[Done] Upgrade CUDA 8.0 in Docker

[gangliao]

Fix #1695

[helinwang]

@gangliao 驱动的问题可以先试试nvidia-docker，如果nvidia-docker也不行的话，应该是我们编译／镜像的问题。如果可以，估计是运行方法的问题。

[helinwang]

我做了如下命令：

sudo rm -rf build third_party
docker build --no-cache -t paddle-cuda-8-new .
docker run -v $PWD:/paddle -e "WITH_GPU=ON" -e "WITH_AVX=ON" paddle-cuda-8-new
docker build -t paddle-cuda-8 -f build/Dockerfile . 
nvidia-docker run -it --entrypoint bash paddle-cuda-8

仍然有

Traceback (most recent call last):
  File "mnist.py", line 1, in <module>
    import paddle.v2 as paddle
  File "/usr/local/lib/python2.7/dist-packages/paddle/v2/__init__.py", line 14, in <module>
    import optimizer
  File "/usr/local/lib/python2.7/dist-packages/paddle/v2/optimizer.py", line 1, in <module>
    import py_paddle.swig_paddle as swig_api
  File "/usr/local/lib/python2.7/dist-packages/py_paddle/__init__.py", line 15, in <module>
    from util import DataProviderWrapperConverter
  File "/usr/local/lib/python2.7/dist-packages/py_paddle/util.py", line 18, in <module>
    import swig_paddle
  File "/usr/local/lib/python2.7/dist-packages/py_paddle/swig_paddle.py", line 28, in <module>
    _swig_paddle = swig_import_helper()
  File "/usr/local/lib/python2.7/dist-packages/py_paddle/swig_paddle.py", line 24, in swig_import_helper
    _mod = imp.load_module('_swig_paddle', fp, pathname, description)
ImportError: libcudart.so.7.5: cannot open shared object file: No such file or directory

git HEAD是最新的，在

commit 929090ed2290b25a1d08959e15a3d6e7a05e1862
Merge: 5961b52 df9993f
Author: QI JUN <qijun1994@hotmail.com>
Date:   Fri Mar 24 18:12:30 2017 +0800

    Merge pull request #1702 from gangliao/python_remove
    
    Fix NOT WITH_PYTHON

git diff:

diff --git a/Dockerfile b/Dockerfile
index ccd43be..c4502e8 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,6 +1,6 @@
 # A image for building paddle binaries
 # Use cuda devel base image for both cpu and gpu environment
-FROM nvidia/cuda:7.5-cudnn5-devel-ubuntu14.04
+FROM nvidia/cuda:8.0-cudnn5-devel-ubuntu14.04
 MAINTAINER PaddlePaddle Authors <paddle-dev@baidu.com>
 
 ARG UBUNTU_MIRROR
diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh
old mode 100644
new mode 100755
index a0da561..5be0eb5
--- a/paddle/scripts/docker/build.sh
+++ b/paddle/scripts/docker/build.sh
@@ -4,7 +4,7 @@ set -e
 
 # Set BASE_IMAGE according to env variables
 if [ ${WITH_GPU} == "ON" ]; then
-  BASE_IMAGE="nvidia/cuda:7.5-cudnn5-runtime-ubuntu14.04"
+  BASE_IMAGE="nvidia/cuda:8.0-cudnn5-runtime-ubuntu14.04"
   # additional packages to install when building gpu images
   GPU_DOCKER_PKG="python-pip python-dev"
 else

@gangliao 能否再看一下／按我贴上来的的命令也试一下？

[helinwang]

LGTM

[typhoonzero]

LGTM.应该是没问题了，暂时没有安装了cuda8的机器，在7.5的机器运行会输出：nvidia-docker | 2017/03/27 13:47:47 Error: unsupported CUDA version: driver 7.5 < image 8.0.61 说明已经正确识别了驱动了。

[gangliao]

@helinwang @typhoonzero 可以访问试试 nvidia-docker run -it --entrypoint bash paddle-cuda-8-gang

[helinwang]

我这里试了一下，还需要改一下cmake/cudnn.cmake:

diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake
index e5b59be..c31f6c5 100644
--- a/cmake/cudnn.cmake
+++ b/cmake/cudnn.cmake
@@ -11,6 +11,7 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
     ${CUDNN_ROOT}
     ${CUDNN_ROOT}/lib64
     ${CUDNN_ROOT}/lib
+    ${CUDNN_ROOT}/lib/x86_64-linux-gnu
     $ENV{CUDNN_ROOT}
     $ENV{CUDNN_ROOT}/lib64
     $ENV{CUDNN_ROOT}/lib

[helinwang]

LGTM++