Compile Tensorflow

Tensorflow packages often assume the processor supports avx2. Our Xeon process does not, and requires tensorflow to be compiled without this flag.

Verify dependencies

sudo apt install python3-dev python3-pip

Install Dependencies

sudo pip3 install numpy

Install Bazel using Bazelisk

sudo npm install -g @bazel/bazelisk

Checkout Tensorflow

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow

Configure

Keep all defaults

./configure

Compile

The jobs parameter may be needed to reduce memory dependencies on some servers

bazel build --jobs 2 //tensorflow/tools/pip_package:build_pip_package

I got compilation errors at one point and some guidance said to add the following to the command to reduce memory consumption:

--local_ram_resources 2048

or

bazel build --jobs 2 --local_ram_resources 2048 //tensorflow/tools/pip_package:build_pip_package

Almost at the end, I got a compile error. I fixed it with this suggestion:

  from tensorflow/core/kernels/pad_op.cc:20:
external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h: In member function 'Eigen::TensorEvaluator<const Eigen::TensorPaddingOp<PaddingDimensions, XprType>, Device>::PacketReturnType Eigen::TensorEvaluator<const Eigen::TensorPaddingOp<PaddingDimensions, XprType>, Device>::packetRowMajor(Eigen::TensorEvaluator<const Eigen::TensorPaddingOp<PaddingDimensions, XprType>, Device>::Index) const [with PaddingDimensions = const Eigen::array<Eigen::IndexPair<long int>, 2ul>; ArgType = const Eigen::TensorMap<Eigen::Tensor<const std::complex<float>, 2, 1, long int>, 16, Eigen::MakePointer>; Device = Eigen::ThreadPoolDevice]':
external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorPadding.h:688:7:internal compiler error: in emit_move_insn, at expr.c:3547values[i] = coeff(index+i);

With this workaround, you can compile tensorflow with gcc 6.3

In external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h:801
(you can find the file in a subdirectory of ~/.cache/bazel )

replace

values[i] = internal::InnerMostDimReducer<Self, Op>::reduce(*this, firstIndex + i * num_values_to_reduce,
num_values_to_reduce, reducer);

by two instructions instead

Self::CoeffReturnType a = internal::InnerMostDimReducer<Self, Op>::reduce(*this, firstIndex + i * num_values_to_reduce,
num_values_to_reduce, reducer);
values[i] = a;