Optimize tensorflow for your MacBook by using SSE3, SSE4.1 SSE4.2 AVX, AVX2 and FMA instructions.

" The TensorFlow library wasn't compiled to use X instructions, but these are available on your machine and could speed up CPU computations."

I've been using tensorflow for a while now and got some messages about speeding up the CPU-computations using SSE3, SSE4.1 SSE4.2 AVX, AVX2 and FMA instructions. In short these are extensions to the x86 instruction set architecture for microprocessors from Intel and AMD and contain optimized instructions for processing matrix or vector operations (which of course is good for tensorflow). The pre-built versions of tensorflow have a wide compatibility which means disabling most of these optional speedups. Hence we will build tensorflow from source in a Conda environment.

Reference to tensorflows guide: https://www.tensorflow.org/install/install_sources

I used my MacBook Pro late 2013 with an 2 GHz Intel Core i7.

To check if your CPU supports the instructions run sysctl machdep.cpu.features and look for the SSE/FMA instructions.

sysctl -a | grep machdep.cpu.leaf7_features and look for the AVX instructions.

What you'll need installed

Xcode

Should be installed, if not it's available from the App Store.

Python

I was using the Anaconda distribution of Python 3.6.2

Homebrew

ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"

Java

Available at oracle

Bazel

Used to compile and package tensorflow. Install by running brew install bazel

Python dependencies

• six
• numpy
• wheel, to manage python compressed packages in .whl format.

conda install six numpy wheel

Install Tensorflow

Create a Conda environment and activate it.

conda create -n tensorflow

source activate tensorflow

Checkout tensorflow from github.

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

Enter local copy

cd tensorflow

Configure build

./configure

for user default just press enter for all.

Build the pip package

bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-msse4.2 -k \ //tensorflow/tools/pip_package:build_pip_package

this will take about 40 minutes.

The build command built a script named build_pip_package. Next we build a .whl file within the /tmp/tensorflow_pkg folder.

bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

Install the pip package, the filename of the .whl file depends on the platform but there should only be one file in the directory so just use tab to autocomplete.

sudo pip install /tmp/tensorflow_pkg/<Your file name>.whl

Next we have to test that tensorflow have been installed correctly. You have to cd out of the tensorflow directory otherwise you'll get a weird error.

python

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))

If the systems outputs Hello, TensorFlow! everything should be correct.

Using tensorflow in a jupyter notebook.

Since tensorflow only exists in the environment you have to been in it to run tensorflow. Hence to use a Jupyter Notebook it also has to be installed inside the environment.

To activate and deactivate the environment use:

source activate tensorflow and source deactivate tensorflow