CLBlast is a modern, lightweight, performant and tunable OpenCL BLAS library written in C++11. It is designed to leverage the full performance potential of a wide variety of OpenCL devices from different vendors, including desktop and laptop GPUs, embedded GPUs, and other accelerators. CLBlast implements BLAS routines: basic linear algebra subprograms operating on vectors and matrices. See [the CLBlast website](https://cnugteren.github.io/clblast) for performance reports on various devices as well as the latest CLBlast news.
The library is not tuned for all possible OpenCL devices: __if out-of-the-box performance is poor, please run the tuners first__. See below for a list of already tuned devices and instructions on how to tune yourself and contribute to future releases of the CLBlast library. See also the [CLBlast feature roadmap](ROADMAP.md) to get an indication of the future of CLBlast.
Building a static version of the library instead of shared one (.dylib/.so/.dll) can be done by disabling the `BUILD_SHARED_LIBS` option when calling CMake. For example:
Like clBLAS and cuBLAS, CLBlast also requires OpenCL device buffers as arguments to its routines. This means you'll have full control over the OpenCL buffers and the host-device memory transfers. CLBlast's API is designed to resemble clBLAS's C API as much as possible, requiring little integration effort in case clBLAS was previously used. Using CLBlast starts by including the C++ header:
Afterwards, any of CLBlast's routines can be called directly: there is no need to initialize the library. The available routines and the required arguments are described in the above mentioned include files and the included [API documentation](doc/clblast.md). The API is kept as close as possible to the Netlib BLAS and the cuBLAS/clBLAS APIs.
To get started quickly, a couple of stand-alone example programs are included in the `samples` subfolder. They can optionally be compiled using the CMake infrastructure of CLBlast by providing the `-DSAMPLES=ON` flag, for example as follows:
For all of CLBlast's APIs, it is possible to optionally set an OS environmental variable `CLBLAST_BUILD_OPTIONS` to pass specific build options to the OpenCL compiler.
There is also a Netlib CBLAS C API available. This is however not recommended for full control over performance, since at every call it will copy all buffers to and from the OpenCL device. Especially for level 1 and level 2 BLAS functions performance will be impacted severely. However, it can be useful if you don't want to touch OpenCL at all. You can set the default device and platform by setting the `CLBLAST_DEVICE` and `CLBLAST_PLATFORM` environmental variables. This API can be used as follows after providing the `-DNETLIB=ON` flag to CMake:
There is also a CUDA API of CLBlast available. Enabling this compiles the whole library for CUDA and thus replaces the OpenCL API. It is based upon the CUDA runtime and NVRTC APIs, requiring NVIDIA CUDA 7.5 or higher. The CUDA version of the library can be used as follows after providing the `-DCUDA=ON -DOPENCL=OFF` flags to CMake:
The CLBlast library is already tuned for the most commonly used OpenCL devices and it's gradually being extended to other devices as well. For unseen devices CLBlast will make use of common-best tuning values for similar architectures (e.g. AMD Fiji) or in general similar devices (e.g. AMD GPUs), so performance might still be decent. The current release of CLBlast is tuned for the following devices:
If your device is not (yet) among this list or if you want to tune CLBlast for specific parameters (e.g. rectangular matrix sizes), you should compile the library with the optional tuners by specifing `-DTUNERS=ON`, for example as follows:
Compiling with `-DTUNERS=ON` will generate a number of tuners, each named `clblast_tuner_xxxxx`, in which `xxxxx` corresponds to a `.opencl` kernel file as found in `src/kernels`. These kernels corresponds to routines (e.g. `xgemm`) or to common pre-processing or post-processing kernels (`copy` and `transpose`). Running such a tuner will test a number of parameter-value combinations on your device and report which one gave the best performance. Running `make alltuners` runs all tuners for all precisions in one go. You can set the default device and platform for `alltuners` by setting the `CLBLAST_DEVICE` and `CLBLAST_PLATFORM` environmental variables.
The tuners output a JSON-file with the results. The best results need to be added to `src/database/kernels/xxxxx.hpp` in the appropriate section. However, this can be done automatically based on the JSON-data using a Python (2.7 or 3.x) script in `scripts/database/database.py`. If you want the found parameters to be included in future releases of CLBlast, please attach the JSON files to the corresponding issue on GitHub or [email the main author](http://www.cedricnugteren.nl).
Alternatively, you can also supply your tuning parameters programmatically through the CLBlast API. This is especially useful if you tune for specific non-standard arguments (e.g. a rectangular or a very small matrix). To do so, you can call the `OverrideParameters` function which will set new parameters for a specific kernel. At the first next call of the target routine, CLBlast will compile a new binary and use it together with the new parameters from then on. Until `OverrideParameters` is called again of course. See the [API documentation](doc/clblast.md#overrideparameters-override-tuning-parameters-auxiliary-function) for more details.
After the kernels are tuned, you can run the `clblast_tuner_routine_xgemm` tuner to optimize the high-level GEMM routine, i.e. selecting which method to use: the direct kernel or the in-direct kernel.
To make sure CLBlast is working correctly on your device (recommended), compile with the tests enabled by specifying `-DTESTS=ON`, for example as follows:
To build these tests, another BLAS library is needed to serve as a reference. This can be either:
* The OpenCL BLAS library [clBLAS](http://github.com/clMathLibraries/clBLAS) (maintained by AMD)
* A regular CPU Netlib BLAS library, e.g.:
- OpenBLAS
- BLIS
- Accelerate
Afterwards, executables in the form of `clblast_test_xxxxx` are available, in which `xxxxx` is the name of a routine (e.g. `xgemm`). Note that CLBlast is tested for correctness against [clBLAS](http://github.com/clMathLibraries/clBLAS) and/or a regular CPU BLAS library. If both are installed on your system, setting the command-line option `-clblas 1` or `-cblas 1` will select the library to test against for the `clblast_test_xxxxx` executables. All tests have a `-verbose` option to enable additional diagnostic output. They also have a `-full_test` option to increase coverage further.
All tests can be run directly together in one go through the `make alltests` target or using CTest (`make test` or `ctest`). In the latter case the output is less verbose. Both cases allow you to set the default device and platform to non-zero by setting the `CLBLAST_DEVICE` and `CLBLAST_PLATFORM` environmental variables. Further options can be supplied through the `CLBLAST_ARGUMENTS` environmental variable (e.g. export CLBLAST_ARGUMENTS="-full_test -cblas 1 -clblas 0" on a UNIX system).
To test the performance of CLBlast and compare optionally against [clBLAS](http://github.com/clMathLibraries/clBLAS), cuBLAS (if testing on an NVIDIA GPU and `-DCUBLAS=ON` set), or a CPU BLAS library (see above for requirements), compile with the clients enabled by specifying `-DCLIENTS=ON`, for example as follows:
The performance tests come in the form of client executables named `clblast_client_xxxxx`, in which `xxxxx` is the name of a routine (e.g. `xgemm`). These clients take a bunch of configuration options and directly run CLBlast in a head-to-head performance test against optionally clBLAS and/or a CPU BLAS library. You can use the command-line options `-clblas 1` or `-cblas 1` to select a library to test against.
On [the CLBlast website](https://cnugteren.github.io/clblast) you will find performance results for various devices. Performance is compared in this case against a tuned version of the clBLAS library and optionally also against cuBLAS. Such graphs can be generated automatically on your own device as well. First, compile CLBlast with the clients enabled. Then, make sure your installation of the reference clBLAS is performance-tuned by running the `tune` executable (shipped with clBLAS). Finally, run the Python/Matplotlib graph-script found in `scripts/benchmark/benchmark.py`. For example, to generate the SGEMM PDF on device 1 of platform 0 from the `build` subdirectory:
Note that the CLBlast library provides pre-tuned parameter-values for some devices only: if your device is not among these, then out-of-the-box performance might be poor. See above under `Using the tuners` to find out how to tune for your device.
In case performance is still sub-optimal or something else is wrong, CLBlast can be build in verbose mode for (performance) debugging by specifying `-DVERBOSE=ON` to CMake.
CLBlast supports almost all the Netlib BLAS routines plus a couple of extra non-BLAS routines. The supported BLAS routines are marked with '✔' in the following tables. Routines marked with '-' do not exist: they are not part of BLAS at all. The different data-types supported by the library are:
The half-precision fp16 format is a 16-bits floating-point data-type. Some OpenCL devices support the `cl_khr_fp16` extension, reducing storage and bandwidth requirements by a factor 2 compared to single-precision floating-point. In case the hardware also accelerates arithmetic on half-precision data-types, this can also greatly improve compute performance of e.g. level-3 routines such as GEMM. Devices which can benefit from this are among others Intel GPUs, ARM Mali GPUs, and NVIDIA's latest Pascal GPUs. Half-precision is in particular interest for the deep-learning community, in which convolutional neural networks can be processed much faster at a minor accuracy loss.
Since there is no half-precision data-type in C or C++, OpenCL provides the `cl_half` type for the host device. Unfortunately, internally this translates to a 16-bits integer, so computations on the host using this data-type should be avoided. For convenience, CLBlast provides the `clblast_half.h` header (C99 and C++ compatible), defining the `half` type as a short-hand to `cl_half` and the following basic functions:
*`half FloatToHalf(const float value)`: Converts a 32-bits floating-point value to a 16-bits floating-point value.
*`float HalfToFloat(const half value)`: Converts a 16-bits floating-point value to a 32-bits floating-point value.
For deployment on Android, there are three options to consider.
First of all, you can use Google's recommended route of installing Android Studio with the NDK, and then use the JNI to interface to the CLBlast library. For this, we refer to the official Android Studio documentation and the online tutorials.
Alternatively, you can cross-compile the library and the test/client/tuner executables directly. To do so, first install the NDK, then find your vendor's OpenCL library (e.g. in `/system/vendor/lib`), get OpenCL headers from the Khronos registry, and invoke CMake as follows:
For any potential issues, first check [cmath 'has not been declared' errors](https://stackoverflow.com/questions/45183525/compilation-error-with-ndk-using-cstatic/46433625). Also, if you are encountering errors such as `#error Bionic header ctype.h does not define either _U nor _CTYPE_U`, make sure CMake is not including system paths.
Contributions are welcome in the form of tuning results for OpenCL devices previously untested or pull requests. See [the contributing guidelines](CONTRIBUTING.md) for more details.
Hardware/software for this project was contributed by:
* [ArrayFire](http://arrayfire.org) for settings up and supporting Jenkins CI correctness tests on 7 platforms
* [JetBrains](https://www.jetbrains.com/clion/) for supply a free CLion IDE license for CLBlast developers
* [Travis CI](https://travis-ci.org/CNugteren/CLBlast/branches) and [AppVeyor](https://ci.appveyor.com/project/CNugteren/clblast) for free automated build tests for open-source projects
* A 20-minute presentation of CLBlast was given at the GPU Technology Conference in May 2017. A recording is available on the [GTC on-demand website](http://on-demand.gputechconf.com/gtc/2017/video/s7280-nugteren-clblast.mp4) (poor audio quality however) and a full slide-set is also available [as PDF](http://on-demand.gputechconf.com/gtc/2017/presentation/s7280-cedric-nugteren-clblast.pdf).
* More in-depth information and experimental results are also available in a scientific paper titled [CLBlast: A Tuned OpenCL BLAS Library](https://arxiv.org/abs/1705.05249) (May 2017). For CLTune, the inspiration for the included auto-tuner, see also the [CLTune: A Generic Auto-Tuner for OpenCL Kernels](https://arxiv.org/abs/1703.06503) paper.
This project started in March 2015 as an evenings and weekends free-time project next to a full-time job for Cedric Nugteren. If you are in the position to support the project by OpenCL-hardware donations or otherwise, please find contact information on the [website of the main author](http://cnugteren.github.io).