Tensorflow

System information

OS - Sierra 10.12
Tensorflow - 1.7.0
Xcode command line tools - 8.2 (Download from here: Xcode - Support - Apple Developer & Switch to different clang version: sudo xcode-select --switch/Library/Developer/CommandLineTools & check version: clang -v)
Cmake - 3.11
Bazel - 0.11.1
CUDA - 8.0.61
cuDNN - 7

Requirements

sudo pip install six numpy wheel
brew install coreutils
brew install gcc

We install gcc for libgomp. You may need to add a link to it, i. e. ln -s /usr/local/Cellar/gcc/7.2.0_1/lib/gcc/7/libgomp.dylib /usr/local/lib/libgomp.dylib

Patching Tensorflow

You can check out my MacOS GPU branch at https://github.com/MatthiasWinkelmann/tensorflow, but it may not always be up-to-date with tensorflow. The basic changes are:

Replace all occurances of align(sizeof(T)) with align(sizeof(T) > 16 ? sizeof(T) : 16)

This currently affects the following files:

tensorflow/core/kernels/bucketize_op_gpu.cu.cc
tensorflow/core/kernels/concat_lib_gpu_impl.cu.cc
tensorflow/core/kernels/depthwise_conv_op_gpu.cu.cc
tensorflow/core/kernels/split_lib_gpu.cu.cc

For example, extern shared __align(sizeof(T))__ unsigned char smem[]; => extern shared __align__(sizeof(T) > 16 ? sizeof(T) : 16) unsigned char smem[];

constexpr

Change:

tensorflow/core/framework/variant.h:

class Variant {
public:
 //constexpr Variant() noexcept = default;
 Variant() noexcept = default;

Steps:

Run ./configure
Add following paths:
- export CUDA_HOME=/usr/local/cuda
- export DYLD_LIBRARY_PATH=/usr/local/cuda/lib
bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
pip install /tmp/tensorflow_pkg/*.whl (File name depends on tensorflow version and python version)

MatthiasWinkelmann/Tensorflow_Build_GPU.md