Tuned OpenCL BLAS
Platform | Build status |
---|---|
Windows | |
Linux/macOS |
Test machine (thanks to ArrayFire) | Test status |
---|---|
clblast-linux-nvidia-a100 | |
clblast-linux-nvidia-k80 | |
clblast-linux-nvidia-p100 | |
clblast-linux-nvidia-t4 | |
clblast-linux-nvidia-v100 | |
clblast-windows-amd-r9 | |
clblast-windows-nvidia-m6000 |
CLBlast is a 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 for performance reports on some devices.
The library is not tuned for all possible OpenCL devices: if out-of-the-box performance is poor, please run the tuners first. See the docs for a list of already tuned devices and instructions on how to tune yourself and contribute to future releases of the CLBlast library.
Use CLBlast instead of clBLAS:
Use CLBlast instead of cuBLAS:
When not to use CLBlast:
CLBlast can be compiled with minimal dependencies (apart from OpenCL) in the usual CMake-way, e.g.:
mkdir build && cd build
cmake ..
make
Detailed instructions for various platforms can be found are here.
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:
#include <clblast.h>
Or alternatively the plain C version:
#include <clblast_c.h>
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. The API is kept as close as possible to the Netlib BLAS and the cuBLAS/clBLAS APIs. For an overview of the supported routines, see here.
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:
cmake -DSAMPLES=ON ..
Afterwards, you can optionally read more about running proper benchmarks and tuning the library.
More detailed documentation is available in separate files:
Known performance related issues:
Other known issues:
Routines returning an integer are currently not properly tested for half-precision FP16: IHAMAX/IHAMIN/IHMAX/IHMIN
Half-precision FP16 tests might sometimes fail based on order multiplication, i.e. (a * b) * c != (c * b) * a
The AMD APP SDK has a bug causing a conflict with libstdc++, resulting in a segfault when initialising static variables. This has been reported to occur with the CLBlast tuners.
The AMD run-time compiler has a bug causing it to get stuck in an infinite loop. This is reported to happen occasionally when tuning the CLBlast GEMM routine.
AMD Southern Island GPUs might cause wrong results with the amdgpu-pro drivers. Do configure CMake with AMD_SI_EMPTY_KERNEL_WORKAROUND
to resolve the issue, see issue #301.
Tests might fail on an Intel IvyBridge GPU with the latest Beignet. Please downgrade Beignet to 1.2.1, see issue #231.
Contributions are welcome in the form of tuning results for OpenCL devices previously untested or pull requests. See the contributing guidelines for more details.
The main contributing authors (code, pull requests, testing) can be found in the list ofGitHub contributors.
Tuning and testing on a variety of OpenCL devices was made possible by:
Hardware/software for this project was contributed by:
Further information on CLBlast is available through the following links:
How to cite this work:
Cedric Nugteren. CLBlast: A Tuned OpenCL BLAS Library. In IWOCL'18: International Workshop
on OpenCL. ACM, New York, NY, USA, 10 pages. 2018. https://doi.org/10.1145/3204919.3204924
This project started in March 2015 as an evenings and weekends free-time project next to a full-time job for Cedric Nugteren. You can find contact information on the website of the main author.