- Supported Platforms
- Getting Started with SYCL-DNN
- Support
- Cross-compilation with ComputeCpp
- Contributions
- Citation
SYCL-DNN is a library implementing various neural network algorithms such as pooling and convolution written using SYCL and C++.
SYCL-DNN currently supports the following operations:
- 2D convolutions
- 2D depthwise convolutions
- 2D max & average pooling
- Relu and tanh activations
The convolution operations have several implementations, including tiled and Winograd kernels. The supported data format is NHWC.
The project is maintained by Codeplay Software.
The master branch of SYCL-DNN is regularly tested with the "Supported" hardware listed on the ComputeCpp Supported Platforms page. SYCL-DNN may also work on other hardware and platforms assuming they implement SPIR or SPIR-V support. SYCL-DNN is primarily tested on Ubuntu 16.04 LTS with the corresponding default package versions. SYCL-DNN will generally match the most recently released ComputeCpp, though it is likely to be compatible with other versions. We test against the most recent version.
- CMake (version 3.5.1 and above)
- OpenCL 1.2-capable hardware and drivers with SPIR 1.2 or SPIR-V support
- OpenCL ICD Loader
- OpenCL headers
- gcc (version 5.4 and above)
- ComputeCpp
- Building documentation requires Doxygen and Graphviz/Dot. Tested against versions 1.8.11 and 2.38.0 respectively.
SYCL-DNN uses CMake as its build system. There are provisions in the CMake files for downloading SYCL-DNN's dependencies automatically, for finding other dependencies and for selecting which bits of SYCL-DNN to build. All these configuration options can be found in the main CMakeLists.txt for the project and will show up in the CMake GUI if you use it. By default, the tests and library will be built, but not the benchmarks.
It is recommended to leave the option SNN_DOWNLOAD_MISSING_DEPS set to
on. This will automatically download the source libraries necessary for
SYCL-DNN to build and run (such as Google Test, Google benchmark and
the Eigen linear algebra library). Even if you already have these on your
machine, downloading them as part of the SYCL-DNN means a more consistent
configuration.
You will need to provide the location of the ComputeCpp install you are
using in the variable ComputeCpp_DIR. It should point to the folder
where bin/, lib/ etc. are. This should be the only argument that is
mandatory, everything else should be optional. The default build type is
Release, though this can be overridden.
The following command shows how to compile SYCL-DNN.
# Setup build environment
mkdir build && cd build
cmake .. -DComputeCpp_DIR=/path/to/computecpp
# Compile SYCL-DNN
make -j$(nproc)The "samples" directory contains sample code for the 2D convolution and pooling operations offered by SYCL-DNN. These binaries are compiled when building SYCL-DNN using CMake.
The SYCL-DNN tests are compiled when building SYCL-DNN using CMake. The following command shows how to run the tests.
# Run the tests
ctest
# If compiled with benchmark support, run just the benchmarks
ctest -C Benchmark -E testBug reports are vital to provide feedback to the developers about what is going wrong with the project, you can raise these using the "Issues" feature in GitHub.
Please make sure that your bug report contains the following information:
- A clear and descriptive title.
- The output of
clinfo | grep -E "Platform ID|Name|Vendor|[Vv]ersion|Profile|Extensions". - The output of
computecpp_info. - The exact steps and commands to run to reproduce the bug.
- The exact error text shown (if applicable), otherwise the behaviour you expected and what you encountered instead.
- Should the problem arise outside the project's test suite then please provide a minimal test to allow us to reproduce the problem.
SYCL-DNN supports cross-compilation targeting a number of devices. However, because of the two-step compilation process used in ComputeCpp, standard CMake toolchain files won't provide enough information to SYCL-DNN's build scripts to work properly.
To that end, two toolchains are available. The first, gcc-generic.cmake, will likely work with any prebuilt GCC toolchain (it is not compatible with those installed through package managers). The second is designed to work with the poky toolchain available as part of the Yocto Linux system.
The first step is to download ComputeCpp for both the host machine you are
running on and for the platform you would like to target. You should make
sure to match the ComputeCpp version for both downloads. Both are required
so that the host can run the compiler binary, while the tools can link
using the target device library. Similarly, acquire a GCC toolchain for
the platform you are targeting. Lastly you should download the OpenCL
headers. They are standard across all platforms, but you cannot specify
the default package-managed location of /usr/include for them, as that
will cause conflicts with other system headers. An easy fix is to download
the headers from GitHub.
Toolchain files cannot make use cache variables set by the user when running CMake, as the cache does not exist when the toolchain is executed. Environment variables are available to toolchain files, however, so they are used to pass information to the toolchain. The gcc-generic.cmake toolchain relies on the following environment variables:
SNN_TARGET_TRIPLE # the triple of the platform you are targeting
SNN_TOOLCHAIN_DIR # The root directory of the GCC you downloaded
SNN_SYSROOT_DIR # The system root, probably (but not necessarily)
# ${SNN_TOOLCHAIN_DIR}/${SNN_TARGET_TRIPLE}/libcCMake can then be invoked in a build directory as follows:
cmake -DComputeCpp_DIR=/path/to/computecpp \
-DComputeCpp_HOST_DIR=/path/to/host/computecpp \
-DOpenCL_INCLUDE_DIR=/path/to/opencl/headers \
`# For cross-compiling, check documentation for your platform` \
-DCOMPUTECPP_BITCODE=[(spir[32|64]|spirv[32|64]|ptx64)] \
-DSNN_BUILD_DOCUMENTATION=OFF \
`# Next options let you install the tests to a zippable folder` \
-DSNN_BUILD_TESTS=ON \
-DSNN_BUILD_BENCHMARKS=ON \
-DSNN_INSTALL_TESTS=ON \
-DSNN_INSTALL_BENCHMARKS=ON \
`# This is the most important part - tells CMake to crosscompile` \
-DCMAKE_TOOLCHAIN_FILE=$PWD/../cmake/toolchains/(gcc-generic|arm-gcc-poky).cmake \
-DCMAKE_INSTALL_PREFIX=packaged-binaries \
-GNinja ../The process for the poky toolchain is similar, save that you only need to
provide the SNN_SYSROOT_DIR environment variable. It should be set to
point to the directory named sysroots in the poky toolchain. You will
likely want COMPUTECPP_BITCODE=spir32. Otherwise, these instructions
should still work.
Please see the file CONTRIBUTING.md for further details if you would like to contribute code, build systems, bug fixes or similar.
If you use SYCL-DNN in your research, please cite the library as follows:
Rod Burns, John Lawson, Duncan McBain, and Daniel Soutar. 2019. Accelerated Neural Networks on OpenCL Devices Using SYCL-DNN. In Proceedings of the International Workshop on OpenCL (IWOCL'19). ACM, New York, NY, USA, Article 10, 4 pages. DOI: https://doi.org/10.1145/3318170.3318183
@inproceedings{Burns:2019:ANN:3318170.3318183,
author = {Burns, Rod and Lawson, John and McBain, Duncan and Soutar, Daniel},
title = {Accelerated Neural Networks on OpenCL Devices Using SYCL-DNN},
booktitle = {Proceedings of the International Workshop on OpenCL},
series = {IWOCL'19},
year = {2019},
isbn = {978-1-4503-6230-6},
location = {Boston, MA, USA},
pages = {10:1--10:4},
articleno = {10},
numpages = {4},
url = {http://doi.acm.org/10.1145/3318170.3318183},
doi = {10.1145/3318170.3318183},
acmid = {3318183},
publisher = {ACM},
address = {New York, NY, USA},
keywords = {GPGPU, OpenCL, SYCL, machine learning, neural networks},
}