Towards High-Performance Flexible FPGA-Based Accelerators for CNN Inference: General Hardware Architecture and End-to-End Deploying Toolflow

License

This project is dual-licensed under the following terms:

Open Source Use

• The source code is released under the GNU Affero General Public License v3.0 (AGPL-3.0).
  • Derivative works must comply with the AGPL's open source requirements, including making their source code publicly available under the same license.
  • For details, see the full AGPL-3.0 license text.

Commercial Licensing

• Commercial use (including but not limited to SaaS offerings, proprietary integrations, or resale) requires a paid commercial license.
• The commercial license grants rights to:
  • Use the code in closed-source products.
  • Modify the code without open-sourcing derivative works.
  • Access priority technical support and updates.
• To obtain a commercial license, contact us at wugang@mail.neu.edu.cn.

Patent Notice

This code is protected by patents (Patent No. ZL202310288488.1).

• Non-commercial use is permitted under AGPL-3.0.
• Commercial use requires explicit patent authorization.

Folder Description

/FPGA_linux : Hardware design code (Verilog and SystemVerilog) and driver code (C/C++).

/Python : Compiler code.

How to reproduce the results.

Taking the latest version as an example.

1. Generate bitstream files and burn them into FPGA.

For example, placing a bitstream file in ./FPGA_inux/bitstreams/M64P64Q16R16S8/ folder.

2. Enter ./FPGA_inux/linux-d river/run/ folder,run the script in the to collect data.

run_model_e2e_perf.sh : End to end delay for collecting various models.

run_model_ins_perf.sh: Used to collect the delay of each instruction for each model.

run_model_build_db.sh: Used to build convolutional databases, it collects the latency of each Conv instruction in both static and dynamic states.

run_test_*: Used for testing various functional components.

All collected raw data is saved in ./FPGA_inux/linux-d river/extvres/.

3. Enter ./Python/ana/，analyzing results.

Firstly, run the script merge files. py and copy all the original experimental results to this folder.

Secondly, run the script run_ana. pyto start analyzing each model, and save the results in the res_ * file under the accelerator corresponding to each model.

Then, run the script plot. py to start plotting the performance and inference delay ratio of each model.