Permalink
Find file Copy path
426 lines (340 sloc) 18.3 KB

Intel(R) Math Kernel Library for Deep Neural Networks (Intel(R) MKL-DNN)

v0.17 beta

Intel(R) Math Kernel Library for Deep Neural Networks (Intel(R) MKL-DNN) is an open-source performance library for deep-learning applications. The library accelerates deep-learning applications and frameworks on Intel architecture. Intel MKL-DNN contains vectorized and threaded building blocks that you can use to implement deep neural networks (DNN) with C and C++ interfaces.

DNN functionality optimized for Intel architecture is also included in Intel Math Kernel Library (Intel MKL). The API in that implementation is not compatible with Intel MKL-DNN and does not include certain new and experimental features.

This release contains performance-critical functions that improve performance of the following deep learning topologies and variations of these:

Application Example topology
Image recognition AlexNet, VGG, GoogleNet, ResNet, MobileNet
Image segmentation FCN, SegNet, MaskRCNN, U-Net
Volumetric segmentation 3D-Unet
Object detection SSD, Faster R-CNN, Yolo
Neural machine translation (experimental) GNMT
Speech recognition (experimental) DeepSpeech
Adversarial networks DCGAN, 3DGAN
Reinforcement learning A3C
Text-to-speech WaveNet

Intel MKL-DNN is used in the following software products:

License

Intel MKL-DNN is licensed under Apache License Version 2.0. This software includes the following third-party components:

Documentation

  • Introduction explains the programming model and basic concepts
  • Reference manual provides detailed functionality description
  • Examples demonstrates use of C and C++ APIs in simple topologies
  • Tutorial provides step-by-step installation instructions and an example walkthrough

Support

Please submit your questions, feature requests, and bug reports on the GitHub issues page.

WARNING The following functionality has preview status and might change without prior notification in future releases:

  • Convolutions with s16 data type in source, weights or destination
  • Convolutions and auxiliary primitives for 3D spatial data
  • RNN, LSTM, and GRU primitives
  • Threading Building Blocks (TBB) support

How to Contribute

We welcome community contributions to Intel MKL-DNN. If you have an idea on how to improve the library:

  • Share your proposal via GitHub issues.
  • Ensure you can build the product and run all the examples with your patch.
  • In the case of a larger feature, create a test.
  • Submit a pull request.

We will review your contribution and, if any additional fixes or modifications are necessary, may provide feedback to guide you. When accepted, your pull request will be merged to the repository.

System Requirements

Intel MKL-DNN supports Intel 64 architecture and compatible architectures. The library is optimized for the systems based on

  • Intel Atom(R) processor with Intel SSE4.1 support
  • 4th, 5th, 6th, 7th, and 8th generation Intel(R) Core(TM) processor
  • Intel(R) Xeon(R) processor E5 v3 family (formerly Haswell)
  • Intel Xeon processor E5 v4 family (formerly Broadwell)
  • Intel Xeon Platinum processor family (formerly Skylake)
  • Intel(R) Xeon Phi(TM) processor x200 product family (formerly Knights Landing)
  • Intel Xeon Phi processor x205 product family (formerly Knights Mill)

and compatible processors.

The software dependencies are:

  • Cmake 2.8.0 or later
  • Doxygen 1.8.5 or later
  • C++ compiler with C++11 standard support
  • Optional dependencies:
    • GNU* OpenMP*, LLVM OpenMP, or Intel OpenMP
    • Threading Building Blocks (TBB) 2017 or later
    • Intel MKL 2017 Update 1 or Intel MKL small libraries

Note Building Intel MKL-DNN with optional dependencies may introduce additional runtime dependencies for the library. For details, refer to the corresponding software system requirements.

The software was validated on RedHat* Enterprise Linux 7 with

  • GNU Compiler Collection 4.8, 5.4, 6.1, 7.2, and 8.1
  • Clang* 3.8.0
  • Intel C/C++ Compiler 17.0, 18.0, and 19.0

on Windows Server* 2012 R2 with

on macOS* 10.13 (High Sierra) with

The implementation uses OpenMP 4.0 SIMD extensions. We recommend using the Intel C++ Compiler for the best performance results.

Installation

Build from source

Download source code

Download Intel MKL-DNN source code or clone the repository to your system.

git clone https://github.com/intel/mkl-dnn.git

Configure build

Intel MKL-DNN uses a CMake-based build system. You can use CMake options to control the build. Along with the standard CMake options such as CMAKE_INSTALL_PREFIX and CMAKE_BUILD_TYPE, you can pass Intel MKL-DNN specific options:

Option Possible Values (defaults in bold) Description
MKLDNN_LIBRARY_TYPE SHARED, STATIC Defines the resulting library type
MKLDNN_THREADING OMP, OMP:INTEL, OMP:COMP, TBB Defines the threading type
WITH_EXAMPLE ON, OFF Controls building the examples
WITH_TEST ON, OFF Controls building the tests
ARCH_OPT_FLAGS compiler flags Specifies compiler optimization flags (see warning note below)
VTUNEROOT path Enables integration with Intel(R) VTune(TM) Amplifier

WARNING

By default, Intel MKL-DNN is built specifically for the processor type of the compiling machine (for example, -march=native in the case of GCC). While this option gives better performance, the resulting library can be run only on systems that are instruction-set compatible with the compiling machine.

Therefore, if Intel MKL-DNN is to be shipped to other platforms (for example, built by Linux distribution maintainers), consider setting ARCH_OPT_FLAGS to "".

For more options and details, check cmake/options.cmake.

Using Intel MKL (optional)

Intel MKL-DNN includes an optimized matrix-matrix multiplication (GEMM) implementation for modern platforms. The library can also take advantage of GEMM functions from Intel MKL to improve performance with older versions of compilers or on older platforms. This behavior is controlled by the MKLDNN_USE_MKL option.

Option Possible Values (defaults in bold) Description
MKLDNN_USE_MKL DEF, NONE, ML, FULL, FULL:STATIC Defines the binary dependency on Intel MKL

The dynamic library with this functionality is included in the repository. If you choose to build Intel MKL-DNN with the binary dependency, download the Intel MKL small libraries using the provided script:

Linux/macOS

cd scripts && ./prepare_mkl.sh && cd ..

Windows*

cd scripts && call prepare_mkl.bat && cd ..

or manually from GitHub release section, and unpack it to the external directory in the repository root. Intel MKL-DNN can also be built with full Intel MKL if the latter is installed on the system. You might need to set the MKLROOT environment variable to the path where the full Intel MKL is installed to help cmake locate the library.

Note

Using Intel MKL small libraries currently works only for Intel MKL-DNN built with OpenMP. Building with Intel TBB requires either the full Intel MKL library or a standalone build.

Using Intel MKL or Intel MKL small libraries will introduce additional runtime dependencies. For additional information, refer to Intel MKL system requirements.

Threading

Intel MKL-DNN is parallelized and can use the OpenMP or TBB threading runtime. OpenMP threading is the default build mode and is recommended for the best performance. TBB support is experimental. This behavior is controlled by the MKLDNN_THREADING option.

Option Possible Values (defaults in bold) Description
MKLDNN_THREADING OMP, OMP:INTEL, OMP:COMP, TBB Defines the threading type
OpenMP

Intel MKL-DNN can use Intel, GNU or CLANG OpenMP runtime. Because different OpenMP runtimes may not be binary compatible, it's important to ensure that only one OpenMP runtime is used throughout the application. Having more than one OpenMP runtime initialized may lead to undefined behavior including incorrect results or crashes.

Intel MKL-DNN library built with the binary dependency will link against the Intel OpenMP runtime included with the Intel MKL small libraries package. The Intel OpenMP runtime is binary compatible with the GNU OpenMP and Clang OpenMP runtimes and is recommended for the best performance results.

Intel MKL-DNN library built standalone will use the OpenMP runtime supplied by the compiler, so as long as both the library and the application use the same compiler, the correct OpenMP runtime will be used.

TBB

TBB support is experimental. Intel MKL-DNN has limited optimizations done for Intel TBB and has some functional limitations if built with Intel TBB.

Functional limitations:

  • Convolution with Winograd algorithm is not supported

Performance limitations (mostly less parallelism than in case of OpenMP):

  • Batch normalization
  • Convolution backward by weights
  • mkldnn_sgemm

WARNING

If the library is built with the full Intel MKL, the user is expected to set the MKL_THREADING_LAYER environment variable to either tbb or sequential in order to force Intel MKL to use Intel TBB for parallelization or to be sequential, respectively. Without this setting, Intel MKL (RT library) tries to use OpenMP for parallelization by default.

Build on Linux/macOS

Ensure that all software dependencies are in place and have at least the minimal supported version.

Configure CMake and create a makefile:

mkdir -p build && cd build && cmake $CMAKE_OPTIONS ..

Build the application:

make

The build can be validated with the unit-test suite:

ctest

The reference manual is provided inline and can also be generated in HTML format with Doxygen:

make doc

Documentation will reside in the build/reference/html folder.

Finally:

make install

will place the header files, libraries, and documentation in /usr/local. To change the installation path, use the option -DCMAKE_INSTALL_PREFIX=<prefix> when invoking CMake.

Build on Windows

Ensure that all software dependencies are in place and have at least the minimal supported version.

NOTE

Building Intel MKL-DNN from a terminal requires using either the Intel Parallel Studio command prompt or the Microsoft* Visual Studio* developer command prompt instead of the default Windows command prompt.

The Intel(R) Parallel Studio command prompt is an item in the Start menu in the Intel Parallel Studio <version> folder that has a Windows Command Prompt icon and a name like Compiler 18.0 Update 5….

The default for building the project for the Intel C++ Compiler is to use the Intel Parallel Studio developer command prompt.

Configure CMake and create a Microsoft Visual Studio solution:

mkdir build & cd build && cmake -G "Visual Studio 15 2017 Win64" ..

For the solution to use Intel C++ Compiler:

cmake -G "Visual Studio 15 2017 Win64" -T "Intel C++ Compiler 18.0" ..

After you have built the initial project using CMake, you can then open the project with Microsoft Visual Studio and build from there. You can also use msbuild command-line tool to build from the command line:

msbuild "Intel(R) MKL-DNN.sln" /p:Configuration=Release [/t:rebuild] /m

where the optional argument /t:rebuild rebuilds the project.

The build can be validated with the unit-test suite:

ctest

Linking Your Application

Linux/macOS

Intel MKL-DNN includes several header files providing C and C++ APIs for the functionality and one or several dynamic libraries depending on how Intel MKL-DNN was built.

Linux

File Description
include/mkldnn.h C header
include/mkldnn.hpp C++ header
include/mkldnn_types.h Auxiliary C header
lib/libmkldnn.so Intel MKL-DNN dynamic library
lib/libmkldnn.a Intel MKL-DNN static library (if built with MKLDNN_LIBRARY_TYPE=STATIC)
lib/libiomp5.so Intel OpenMP* runtime library (if built with MKLDNN_USE_MKL=ML)
lib/libmklml_gnu.so Intel MKL small library for GNU OpenMP runtime (if built with MKLDNN_USE_MKL=ML)
lib/libmklml_intel.so Intel MKL small library for Intel OpenMP runtime (if built with MKLDNN_USE_MKL=ML)

macOS

File Description
include/mkldnn.h C header
include/mkldnn.hpp C++ header
include/mkldnn_types.h Auxiliary C header
lib/libmkldnn.dylib Intel MKL-DNN dynamic library
lib/libmkldnn.a Intel MKL-DNN static library (if built with MKLDNN_LIBRARY_TYPE=STATIC)
lib/libiomp5.dylib Intel OpenMP* runtime library (if built with MKLDNN_USE_MKL=ML)
lib/libmklml_gnu.dylib Intel MKL small library for GNU OpenMP runtime (if built with MKLDNN_USE_MKL=ML)
lib/libmklml_intel.dylib Intel MKL small library for Intel OpenMP runtime (if built with MKLDNN_USE_MKL=ML)

Linkline examples below assume that Intel MKL-DNN is installed in the directory defined in the MKLDNNROOT environment variable.

g++ -std=c++11 -I${MKLDNNROOT}/include -L${MKLDNNROOT}/lib simple_net.cpp -lmkldnn
clang -std=c++11 -I${MKLDNNROOT}/include -L${MKLDNNROOT}/lib simple_net.cpp -lmkldnn
icpc -std=c++11 -I${MKLDNNROOT}/include -L${MKLDNNROOT}/lib simple_net.cpp -lmkldnn

WARNING

Using the GNU compiler with the -fopenmp and -liomp5 options will link the application with both the Intel and GNU OpenMP runtime libraries. This will lead to undefined behavior in the application.

NOTE

Applications linked dynamically will resolve the dependencies at runtime. Make sure that the dependencies are available in the standard locations defined by the operating system, in the locatons listed in LD_LIBRARY_PATH (Linux), DYLD_LIBRARY_PATH (macOS) environment variables, or rpath mechanism.

Windows

Intel MKL-DNN includes several header files providing C and C++ APIs for the functionality and one or several dynamic libraries depending on how Intel MKL-DNN was built.

File Description
bin\libmkldnn.dll Intel MKL-DNN dynamic library
bin\libiomp5.dll Intel OpenMP* runtime library (if built with MKLDNN_USE_MKL=ML)
bin\libmklml.dll Intel MKL small library (if built with MKLDNN_USE_MKL=ML)
include\mkldnn.h C header
include\mkldnn.hpp C++ header
include\mkldnn_types.h Auxiliary C header
lib\libmkldnn.lib Intel MKL-DNN import library
lib\libiomp5.lib Intel OpenMP* runtime import library (if built with MKLDNN_USE_MKL=ML)
lib\libmklml.lib Intel MKL small library import library (if built with MKLDNN_USE_MKL=ML)

To link the application from the command line, set up the LIB and INCLUDE environment variables to point to the locations of the Intel MKL-DNN headers and libraries. The Linkline examples below assume that Intel MKL-DNN is installed in the directory defined in the MKLDNNROOT environment variable.

set INCLUDE=%MKLDNNROOT%\include;%INCLUDE%
set LIB=%MKLDNNROOT%\lib;%LIB%
icl /Qstd=c++11 /qopenmp simple_net.cpp mkldnn.lib
cl simple_net.cpp mkldnn.lib

Refer to Microsoft Visual Studio documentation on linking the application using MSVS solutions.

NOTE Applications linked dynamically will resolve the dependencies at runtime. Make sure that the dependencies are available in the standard locations defined by the operating system or in the locatons listed in the PATH environment variable.


Legal Information