Using 2D Convolutions with SpiNNaker

Follow me on GitHub

Populations in PyNN on SpiNNaker are typically 1-dimensional. This can make handling of 2D data on the platform inefficient. To address this, there are now branches that support the partitioning and placement of Populations in 2D. This means that connections between these Populations can be handled more efficiently, depending on that connectivity; in particular, convolutional connectivity is handled with a better division of labour between cores, and fewer unncessary

In addition to the better routing of spike packets, the software updates also include better handling of the convolution processing itself. In particular, the core hold the kernel weights in local memory and applies them to the incoming data dynamically. This avoids the need to transfer data from SDRAM, and so speeds up the processing significantly.

The SpiNNaker software current supports 2D representations and convolutions on the extdev_fpgas branches of the git repositories. These can be used by following the instructions here and then switching the following branches to extdev_fpgas (e.g. cd <module>; git checkout extdev_fpgas):

  • SpiNNMachine
  • SpiNNFrontEndCommon
  • sPyNNaker
  • JavaSpiNNaker (optional: only if you have use_java=True in your config file)

Once the branches are on the correct version of the software run the following:

  • SupportScripts/
  • mvn -f JavaSpiNNaker -DskipTests=True clean package

Once support is in place, 2D Populations can be used as in the following example:

import spynnaker8 as p
import numpy
from import Grid2D

# The rectangle of neurons per core


# Set the number of neurons per core to a rectangle
p.set_number_of_neurons_per_core(p.IF_curr_exp, (SUB_WIDTH, SUB_HEIGHT))

# Make a kernel and convolution connector
k_shape = numpy.array([5, 5], dtype='int32')
k_size =
kernel = (numpy.arange(k_size) - (k_size / 2)).reshape(k_shape) * 0.1
conn = p.ConvolutionConnector(kernel)

# Start with an input shape, and deduce the output shape
in_shape = (11, 11)
out_shape = conn.get_post_shape(in_shape)
n_input =
n_output =

# Make a 2D source that spikes at the middle if the input shape
spike_idx = ((in_shape[1] // 2) * in_shape[0]) + (in_shape[1] // 2)
spike_times = [[1.0] if i == spike_idx else [] for i in range(n_input)]

# Note the structure=Grid2D, which makes this a 2D population
src = sim.Population(
    n_input, sim.SpikeSourceArray, {'spike_times': spike_times}, 
    label='input spikes', structure=Grid2D(in_shape[0] / in_shape[1]))
# Make a 2D target Population and record it
output = sim.Population(
    n_output, sim.IF_curr_exp(), label="out",
   structure=Grid2D(out_shape[0] / out_shape[1]))

# Connect the two populations with the convolution.  Note the use
# of the synapse_type to ensure fast convolutional processing is
# done
p.Projection(src, output, conn, p.Convolution())