MIOpen

AMD's library for high performance machine learning primitives. Sources and binaries can be found at MIOpen's GitHub site. The latest released documentation can be read online here.

MIOpen supports two programming models -

1. OpenCL
2. HIP

Prerequisites

• A ROCm enabled platform, more info here
• Base software stack, which includes
  • OpenCL - OpenCL libraries and header files
  • HIP -
    • HIP and HCC libraries and header files
    • clang-ocl -- required
• MIOpenGEMM to enable various functionalities including transposed and dilated convolutions. This is optional on the HIP backend. Users can enable this library using the cmake configuration flag -DMIOPEN_USE_MIOPENGEMM=On.
• ROCm cmake modules can be installed from here
• Half - IEEE 754-based half-precision floating point library
• Boost at least version 1.58
  • MIOpen uses boost-system and boost-filesystem packages to enable persistent kernel cache
• rocBlas Minimum version branch master-rocm-2.10

Installing MIOpen with pre-built packages

MIOpen can be installed on Ubuntu using apt-get.

For OpenCL backend: apt-get install miopen-opencl

For HIP backend: apt-get install miopen-hip

Currently both the backends cannot be installed on the same system simultaneously. If a different backend other than what currently exists on the system is desired, please uninstall the existing backend completely and then install the new backend.

Installing MIOpen kernels package

MIOpen provides an optional pre-compiled kernels package to reduce the startup latency. To install the kernels package for your GPU architecture, use the following command:

apt-get install miopen-kernels-<arch>-<num cu>

Where <arch> is the GPU architecture ( for example, gfx900, gfx906 ) and <num cu> is the number of CUs available in the GPU (for example 56 or 64 etc).

Not installing these packages would not impact the functioning of MIOpen, since MIOpen will compile these kernels on the target machine once the kernel is run. However, the compilation step may significantly increase the startup time for different operations.

Installing the dependencies

The dependencies can be installed with the install_deps.cmake, script: cmake -P install_deps.cmake

This will install by default to /usr/local but it can be installed in another location with --prefix argument:

cmake -P install_deps.cmake --prefix /some/local/dir

This prefix can used to specify the dependency path during the configuration phase using the CMAKE_PREFIX_PATH.

MIOpen's HIP backend uses rocBlas by default. Users can install rocBlas minimum release by using apt-get install rocblas. To disable using rocBlas set the configuration flag -DMIOPEN_USE_ROCBLAS=Off. rocBlas is not available for the OpenCL backend.

Installing minimum dependencies in ROCm environment

Users who are working in a fully installed and up to date ROCm environment may not wish to additionally install rocm-cmake, clang-ocl, MIOpenGEMM, or rocBLAS. This can be done by simply inserting the command --minimum into the cmake command as shown below:

cmake -P install_deps.cmake --minimum --prefix /some/local/dir

This will build the Boost and half libraries.

Building MIOpen from source

Configuring with cmake

First create a build directory:

mkdir build; cd build;

Next configure cmake. The preferred backend for MIOpen can be set using the -DMIOPEN_BACKEND cmake variable.

For OpenCL, run:

cmake -DMIOPEN_BACKEND=OpenCL ..

The above assumes that OpenCL is installed in one of the standard locations. If not, then manually set these cmake variables:

cmake -DMIOPEN_BACKEND=OpenCL -DMIOPEN_HIP_COMPILER=<hip-compiler-path> -DOPENCL_LIBRARIES=<opencl-library-path> -DOPENCL_INCLUDE_DIRS=<opencl-headers-path> ..

And an example setting the dependency path for an envirnment in ROCm 3.5 and later:

cmake -DMIOPEN_BACKEND=OpenCL -DMIOPEN_HIP_COMPILER=/opt/rocm/llvm/bin/clang++ -DCMAKE_PREFIX_PATH=/some/local/dir ..

For ROCm 3.3 and earlier:

cmake -DMIOPEN_BACKEND=OpenCL -DCMAKE_PREFIX_PATH=/some/local/dir ..

For the HIP backend on ROCm 3.3 and earlier, run:

Set the C++ compiler to hcc.

export CXX=<location-of-hcc-compiler>
cmake -DMIOPEN_BACKEND=HIP -DCMAKE_PREFIX_PATH="<hip-installed-path>;<hcc-installed-path>;<clang-ocl-installed-path>;<miopen-dependency-path>" ..

An example cmake step can be:

CXX=/opt/rocm/hcc/bin/hcc cmake -DMIOPEN_BACKEND=HIP -DCMAKE_PREFIX_PATH="/opt/rocm/hcc;/opt/rocm/hip;/some/local/dir" ..

For the HIP backend on ROCm 3.5 and later, run:

Set the C++ compiler to clang++.

export CXX=<location-of-clang++-compiler>
cmake -DMIOPEN_BACKEND=HIP -DCMAKE_PREFIX_PATH="<hip-installed-path>;<rocm-installed-path>;<miopen-dependency-path>" ..

An example cmake step can be:

CXX=/opt/rocm/llvm/bin/clang++ cmake -DMIOPEN_BACKEND=HIP -DCMAKE_PREFIX_PATH="/some/local/dir" ..

Note: When specifying the path for the CMAKE_PREFIX_PATH variable, do not use the ~ shorthand for the user home directory.

Setting Up Locations

By default the install location is set to '/opt/rocm', this can be set by using CMAKE_INSTALL_PREFIX:

cmake -DMIOPEN_BACKEND=OpenCL -DCMAKE_INSTALL_PREFIX=<miopen-installed-path> ..

System Performance Database and User Database

The default path to the System PerfDb is miopen/share/miopen/db/ within install location. The default path to the User PerfDb is ~/.config/miopen/. For development purposes, setting BUILD_DEV will change default path to both database files to the source directory:

cmake -DMIOPEN_BACKEND=OpenCL -DBUILD_DEV=On ..

Database paths can be explicitly customized by means of MIOPEN_SYSTEM_DB_PATH (System PerfDb) and MIOPEN_USER_DB_PATH (User PerfDb) cmake variables.

More information about the performance database can be found here.

Persistent Program Cache

MIOpen by default caches the device programs in the location ~/.cache/miopen/. In the cache directory there exists a directory for each version of MIOpen. Users can change the location of the cache directory during configuration using the flag -DMIOPEN_CACHE_DIR=<cache-directory-path>.

Users can also disable the cache during runtime using the environmental variable set as MIOPEN_DISABLE_CACHE=1.

For MIOpen version 2.3 and earlier

If the compiler changes, or the user modifies the kernels then the cache must be deleted for the MIOpen version in use; e.g., rm -rf ~/.cache/miopen/<miopen-version-number>. More information about the cache can be found here.

For MIOpen version 2.4 and later

MIOpen's kernel cache directory is versioned so that users' cached kernels will not collide when upgrading from earlier version.

Changing the cmake configuration

The configuration can be changed after running cmake by using ccmake:

ccmake .. OR cmake-gui: cmake-gui ..

The ccmake program can be downloaded as the Linux package cmake-curses-gui, but is not available on windows.

Building the library

The library can be built, from the build directory using the 'Release' configuration:

cmake --build . --config Release OR make

And can be installed by using the 'install' target:

cmake --build . --config Release --target install OR make install

This will install the library to the CMAKE_INSTALL_PREFIX path that was set.

Building the driver

MIOpen provides an application-driver which can be used to execute any one particular layer in isolation and measure performance and verification of the library.

The driver can be built using the MIOpenDriver target:

cmake --build . --config Release --target MIOpenDriver OR make MIOpenDriver

Documentation on how to run the driver is here.

Running the tests

The tests can be run by using the 'check' target:

cmake --build . --config Release --target check OR make check

A single test can be built and ran, by doing:

cmake --build . --config Release --target test_tensor
./bin/test_tensor

Building the documentation

HTML and PDF documentation can be built using:

cmake --build . --config Release --target doc OR make doc

This will build a local searchable web site inside the ./MIOpen/doc/html folder and a PDF document inside the ./MIOpen/doc/pdf folder.

Documentation is built using generated using Doxygen and should be installed separately.

HTML and PDFs are generated using Sphinx and Breathe, with the ReadTheDocs theme.

Requirements for both Sphinx, Breathe, and the ReadTheDocs theme can be filled for these in the MIOpen/doc folder:

pip install -r ./requirements.txt

Depending on your setup sudo may be required for the pip install.

Formatting the code

All the code is formatted using clang-format. To format a file, use:

clang-format-3.8 -style=file -i <path-to-source-file>

Also, githooks can be installed to format the code per-commit:

./.githooks/install

Installing the dependencies manually

If Ubuntu v16 is used then the Boost packages can also be installed by:

sudo apt-get install libboost-dev
sudo apt-get install libboost-system-dev
sudo apt-get install libboost-filesystem-dev

Note: MIOpen by default will attempt to build with Boost statically linked libraries. If it is needed, the user can build with dynamically linked Boost libraries by using this flag during the configruation stage:

-DBoost_USE_STATIC_LIBS=Off

however, this is not recommended.

The half header needs to be installed from here.

Using docker

The easiest way is to use docker. You can build the top-level docker file:

docker build -t miopen-image .

Then to enter the development environment use docker run, for example:

docker run -it -v $HOME:/data --privileged --rm --device=/dev/kfd --device /dev/dri:/dev/dri:rw  --volume /dev/dri:/dev/dri:rw -v /var/lib/docker/:/var/lib/docker --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined miopen-image

Prebuilt docker images can be found on ROCm's public docker hub here.

Citing MIOpen

MIOpen's paper can be accessed on arXiv:
MIOpen: An Open Source Library For Deep Learning Primitives

Citation BibTeX

@misc{jeh2019miopen,
    title={MIOpen: An Open Source Library For Deep Learning Primitives},
    author={Jehandad Khan and Paul Fultz and Artem Tamazov and Daniel Lowell and Chao Liu and Michael Melesse and Murali Nandhimandalam and Kamil Nasyrov and Ilya Perminov and Tejash Shah and Vasilii Filippov and Jing Zhang and Jing Zhou and Bragadeesh Natarajan and Mayank Daga},
    year={2019},
    eprint={1910.00078},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}