A faster SNN simulator which supports multiple GPGPUs.

The simulator only intergrates the LIF model and support for clusters is under development.

When using this software, please use the following citation:

@ARTICLE{9091320,
  author={Qu, Peng and Zhang, Youhui and Fei, Xiang and Zheng, Weimin},
  journal={IEEE Transactions on Parallel and Distributed Systems},
  title={High Performance Simulation of Spiking Neural Network on GPGPUs},
  year={2020},
  volume={31},
  number={11},
  pages={2510-2523},
  doi={10.1109/TPDS.2020.2994123}}

• g++ 5.4.0
• CUDA 9.0
• mpich 3.2-6build1
• libblas 3.6.0

Tutorial:

A typical SNN is defined as follows:

	#include "include/BSim.h"

	...

	Network c;
	// Define a population of N1 LIF neurons.
	Population<LIF_brian> *p0 = c.createPopulation(N1, LIF_brian(....));
	// Define a population of N2 LIF neurons.
	Population<LIF_brian> *p1 = c.createPopulation(N2, LIF_brian(....));

	...

	float * weight = getRandomArray(N1*N2);
	float * delay = getRandomArray(N1*N2);
	// Define an all-to-all projection between the two population
	c.connect(p0, p1, weight, delay, NULL, N1*N2);

	// Define the simulator and set the simulation step to 0.1ms
	SGSim sg(&c, 1e-4 /*dt*/);
	// Run the simulation for 0.1s
	sg.run(0.1)

To customize neuron/synapse models, one can refer to the code under neuron/synapse directory. We will provide interfaces that are easier to use later.

Compiling:

Put your customized examples under test/gpu and run the build/bin/build.sh script

$ bash ./build.sh release float

The first parameter is release/log/debug. Release parameter provides the release build and records the overall firing rate of each neuron in GFire.log file. Log parameter further records the membrane voltage in g_v.data file. Debug paramter provides the debug build.

The sceond paramter is float/double, which indicates whether to use single/double floating point numbers.

Samples:

At the moment, the following example SNNs are provides:

• test/gpu/standard_test.cpp: It provides a CUBA IF network, which is a forward network along with several cross-population projections. About 80% of all the populations only connect with its front and rear populations, and the rest 20% also connect with remote populations. Other factors, such as the number of populations, the average number of neurons in one population, the firing rate, are all configurable.

• Toy examples

Tests:

After compiling, the binary files are generated under test/build/bin directory. And their names are kept the same as the source files.

For standard_test sample:

• 100Hz firing rate:

$ ./standard_test number_of_populations number_of_neurons_per_population 0.7 0.5 0.6 0.3 6

• 500Hz firing rate:

$ ./standard_test number_of_populations number_of_neurons_per_population 0.7 0.9 0.6 0.2 6

• 2000Hz firing rate:

$ ./standard_test number_of_populations number_of_neurons_per_population 1.3 1 2 1 50

Branches:

Currently, different branches provide different optimization methods and/or for different devices:

• master: It provides most of the stable optimization methods.
• c_and_s: It provides both the cross-population/-projection parallelism exploration and sparsity aware load balance.
• no_s_synapse: It drops the sparsity aware load balance for synapses.
• no_s_neuron: It drops the sparsity aware load balance for neurons.
• dgx1: It supports multiple GPGPUs on NVIDIA DGX-1 server.
• develop: The under development version.