FlexCNN

Publication

• Suhail Basalama, Atefeh Sohrabizadeh, Jie Wang, Licheng Guo, Jason Cong. FlexCNN: An End-to-End Framework for Composing CNN Accelerators on FPGA. In TRETS, 2022.

• Atefeh Sohrabizadeh, Jie Wang, Jason Cong. End-to-End Optimization of Deep Learning Applications. In FPGA, 2020.

About

This repo contains the codes for building FlexCNN, an accelerator for running CNNs on FPGA, described in the papers above. As mentioned in the papers, you can further integrate FlexCNN to TensorFlow and offload CNN computation of your application to FPGA.

The latest version of FlexCNN is tested on U-Net, E-Net, and VGG16.

Content

1. Hardware and Operating System
2. Requirements and Dependencies
3. Testing and Deployment
4. Citation

Hardware and Operating System

Development

For development, the OS should be Ubuntu 18.04.6 LTS

Requirements and Dependencies

Requirements

FlexCNN generate Vitis HLS code or TAPA HLS code (TAPA code is optimized for better P&R). The netwrorks are tested using Vitis 2021.2. To install TAPA, follow this tutorial.

Dependencies

You should have Python 3.8 installed with the libraries in requirements.txt

Testing and Deployment

For testing and depolyment, despite the os requirement above, the server/PC should also be equipped with Alveo U250 or U280 Data Center Accelerator Cards.

Run the Tested CNNs

1. source env.sh
       mkdir $STREAM_VSA_PATH/auto_compile/data/onnx

2. Download the ONNX file for U-Net, E-Net, or VGG16, and put it in

   $STREAM_VSA_PATH/auto_compile/data/onnx/

3. Generate the the design for CNN specified in generate_design.sh:

   ./generate_design.sh

   echo 13 | $STREAM_VSA_PATH/auto_compile/run.sh ENet ENet 32 u250 4 TAPA_1

   This command runs the framework for ENet targeting 32-bit float implementation on u250 generating TAPA code. '4' is used for memory allocation between BRAMs and URAMs.

   This will generate a design under the ''designs'' directory

4. To run the C-simultation, run:

   cd designs/TAPA_1_ENet_8_9_8_MEM_4_32
   source env.sh # to set up TAPA and Xilinx libraries
   cd sim
   make csim_all
   #./host.exe start_instruction_id end_instruction_id
   ./host.exe 1 89 # for ENet

   C-simulation should be outputed to sim/outputs/csim_output.log

5. To generate the bitstream, run:

   cd designs/TAPA_1_ENet_8_9_8_MEM_4_32
   source env.sh # to set up TAPA and Xilinx libraries
   cd sim
   make hw_all
   #./host.exe start_instruction_id end_instruction_id
   ./host.exe 1 89 --bitstream=path_to/your_bitstream.xclbin # for ENet

Citation

If you find any of the ideas/codes useful for your research, please cite our papers:

@article{basalama2022flexcnn,
    title={FlexCNN: An End-to-End Framework for Composing CNN Accelerators on FPGA},
    author={Basalama, Suhail and Sohrabizadeh, Atefeh and Wang, Jie and Guo, Licheng and Cong, Jason},
    journal={ACM Transactions on Reconfigurable Technology and Systems},
    year={2022},
    publisher={ACM New York, NY}
}

@inproceedings{sohrabizadeh2020end,
  title={End-to-End Optimization of Deep Learning Applications},
  author={Sohrabizadeh, Atefeh and Wang, Jie and Cong, Jason},
  booktitle={The 2020 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays},
  pages={133--139},
  year={2020}
}