GeForce GTX 1080 + CUDA 8.0 + Ubuntu 16.04 + Caffe/TensorFlow

Install cuda 8.0 referring here.

As the default gcc in Ubuntu 16.04 is very new, if you have compiling error similar to #error -- unsupported GNU version! gcc versions later than 5.3 are not supported!

Try to uncomment the #error line in file /usr/local/cuda/include/host_config.h

#if __GNUC__ > 5 || (__GNUC__ == 5 && __GNUC_MINOR__ > 3)

//#error -- unsupported GNU version! gcc versions later than 5.3 are not supported!

#endif /* __GNUC__ > 5 || (__GNUC__ == 5 && __GNUC_MINOR__ > 1) */

Update nvidia driver

You may need to remove the old driver version installed with cuda-8.0 and install a newer version of driver, if you have error similar to modprobe: ERROR: could not insert 'nvidia_361_uvm': Invalid argument

Please refer here to remove old driver, then install a new one:

sudo apt-get purge nvidia-*
dkms status
sudo dkms remove bbswitch/0.8 -k 4.4.0-31-generic #do this based on your `dkms status`
sudo ./ # download the compatible driver (GTX 1080 in my case) from nvidia website and install it
sudo reboot # reboot if the driver was not updated
./cuda8-0-samples/bin/x86_64/linux/release/deviceQuery # compile the sample code and check if it works

Install cudnn V5

cudnn v4 is NOT supported in 1080. You may compile and run, but will not converge.

Install prerequisities of Caffe

Because the gcc version is Ubuntu 16.04 is very new, If any prerequisity installed from apt-get does not work, uninstall it, compile and install it by the default gcc (5.4) from the source code. The prerequisities that may have problems include: protobuf and opencv. E.g. if you have protobuf error similar to

.build_release/lib/ undefined reference to `google::protobuf::io::CodedOutputStream::WriteVarint64ToArray(unsigned long long, unsigned char*)'

Try to uninstall the protobuf installed by apt-get, the one installed by apt-get might have been compiled by a older gcc version, so its shared libraries may not be compatible with your default gcc:

sudo apt-get purge libprotobuf-dev protobuf-compiler

then compile the protobuf-2.5.0 from src and install it. Please config the default gcc (5.4 in my case) when you compile protobuf:

./configure --prefix=/your/path/ CC=/usr/bin/gcc
make check
make install

 Compile and test Caffe here.

Please also refer here.

If you use anaconda, and get error like awk: symbol lookup error: $HOME/anaconda2/lib/ undefined symbol: PC

try to remove readline in anaconda lib so as to use the default system one.

conda remove --force readline

Compile and test TensorFlow here.

If there is error like ERROR: /mnt/tmp/tensorflow/tensorflow/core/BUILD:87:1: //tensorflow/core:protos_all_py: no such attribute 'imports' in 'py_library' , newer version (e.g. 0.2.3) of bazel can solve this problem.

git clone
cd bazel/
git tag -l
git checkout tags/0.2.3 # or 0.3.0 etc.

“This will create a bazel binary in bazel-bin/src/bazel. This binary is self-contained, so it can be copied to a directory on the PATH (e.g., /usr/local/bin) or used in-place. ”

Use which bazel to make sure your bazel is updated.

Then build and install:

bazel build -c opt --config=cuda //tensorflow/cc:tutorials_example_trainer
# test
bazel-bin/tensorflow/cc/tutorials_example_trainer --use_gpu

# build pip package with gpu support
bazel build -c opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

bazel-bin/tensorflow/tools/pip_package/build_pip_package ~/tmp/tensorflow_pkg

# The name of the .whl file will depend on your platform.
sudo pip install ~/tmp/tensorflow_pkg/tensorflow-0.9.0-py2-none-any.whl

mkdir _python_build
cd _python_build
ln -s ../bazel-bin/tensorflow/tools/pip_package/build_pip_package.runfiles/* .
ln -s ../tensorflow/tools/pip_package/* .
python develop

cd tensorflow/models/image/mnist

Known issues

  • TensorFlow compiling @ RHEL
ERROR: /home/wwen/github/tensorflow/tensorflow/core/kernels/BUILD:1529:1: undeclared inclusion(s) in rule '//tensorflow/core/kernels:depth_space_ops_gpu':
this rule is missing dependency declarations for the following files included by 'tensorflow/core/kernels/':

Solution: add cuda include path to third_party/gpus/crosstool/CROSSTOO by adding a line similar to cxx_builtin_include_directory: "/usr/local/cuda-7.5/include"

  • bazel compiling @ RHEL

If no JAVA_HOME is found when installing bazel from source code, install java sdk and make sure the default one is the one you want

sudo yum install java-1.8.0-openjdk-devel
/usr/sbin/alternatives --config java

Add JAVA_HOME variables in ~/.bashrc:

export JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk
export PATH=$PATH:$JAVA_HOME/bin
  • TensorFlow WORK_DIR issue
bazel-bin/inception/download_and_preprocess_imagenet "/fdata/imagenet-data"

And get error: bazel-bin/inception/download_and_preprocess_imagenet: line 66: bazel-bin/inception/download_and_preprocess_imagenet.runfiles/inception/data/ No such file or directory, change WORK_DIR in ./inception/data/ to WORK_DIR="$0.runfiles/__main__/inception"