally found and fixed the real source of the issue in the latest master.