Arch Linux User Repository

Umm, the prepare() function is kind-of invalid. You can't set environment variables from prepare(), I mean, you can set it, but you can't expect them to be visible in build() stage. The build fails when prepare() and build() are run in separate stages - like first makepkg -o and then followed by separate makepkg -e - makepkg -e will cause configure.py ask for some settings, because prepare() was not run. The build most notably fails to build properly when using yay AUR helper.

Consider applying like the following trivial patch to just move variable settings to be inside build():

--- i/PKGBUILD
+++ w/PKGBUILD
@@ -42,7 +42,9 @@ prepare() {

   [ -d ${srcdir}/tensorflow-cuda ] && rm -rf ${srcdir}/tensorflow-cuda
     cp -r ${srcdir}/tensorflow ${srcdir}/tensorflow-cuda
+}

+_set_build_env() {
   # make sure the proxy variables are in all caps, otherwise bazel ignores them
   export HTTP_PROXY=`echo $http_proxy | sed -e 's/\/$//'`
   export HTTPS_PROXY=`echo $https_proxy | sed -e 's/\/$//'`
@@ -96,6 +98,7 @@ prepare() {
 }

 build() {
+  _set_build_env
   cd ${srcdir}/tensorflow
   export CC_OPT_FLAGS="-march=native -O3"
   export TF_NEED_CUDA=0

Thanks!

have to make symlink in /usr/bin

sudo ln -s bazel bazel-3.0.0-linux-x86_64

How to compile it with modern CPU's instructions support ?

Had to downgrade from Bazel 2.0.0 to 1.2.1. Otherwise, it fails with a Python "subprocess" error. This is not caused by the PKGBUILD, however.

Fails when using Bazel 0.13 java.lang.RuntimeException: Unrecoverable error while evaluating node 'PACKAGE:tensorflow/core/kernels'

Worked with Bazel 0.12. Refer to https://github.com/tensorflow/tensorflow/issues/19015

The build fails at boringssl.

Adding --copt=-O to the bazel options fixes the build.
See also: https://stackoverflow.com/questions/40373686/freeze-graph-py-throws-an-error-during-build/40374431#40374431

I can confirm that cudnn is necessary. There an issue on tensorflow repository asking for support cudnn 7 and cuda 9 (https://github.com/tensorflow/tensorflow/issues/12052).

I think cudnn should be cudnn6 now?

python-tensorflow-cuda-git seems to depend on libglvnd which conflicts with my video driver. Is that necessary? tensorflow-cuda-git does not require it.