Intel® Optimization for TensorFlow* Installation Guide

Published: 08/09/2017  

Last Updated: 07/26/2021

By PREETHI VENKATESH, Jing Xu, and Hung-Ju Tsai

TensorFlow* is a widely-used machine learning framework in the deep learning arena, demanding efficient utilization of computational resources. In order to take full advantage of Intel® architecture and to extract maximum performance, the TensorFlow framework has been optimized using oneAPI Deep Neural Network Library (oneDNN) primitives, a popular performance library for deep learning applications. For more information on the optimizations as well as performance data, see this blog post TensorFlow* Optimizations on Modern Intel® Architecture .

Anaconda* has now made it convenient for the AI community to enable high-performance-computing in TensorFlow. Starting from TensorFlow v1.9, Anaconda has and will continue to build TensorFlow using oneDNN primitives to deliver maximum performance in your CPU.

This install guide features several methods to obtain Intel Optimized TensorFlow including off-the-shelf packages or building one from source that are conveniently categorized into Binaries, Docker Images, Build from Source

For more details of those releases, users could check Release Notes of Intel Optimized TensorFlow.

Now, Intel Optimization for Tensorflow is also available as part of Intel® AI Analytics Toolkit. Download and Install to get separate conda environments optimized with Intel's latest AI accelerations. Code samples to help get started with are available here.

The oneAPI Deep Neural Network Library (oneDNN) optimizations are also now available in the official x86-64 TensorFlow after v2.5. Users can enable those CPU optimizations by setting the the environment variable TF_ENABLE_ONEDNN_OPTS=1 for the official x86-64 TensorFlow after v2.5. There is a comparison table between those two releases in the additional information session.

Supported Installation Options

NOTE : Users can start with pip wheel installation from Intel Channel if no preference.

Anaconda 

  • Linux: Main Channel (v2.5) | Intel Channel (v2.5) | Intel AI Analytics Toolkit (v2.5)
  • Windows: Main Channel (v2.5)
  • MacOS: Main Channel (v2.0)

PIP Wheels

  • Linux: Py37 | Py38 | Py39 (v2.7)
  • Windows: Py37 | Py38 | Py39 (v2.7)

Docker Containers

  • Linux: Intel containers (v2.6) | Google DL containers (v2.7)

Build from source

  • Linux | Windows

 

Installation Options

1. Binaries

Get Intel® Optimization for TensorFlow* Pre-Built Images

Available for Linux*, Windows*, MacOS*

OS TensorFlow* version
Linux* 2.5.0
Windows* 2.5.0
MacOS* 2.0.0

 

Installation instructions:

If you don't have conda package manager, download and install Anaconda

Linux and MacOS

Open Anaconda prompt and use the following instruction

conda install tensorflow

In case your anaconda channel is not the highest priority channel by default(or you are not sure), use the following command to make sure you get the right TensorFlow with Intel optimizations

conda install tensorflow -c anaconda

Windows

Open Anaconda prompt and use the following instruction

conda install tensorflow-mkl

(or)

conda install tensorflow-mkl -c anaconda

Besides the install method described above, Intel Optimization for TensorFlow is distributed as wheels, docker images and conda package on Intel channel. Follow one of the installation procedures to get Intel-optimized TensorFlow.

 

Note: All binaries distributed by Intel were built against the TensorFlow version tags in a centOS container with gcc 4.8.5 and glibc 2.17 with the following compiler flags (shown below as passed to bazel*)

--cxxopt=-D_GLIBCXX_USE_CXX11_ABI=0 --copt=-march=corei7-avx --copt=-mtune=core-avx-i --copt=-O3 --copt=-Wformat --copt=-Wformat-security --copt=-fstack-protector --copt=-fPIC --copt=-fpic --linkopt=-znoexecstack --linkopt=-zrelro --linkopt=-znow --linkopt=-fstack-protector

Available for Linux*

TensorFlow* version: 2.5

Installation instructions:

Open Anaconda prompt and use the following instruction. Available for Python 3.7, 3.8 and 3.9.

conda install tensorflow -c intel

Available for Linux*

TensorFlow* version: 2.5

Installation instructions:

There are multiple options provided to download Intel® AI Analytics Toolkit, including Conda, online/offline installer, repositories and containers.

All available download and installation guides can be found here

Available for Linux* and Windows*

TensorFlow version: 2.7

Installation instructions:

Note: For TensorFlow versions 1.13, 1.14 and 1.15 with pip > 20.0, if you experience invalid wheel error, try to downgrade the pip version to < 20.0

For e.g

python -m pip install --force-reinstall pip==19.0

Run the below instruction to install the wheel into an existing Python* installation. Python versions supported are  3.7, 3.8, 3.9

pip install intel-tensorflow==2.7.0

If your machine has AVX512 instruction set supported please use the below packages for better performance.

pip install intel-tensorflow-avx512==2.7.0 # linux only

Pip packages are posted on Google Cloud and AWS for easy access to customers.

Python Version Minimum instruction set required Command with wheels from Google Cloud Storage

python 3.7

AVX

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow-2.7.0-cp37-cp37m-manylinux2010_x86_64.whl

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow-2.7.0-cp37-cp37m-win_amd64.whl

 

AVX-512

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow_avx512-2.7.0-cp37-cp37m-manylinux2010_x86_64.whl

python 3.8

AVX

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow-2.7.0-cp38-cp38-manylinux2010_x86_64.whl

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow-2.7.0-cp38-cp38-win_amd64.whl

 

AVX-512

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow_avx512-2.7.0-cp38-cp38-manylinux2010_x86_64.whl

python 3.9

AVX

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow-2.7.0-cp39-cp39-manylinux2010_x86_64.whl

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow-2.7.0-cp39-cp39-win_amd64.whl

 

AVX-512

pip install https://storage.googleapis.com/intel-optimized-tensorflow/2.7.0/intel_tensorflow_avx512-2.7.0-cp39-cp39-manylinux2010_x86_64.whl



Note: If your machine has AVX-512 instruction set supported, please download and install the wheel file with AVX-512 as minimum required instruction set from the table above.

Note: If you ran into the following Warning on ISA above AVX2, please download and install the wheel file with AVX-512 as minimum required instruction set from the table above.

I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX2 AVX512F FMA To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.

 

Note: If you run a release with AVX-512 as minimum required instruction set on a machine without AVX-512 instruction set support, you will run into "Illegal instruction (core dumped)" error.

Note than for 1.14.0 install we have fixed a few vulnerabilities and the corrected versions can be installed using the below commands. We identified new CVE issues from curl and GCP support in the previous pypi package release, so we had to introduce a new set of fixed packages in PyPI

Available for Linux* here

 

Available for Linux*

TensorFlow version: 2.7

Installation instructions:

Run the below instruction to install the wheel into an existing Python* installation. Python versions supported are 3.7, 3.8, 3.9

pip install tensorflow==2.7.0

Users can enable those oneDNN CPU optimizations by setting the the environment variable TF_ENABLE_ONEDNN_OPTS to 1.

export TF_ENABLE_ONEDNN_OPTS=1 

Please check #Additional Info for differences between Intel® Optimization for TensorFlow* and official TensorFlow*.

2. Docker Images

Get Intel® Optimization for TensorFlow* Docker Images

Starting version 1.14, Google released DL containers for TensorFlow on CPU optimized with oneDNN by default. The TensorFlow v1.x CPU container names are in the format "tf-cpu.", TensorFlow v2.x CPU container names are in the format "tf2-cpu." and support Python3. Below are sample commands to download the docker image locally and launch the container for TensorFlow 1.15 or TensorFlow 2.6. Please use one of the following commands at one time.

# TensorFlow 1.15

docker run -d -p 8080:8080 -v /home:/home gcr.io/deeplearning-platform-release/tf-cpu.1-15

# TensorFlow 2.7

docker run -d -p 8080:8080 -v /home:/home gcr.io/deeplearning-platform-release/tf2-cpu.2-7

This command will start the TensorFlow 1.15 or TensorFlow 2.7 with oneDNN enabled in detached mode, bind the running Jupyter server to port 8080 on the local machine, and mount local /home directory to /home in the container. The running JupyterLab instance can be accessed at localhost:8080.

To launch an interactive bash instance of the docker container, run one of the below commands.

# TensorFlow 1.15

docker run -v /home:/home -it gcr.io/deeplearning-platform-release/tf-cpu.1-15 bash

# TensorFlow 2.7

docker run -v /home:/home -it gcr.io/deeplearning-platform-release/tf2-cpu.2-7 bash

 

Available Container Configurations

You can find all supported docker tags/configurations here.

Tensorflow Version: 2.6.0

These docker images are all published at http://hub.docker.com in intel/intel-optimized-tensorflow and intel/intel-optimized-tensorflow-avx512 namespaces and can be pulled with the following command:

 

# intel-optimized-tensorflow

docker pull intel/intel-optimized-tensorflow

# intel-optimized-tensorflow-avx512

docker pull intel/intel-optimized-tensorflow-avx512:latest

 

For example, to run the data science container directly, simply

# intel-optimized-tensorflow

docker run -it -p 8888:8888 intel/intel-optimized-tensorflow

# intel-optimized-tensorflow-avx512

docker run -it -p 8888:8888 intel/intel-optimized-tensorflow-avx512:latest

And then go to your browser on http://localhost:8888/


For those who want to navigate through the browser, follow the links:

 

Available Container Configurations

You can find all supported docker tags/configurations for intel-optimized-tensorflow and intel-optimized-tensorflow-avx512.

To get the latest Release Notes on Intel® Optimization for TensorFlow*, please refer this article

More containers for Intel® Optimization for TensorFlow* can be found at the Intel® oneContainer Portal.

3. Build from Source

Build TensorFlow from Source with Intel oneAPI oneDNN library

Building TensorFlow from source is not recommended. However, if instructions provided above do not work due to unsupported ISA, you can always build from source.

Building TensorFlow from source code requires Bazel installation, refer to the instructions here, Installing Bazel.

Installation instructions:

  1. Ensure numpy, keras-applications, keras-preprocessing, pip, six, wheel, mock packages are installed in the Python environment where TensorFlow is being built and installed.
  2. Clone the TensorFlow source code and checkout a branch of your preference
    • git clone https://github.com/tensorflow/tensorflow
    • git checkout r2.7
  3. Run "./configure" from the TensorFlow source directory
  4. Execute the following commands to create a pip package that can be used to install the optimized TensorFlow build.
    • PATH can be changed to point to a specific version of GCC compiler:

      export PATH=/PATH//bin:$PATH

    • LD_LIBRARY_PATH can also be to new:

      export LD_LIBRARY_PATH=/PATH//lib64:$LD_LIBRARY_PATH

    • Set the compiler flags support by the GCC on your machine to build TensorFlow with oneDNN.

      bazel build --config=mkl -c opt --copt=-march=native //tensorflow/tools/pip_package:build_pip_package

       
      • If you would like to build the binary against certain hardware, ensure appropriate "march" and "mtune" flags are set. Refer the gcc online docs to know the flags supported by your GCC version.

        bazel build --config=mkl --cxxopt=-D_GLIBCXX_USE_CXX11_ABI=0 --copt=-march=sandybridge --copt=-mtune=ivybridge --copt=-O3 //tensorflow/tools/pip_package:build_pip_package

         

      • Alternatively, if you would like to build the binary against certain instruction sets, set appropriate "Instruction sets" flags:

        bazel build --config=mkl -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mavx512f --copt=-mavx512pf --copt=-mavx512cd --copt=-mavx512er //tensorflow/tools/pip_package:build_pip_package

        Flags set above will add AVX, AVX2 and AVX512 instructions which will result in "illegal instruction" errors when you use older CPUs. If you want to build on older CPUs, set the instruction flags accordingly.

      • Users could enable additional oneDNN features by passing a "--copt=-Dxxx" build option.  For example, enable ITT_TASKS feature from oneDNN by using below build instruction. User could refer to oneDNN build options for more details.

        bazel build --config=mkl -c opt --copt=-march=native --copt=-DDNNL_ENABLE_ITT_TASKS=True //tensorflow/tools/pip_package:build_pip_package
  5. Install the optimized TensorFlow wheel
    • bazel-bin/tensorflow/tools/pip_package/build_pip_package ~/path_to_save_wheel
    • pip install --upgrade --user ~/path_to_save_wheel/

* Prior to TensorFlow 2.3

Prerequisites

Install the below Visual C++ 2015 build tools from https://visualstudio.microsoft.com/vs/older-downloads/

  • Microsoft Visual C++ 2015 Redistributable Update 3
  • Microsoft Build Tools 2015 Update 3

Installation

  1. Refer to Linux Section and follow Steps 1 through 3
  2. To build TensorFlow with oneDNN support, we need two additional steps.
    • Link.exe on  Visual Studio 2015 causes the linker issue when /WHOLEARCHIVE switch is used. To overcome this issue, install the hotfix to your Visual C++ compiler available at https://support.microsoft.com/en-us/help/4020481/fix-link-exe-crashes-with-a-fatal-lnk1000-error-when-you-use-wholearch  
    • Add a PATH environment variable to include MKL runtime lib location that will be created during the build process. The base download location can be specified in the bazel build command by using the --output_base option, and the oneDNN libraries will then be downloaded into a directory relative to that base              
      • set PATH=%PATH%;output_dir\external\mkl_windows\lib

         3. Bazel build with the with "mkl" flag and the "output_dir" to use the right mkl libs

             bazel --output_base=output_dir build --config=mkl --config=opt //tensorflow/tools/pip_package:build_pip_package

          4. Install the optimized TensorFlow wheel

      bazel-bin\tensorflow\tools\pip_package\build_pip_package C:\temp\path_to_save_wheel

      pip install C:\temp\path_to_save_wheel\

 

* TensorFlow 2.3 and newer:

Prerequisites

  • python 3.5 or 3.6 for Windows. Select pip as an optional feature and add it to your %PATH% environmental variable.
  • Tensorflow dependent packages (check out the dependent packages listed in setup.py)
    • pip3 install six numpy wheel

    • pip3 install keras_applications==1.0.6 --no-deps

    • pip3 install keras_preprocessing==1.0.5 --no-deps

       

  • MSYS2
    • (Required for Bazel) MSYS2 is a software distro and building platform for Windows. It contains Bash and common Unix tools (like grep, tar, git).
    • Bazel needs packages installed using the msys2 terminal (note that proxy vars need to be set otherwise these installs won’t work)
    • Open MSYS2 terminal and run the command

      pacman -S zip unzip patch diffutils git

  • Bazel
  • Install Visual C++* Build Tools 2019. It comes with Visual Studio* 2019 but can be installed separately. Go to the Visual Studio Downloads, download and install the following:

    • Microsoft Visual C++ 2019 Redistributable from Other Tools and Frameworks

    • Microsoft Build Tools 2019 from Tools for Visual Studio 2019

Installation

  1. Set the following environment variables:
    1. BAZEL_SH: C:\msys64/usr\bin\bash.exe

    2. BAZEL_VS: C:\Program Files (x86)\Microsoft Visual Studio

    3. BAZEL_VC: C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC

  2. Note: For compile time reduction, please set:

    set TF_VC_VERSION=16.6

    More details can be found here.

  3. Add to the PATH environment variable to include
    1. python path, e.g. C:\Program Files\Python36

    2. oneDNN runtime lib location that will be created during the build process, e.g. D:\output_dir\external\mkl\windows\lib

    3. the Bazel path, e.g. C:\Program Files\Bazel-3.2.0

    4. MSYS2 path, e.g. C:\msys64;C:\msys64/usr\bin

    5. Git path, e.g. C:\Program Files\Git\cmd;C:\Program Files\Git/usr\bin

      set PATH=%PATH%;C:\Program Files\Python36;D:\output_dir\external\mkl_windows\lib;C:\Program Files\Bazel-3.2.0;C:\msys64;C:\msys64/usr\bin;C:\Program Files\Git\cmd;C:\Program Files\Git/usr\bin
  4. Download the TensorFlow source code, checkout the release branch, and configure the build:
    • git clone https://github.com/Intel-tensorflow/tensorflow.git
    • cd tensorflow
    • git checkout r2.3-windows
    • python ./configure.py
  5. Set the oneDNN output directory location outside TensorFlow home directory to avoid infinite symlink expansion error. Then add the path to the oneDNN output directory to the system PATH:   
    • set OneDNN_DIR=\one_dnn_dir

    • set PATH=%OneDNN_DIR%;%PATH%

  6. Build TensorFlow from source with oneDNN. Navigate to the TensorFlow root directory tensorflow and run the following bazel command to build TensorFlow oneDNN from Source:
    • bazel --output_base=%OneDNN_DIR% build --announce_rc --config=opt --config=mkl --action_env=PATH=""  --define=no_tensorflow_py_deps=true  tensorflow/tools/pip_package:build_pip_package

Note: Based on bazel issue #7026 we set --action_env=PATH=. Open cmd.exe, run echo %PATH% and copy the output to the value of --action_env=PATH=. If found, please use single quotes with folder names of white spaces.

 

Additional Informatoin

Once Intel-optimized TensorFlow is installed, running the below command must print "True" if oneDNN optimizations are present.

import tensorflow as tf

import os

def get_mkl_enabled_flag():

    mkl_enabled = False
    major_version = int(tf.__version__.split(".")[0])
    minor_version = int(tf.__version__.split(".")[1])
    if major_version >= 2:
        if minor_version < 5:
            from tensorflow.python import _pywrap_util_port
        else:
            from tensorflow.python.util import _pywrap_util_port
            onednn_enabled = int(os.environ.get('TF_ENABLE_ONEDNN_OPTS', '0'))
        mkl_enabled = _pywrap_util_port.IsMklEnabled() or (onednn_enabled == 1)
    else:
        mkl_enabled = tf.pywrap_tensorflow.IsMklEnabled()
    return mkl_enabled

print ("We are using Tensorflow version", tf.__version__)
print("MKL enabled :", get_mkl_enabled_flag())

  1. Intel-optimized TensorFlow enables oneDNN calls by default. For v2.4 and previous version, If at any point you wish to disable Intel MKL primitive calls, this can be disabled by setting TF_DISABLE_MKL flag to 1 before running your TensorFlow script.

    export TF_DISABLE_MKL=1       

           However, note that this flag will only disable oneDNN calls, but not MKL-ML calls. 

            Although oneDNN is responsible for most optimizations, certain ops are optimized by MKL-ML library, including matmul, transpose, etc. Disabling MKL-ML calls are not supported by TF_DISABLE_MKL flag at present and Intel is working with Google to add this functionality

       2. CPU affinity settings in Anaconda's TensorFlow: If oneDNN enabled TensorFlow is installed from the anaconda channel (not Intel channel), the "import tensorflow" command sets the KMP_BLOCKTIME and OMP_PROC_BIND environment variables if not already set. However, these variables may have effects on other libraries such as Numpy/Scipy which use OpenMP or oneDNN. Alternatively, you can either set preferred values or unset them after importing TensorFlow. More details available in the TensorFlow GitHub issue

            import tensorflow # this sets KMP_BLOCKTIME and OMP_PROC_BIND

            import os

            # delete the existing values

            del os.environ['OMP_PROC_BIND']

           del os.environ['KMP_BLOCKTIME']

 

Although official TensorFlow has oneDNN optimizations by default, there are still some major differences betwteen Intel Optimization for Tensorflow and official TensorFlow

 

Here is a comparison table For TensorFlow v2.6 and v2.7.

  Intel Optimization for Tensorflow official TensorFlow (Running on Intel CPUs)
oneDNN optimiziations   Enabled by default Enable by setting environment variable TF_ENABLE_ONEDNN_OPTS=1 at runtime
OpenMP Optimizations Enabled by default N/A. use eigen thread pool instead
Native Layout Format  Enabled by default. Disable the feature by setting the env-variable TF_ENABLE_MKL_NATIVE_FORMAT=0 Enable by setting environment variable TF_ENABLE_ONEDNN_OPTS=1 at runtime
int8 support from oneDNN Enabled by default Enable by setting environment variable TF_ENABLE_ONEDNN_OPTS=1 at runtime

 

Here is a comparison table for TensorFlow v2.5.

  Intel Optimization for Tensorflow official TensorFlow (Running on Intel CPUs)
oneDNN optimiziations   Enabled by default Enable by setting environment variable TF_ENABLE_ONEDNN_OPTS=1 at runtime
OpenMP Optimizations Enabled by default N/A. use eigen thread pool instead
Native Layout Format  Enabled by default. Disable the feature by setting the env-variable TF_ENABLE_MKL_NATIVE_FORMAT=0 Enable by setting environment variable TF_ENABLE_ONEDNN_OPTS=1 at runtime
int8 support from oneDNN Enabled by setting the env-variable TF_ENABLE_MKL_NATIVE_FORMAT=0 Not supported

Support

If you have further questions or need support on your workload optimization, Please submit your queries at the TensorFlow GitHub issues with the label "comp:mkl" or the Intel AI Frameworks forum.

 

Useful Resources

Category Links
Installation & releases Intel® AI Analytics Toolkit
TensorFlow in Anaconda
Intel TensorFlow Installation Guide
Build TensorFlow from source on Windows
Intel-optimized TensorFlow on AWS
Intel oneContainer
Performance General BKMs to maximize performance
Topology specific BKMs and tutorials
Intel Model Zoo with pretrained models
Getting Started with AutoMixedPrecisionMkl
Optimize pre-trained model

 

 

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.