ICONS 2023 tutorial - From Training to Deployment: Build a Spiking FPGA on a $300 Budget

This is the code used in the tutorial. Both the Python scripts for the quantization aware training (QAT) and the high level synthesis parts can be found in this repository.

Training

In python you can find two Jupyter notebooks: the cheatsheet, which is the pre-compiled one, that you can read during the tutorial to follow along; the other one, instead, is empty and you can use it to follow Jason while he codes up during the tutorial. Try it!

The tutorial given by Jason is focused on QAT for SNNs.

Deploying

In cc, the C++ code used by Fabrizio is contained. In particular, in cc/include/hls4nm/lif.h you can find the C++ class describing the fully connected LIF layer. In hls-mlp, a 728-1024-10 LIF multy-layer percerptron implemented in C++ is contained. Be aware, the compilation can take a bit of time! Experiment a bit, add layers and see what the synthesizer puts out!

For more information about Vitis HLS, visit https://docs.xilinx.com/r/en-US/ug1399-vitis-hls/Introduction. If you want to download the latest version of Vitis HLS and Vivado, visit https://www.xilinx.com/support/download/index.html/content/xilinx/en/downloadNav/vitis.html. For the HLS part, before running the scripts, you should set the environment by running source <path-to-xilinx-programs>/Vitis_HLS/2023.1/settings64.sh.

Using the C++ code

The code structure is the following:

cc
└── include
    └── hls4nm
        ├── lif.h # C++ class containing the LIF fully connected layer.
        ├── params.h # The LIF parameters.
        └── utils.h # Some utilities to print out spikes, generate random weights to perform some simulations and syntheses.

hls-mlp
├── dut.cc # The component being instantiated (i.e., the top entity containing the full neural network).
├── dut.h # Neural network header.
├── gen_weights.cc # A C++ file to generate the weights for the network.
├── hls_cosim.tcl # TCL script to perform co-simulation, i.e., simulate the synthesised RTL in Vivado and bring the results to C++.
├── hls_sim.tcl # C simulation performed in Vitis HLS.
├── hls_syn.tcl # Synthesis from C++ to RTL performed in Vitis HLS.
├── hls_vivado.tcl # Export to a Vivado IP of the synthesised code, so that you can put it on your FPGA.
├── main.cc # The main file instantiating the neural network, providing inputs to it and reading the output spikes.
├── main.h # The simulation parameters (sizes of the layers, timesteps...).
└── Makefile # A Makefile to make ti easier to compile stuff and use Vitis HLS from the command line.

Take a look at hls-mlp/Makefile and modify according to the path to your AMD Xilinx tools.