这几天终于把tensorflow安装上了，中间遇到过不少的问题，这里记录下来。供大家想源码安装的参考。
安装环境：POWER8处理器，Docker容器Ubuntu14.04镜像。
Build Bazel<v0.1.0> for IBM POWER8 CPU from Source Code
My computer's os is ubuntu 14.04 , and I want to install bazel, There is only java7 available, so I choose bazel 0.1.0.
Here are the steps that I succeed to install bazel 0.1.0 on my computer:
1. Get source code
git clone https://github.com/bazelbuild/bazel.git
git checkout 0.1.0
2. Install protobuf
Check which version is needed:
opuser@nova:~/tensorflow/bazel-0.1.0$ ls third_party/protobuf/
LICENSE protobuf-java-3.0.0-alpha-3.jar protoc-linux-x86_64.exe protoc-osx-x86_32.exe protoc-windows-x86_32.exe
README.md protoc-linux-x86_32.exe protoc-linux-x86_64.real protoc-osx-x86_64.exe protoc-windows-x86_64.exe
protobuf-java-3.0.0-alpha-3.jar is the jar file bazel uses. so we need to install protobuf with version <3.0.0-alpha-3> , you can refer to "Build Protobuf<v3.0.0-alpha-3> for IBM POWER8 CPU from Source Code" to build protobuf .
3. Create a link protoc-linux-x86_64.exe to protoc binary file
such as : ln -s /usr/bin/protoc $(pwd)/ third_party/protobuf/protoc-linux-x86_64.exe
4. Install other dependencies
4.1 sudo apt-get install zlib1g-dev zip unzip openjdk-7-jre openjdk-7-jdk
4.2 add java path to ~/.bashrc
export JAVA_HOME=/usr/lib/jvm/java-7-openjdk-ppc64el
export JRE_HOME=${JAVA_HOME}/jre
export CLASSPATH=.:${JAVA_HOME}/lib:${JRE_HOME}/lib
export PATH=${JAVA_HOME}/bin:$PATH
5. Modify bazel source code to support POWER cpu
5.1 in src/main/java/com/google/devtools/build/lib/util/CPU.java
in enum "public enum CPU" add
POWER("power", ImmutableSet.of("power", "power64le", "power64be", "ppc")),
5.2 in /src/main/java/com/google/devtools/build/lib/analysis/config/BuildConfiguration.java
in function "public static class HostCpuConverter implements Converter<String>"add:
case POWER:
return "power";
5.3 in tools/cpp/CROSSTOOL, add
default_toolchain {
cpu: "power"
toolchain_identifier: "local_linux"
}
5.4 in third_party/BUILD, add
config_setting(
name = "power",
values = {"host_cpu": "power"},
)
5.5 in scripts/bootstrap/buildenv.sh, add
MACHINE_IS_64BIT='yes'
6. Compile bazel
./compile // better to see step 7 before run ./compile
after a long time, a message displayed:
Build successful! Binary is here: /home/opuser/tensorflow/tools/bazel-0.1.0/output/bazel
7. Problems during compiling
It is not easy to build bazel from source code. not to mention that I build it in a docker container. there are mainly two problems.
7.1 namespace-sandbox
<error: mount() permission denied!>
bazel may choose to build a sanbox which uses namespace . but there are priority strictions in container. to avoid this problem, you can choose not to build sandbox .
I refer to dcreager's solution posted at https://github.com/bazelbuild/bazel/issues/418. create a bazelrc and use command "BAZELRC=/tmp/bazelrc ./compile.sh" to compile the code. the bazelrc file contains:
startup --batch
build --spawn_strategy=standalone --genrule_strategy=standalone
7.2 openjdk-7
<error: unrecognized option -client!>
The java installed on my computer does not support -client option, but -server option is ok. just as the issue opened at
https://github.com/bazelbuild/bazel/issues/56
Although there is a commit about this:
https://github.com/bazelbuild/bazel/commit/1fa163c42ee26a72c21c118ec820d677a431ee49#commitcomment-13207150.
But it does not work to me. so I changed all the -client option to -server.
8. Reference
Can't run Bazel as part of a Docker build due to sandboxing issues
https://github.com/bazelbuild/bazel/issues/418
Compiling the Java example with SAP JVM does not work
https://github.com/bazelbuild/bazel/issues/56
Add jvm_opts attribute to the java_toolchain rule
https://github.com/bazelbuild/bazel/commit/1fa163c42ee26a72c21c118ec820d677a431ee49#commitcomment-13207150.