Skip to content

Latest commit

 

History

History
21 lines (15 loc) · 1.75 KB

README.md

File metadata and controls

21 lines (15 loc) · 1.75 KB

HDIFF_OCAPI

OpenCAPI FPGA accelerator for High Order Stencils (HOL)

  • This repository includes the implementation of an FPGA accelerator for a High Order Stencil (HOL), specifically the Horizontal Diffusion (HDIFF) from COSMO whether model.
  • The accelerator has been implemented using the OC-Accel Transprecision Framework.
  • C code computes HDIFF by moving data from the CPU host memory to FPGA over OCAPI
  • There is an option to have multiple parallel accelerators. We do procession in ROW and COLUMN while the DEPTH remains constant.

Cite

  • G. Singh, D. Diamantopoulos, C. Hagleitner, S. Stuijk and H. Corporaal, "NARMADA: Near-Memory Horizontal Diffusion Accelerator for Scalable Stencil Computations," 2019 29th International Conference on Field Programmable Logic and Applications (FPL), Barcelona, Spain, 2019, pp. 263-269, doi: 10.1109/FPL.2019.00050.
  • G. Singh, D. Diamantopoulos, S. Stuijk, C. Hagleitner, H. Corporaal, "Low Precision Processing for High Order Stencil Computations", SAMOS 2019, pp. 403-415

Acknowledgement

  • The initial software code was provided by CSCS, under the course of NeMeCo project and Ronald P. Luijten' Sabbatical at ETH.
  • The novel HW-SW co-design using Oc-Accel framework was done by Gagandeep Singh, during his research visit at IBM Zurich lab.
  • This work was funded by the European Union’s H2020 research and innovation program under grant agreement No 732631, project OPRECOMP. For details visit http://oprecomp.eu/.

License

The source code in this repository is licensed under the terms of the Apache License (Version 2.0), unless otherwise specified. See LICENSE for details.