Ctrl+K

Advanced

Advanced#

Under the hood#

The DynapcnnNetwork converts a given model into a sequence of DVSLayer (at most 1) and DynapcnnLayers.

digraph { subgraph cluster { node [shape=polygon, sides=4] label = "DynapcnnNetwork"; DVSLayer -> "DynapcnnLayer[0]" -> "DynapcnnLayer[1]" -> "..."; } }

A ConfigBuilder then converts this model to a config object when make_config method is called. make_config is called internally when a user calls the to method.

digraph { node [shape=polygon, sides=4] rankdir=LR subgraph cluster { label = "ConfigBuilder" DynapcnnNetwork -> "Samna Config" } }

ChipFactory is used to fetch the appropriate ConfigBuilder for a given device.

digraph { node [shape=polygon, sides=4] ChipFactory -> ConfigBuilder }

The config object#

The config object is a nested data structure in samna. Each device type has its own structure of the Config object.

Memory constraints#

Each spiking CNN core (DynapcnnLayer) comprises three memory blocks:

  • Kernel: To store the weights of the convolution

  • Bias: To store the biases

  • Neuron: To store the neuron states

The physical devices have a limited amount of memory available for each of its CNN cores/layers. Depending on the device architecture, each core could have a different amount of memory available. When a model is deployed on a device, one needs to ensure that each of the layers has the required amount of momory available to it. This is take care of by a mapping algorithm when chip_layers_ordering=”auto” option is set while calling make_config().

Attributes of interest#

Knowing the mapping of the various layers of the model to the layers of the chip is crucial. DynapcnnNetwork.chip_layers_ordering is a list of chip layer indices where a model was mapped. This is useful when generating or interpreting events from samna, where the layer attribute refers to the layer on the chip.

It is important to note here that the chip_layers_ordering is only pertinent to DynapcnnLayer and does not include the DVSLayer. This is because there is no ambiguity as to where the DVSLayer is located on the chip. The DynapcnnLayer layers on the other hand have multiple potential core locations. chip_layers_ordering helps specify where each of these layers is concretely placed.

Conversion between raster and spike streams#

You can use the convenience methods raster_to_events() or xytp_to_events() of the ChipFactory to generate Spike sequences of the appropriate type.

Samna: The interface library to the chip#

SynSense develops Samna, a library that handles the communication and configuration of the chip. You will find further examples and API reference of Samna on its documentation page. Documentation available here.

Contents