This project showcases the custom quantised CNN hardware inference architecture for SoC-based wireless radio receivers like the AMD RFSoC.
The CNN model receives data samples from the RFSoC ADC and classifies the modulation scheme the received signal is encoded with.
Comparing 4 levels of quantisation:
- 16-bit weights & 16-bit activations
- 8-bit weights & 16-bit activations
- 4-bit weights & 16-bit activations
- 2-bit weights & 16-bit activations
The demonstration shows a comparison between 4 models with different weight quantisations. The demo has two notebooks, one with models deployed with Post-Training Quantisation (PTQ) and another with the model's weights trained with Quantised-Aware Training (QAT) using Brevitas. Trained weights were exported into MATLAB & Simulink and the accelerator was built using HDL Coder.
The signal is received via an RF loopback cable connected to the RF-DAC. The modulates signal has been distorted by passing it through a simulated Rician multipath channel, adding Gaussian noise at levels ranging from -20dB to 24dB, and applying time, frequency, clock offsets.
Data set construction:
The project should ideally be cloned onto a host PC first. This way the Git LFS file can be pulled properly.
- Make sure you have Git LFS installed
sudo apt install git-lfs- Clone the repo
git clone https://github.com/axdy/rfsoc_quant_amc.git- Pull the LFS file (
transmit_test_SNR.pkl~580MB)
cd rfsoc_quant_amc
git lfs pull
git lfs checkout- Compress the project into a tar.gz
tar czf rfsoc_quant_amc.tar.gz --exclude=.git --exclude=.gitattributes -C "$(pwd)" .- Copy the tar file to the notebooks folder of the PYNQ session on the RFSoC. (Wait for the file to copy over!)
Tested with PYNQ 3.0.1
Open a terminal session and run:
- Install the package
cd /home/xilinx/jupyter_notebooks
pip3 install rfsoc_quant_amc.tar.gz- Run the example notebooks within the
rfsoc_quant_amcfolder. Make sure the RFSoC is configured as described in Hardware Setup
The demo performs best with a Nooelec VeGA Low-Noise Amplifier with a low-pass filter in loopback between DAC_A and ADC_A.
Board specific hardware setups:
The following software is required to use the project files in this repository.
- Vivado Design Suite 2020.2
This project can be built with Vivado from the command line. Open Vivado 2020.2 and execute the following into the tcl console:
cd /<repository-location>/boards/<board-name>/rfsoc_quant_amc/
Make the Vivado project by running the make commands below:
make all
The work demonstrated in this repository has been published in the following paper:
Andrew Maclellan,
"Streaming-based CNN Architecture for Physical Layer Wireless Receivers: Deployment-Oriented Dataset Generation and Quantisation Methods," PhD Thesis, doi:10.48730/4b4r-2d95
Andrew Maclellan, Louise H. Crockett, and Robert W. Stewart,
"RFSoC Modulation Classification With Streaming CNN: Data Set Generation & Quantized-Aware Training,"
IEEE Open Journal of Circuits and Systems, vol. 6, pp. 38–49, 2025.
doi:10.1109/OJCAS.2024.3509627
If you use this repository, please cite the above paper.
Formal citation metadata is also available via GitHub’s “Cite this repository” feature.
BSD-3 Clause