Jetson AGX Xavier 0.15.0 build requires std=c++14 in the crosstool chain configuration

[girgink]

System information

• Jetson AGX Xavier (ARMv8) custom L4T based on Ubuntu 20.04
• TensorFlow 2.7.0 (source)
• TensorFlow-Addons 0.15.0 (source)
• Python 3.8
• Is GPU used? Yes

Describe the bug Because Jetson AGX Xavier has CUDA 10.2, the system includes build_deps/toolchains/gpu/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc.tpl, which only defines rules for C++11. However TF 2.7.0 uses C++14 (e.g. in tensorflow/core/framework/tensor_types.h, MaybeWith32BitIndexingImpl<Eigen::GpuDevice>.operator()). Therefore build fails.

Code to reproduce the issue

TF_NEED_CUDA="1" \
TF_CUDA_VERSION="10.2" \
TF_CUDNN_VERSION="8" \
CUDNN_INSTALL_PATH="/usr/lib/aarch64-linux-gnu" \
python ./configure.py

bazel build build_pip_pkg

Other info / logs

Replacing all c++11 with c++14 in build_deps/toolchains/gpu/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc.tpl solved the issue. Just in case, I also replaced all c++11 with c++14 in build_deps/toolchains/gpu/crosstool/cc_toolchain_config.bzl.tpl. I see that CUDA 11 configuration files include C++14 rules. It can be good to add them also other CUDA configurations.

[bhack]

Have you checked https://github.com/tensorflow/addons/pull/2605 ?

[girgink]

I didn't see that, thanks.

[bhack]

Let me know if we could close this

[girgink]

Sorry for the delay. Yes #2605 is about the same issue, please feel free to close this.