/FBLAS

BLAS implementation for Intel FPGA

Primary LanguageC++BSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

FBLAS

FBLAS is a porting of the BLAS numerical library (http://www.netlib.org/blas/) for Intel FPGA platform. For more details, see our paper.

 

Code

Requirements

The library depends on:

Installation

After cloning this repository, make sure you clone the rapidjson submodule dependency, by executing the following command:

git submodule update --init
pip install -r codegen/requirements.txt

The FBLAS library

FBLAS provides two layers of abstraction:

  • HLS modules, which can be integrated into existing hardware designs. They implement BLAS routines (DOT, GEMV, GEMM, etc.). Modules have been designed with compute performance in mind, exploiting the spatial parallelism and fast on-chip memory on FPGAs and have a streaming interface: data is received and produced using channels. In this way, they can be composed and communicate using on-chip resources rather than off-chip device RAM;

  • a high-level Host API conforming to the classical BLAS interface that allows the user to invoke routines directly from a host program. No prior knowledge on FPGA architecture and/or tools is needed. The user writes a standard OpenCL program: she is responsible to transferring data to and from the device, she can invoke the desired FBLAS routines working on the FPGA memory, and then she copies back the result from the device.

For further information on how to use the library, please refer to the wiki.

Publication

If you use FBLAS, please cite us:


@inproceedings{fblas,
    author = {De Matteis, Tiziano and de Fine Licht, Johannes and Hoefler, Torsten},
    title = {FBLAS: Streaming Linear Algebra on FPGA},
    year = {2020},
    isbn = {9781728199986},
    publisher = {IEEE Press},
    booktitle = {Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis},
    articleno = {59},
    numpages = {13},
    keywords = {high level synthesis, spatial architectures, hardware library},
    location = {Atlanta, Georgia},
    series = {SC '20}
}

Concerning the Artifact Evaluation of the paper, you will find detailed information in subfolder evaluation.

Contact

FBLAS can be used to build numerical applications, and be modified to include new features. Contributions, comments, and issues are welcome!

License

FBLAS is published under the New BSD license, see LICENSE.