This repository contains a cycle-accurate SystemC Transaction-level Modeling (TLM) formalism (software) and a Vivado hardware implementation of a sparsity-aware spiking neural network (SNN). The network uses leaky integrate-and-fire (LIF) neurons and event-driven computation. The network parameters and input spikes are parsed from files generated by snnTorch. The model is capable of flexibly assigning hardware resources based on workload characteristics of the neural net.
The related paper has been accepted in IEEE JETCAS'23 and is also available on arXiv.
The software uses some global variables and constants defined in "defines.h". You can modify them according to your needs. Some of them are:
- INPUT_LAYER_SIZE: The number of neurons in the input layer.
- LAYER0_SIZE_LOGICAL: The number of neurons in the first hidden layer.
- LAYER1_SIZE_LOGICAL: The number of neurons in the second hidden layer.
- LAYER2_SIZE_LOGICAL: The number of neurons in the output layer.
- L0_NEURON_UNIT_SIZE: The neuron unit size in the first hidden layer.
- L1_NEURON_UNIT_SIZE: The neuron unit size in the second hidden layer.
- L2_NEURON_UNIT_SIZE: The neuron unit size in the output layer.
- TIME_STEP: The number of time steps for the spike train length.
To run the SystemC software, follow these steps:
- Make sure you have SystemC installed on your system and set the
SYSTEMCenvironment variable to point to its installation directory. - Run
./custom_run.shto execute the shell script that will generate neuron modules, parse data files, compile and run the SNN model. - Wait for the simulation to finish and check the output file
trace_log.txtfor the simulation time and output spikes.
Alternatively, you can run each step manually using the commands in custom_run.sh.
MIT