streamnet

A directed netlist representation for stream processing

What is a streamnet object?

A streamnet object (for lack of an existing term) is essentially a directed graph composed of a set of elements, a set of external input ports, a set of external output ports where:

• Elements are nodes with labeled input and output pins. They can be zero, in which case they become internal sources or sinks of information.
• External input ports are nodes with an indegree of zero
• External output ports are nodes with an outdegree of zero

These nodes are connected by directed edges with the following restriction:

• Input pins in each element must have an indegree of one
• There are no restrictions in the outdegree of output pins, including zero

Note that every streamnet object is also an element.

What does this streamnet represent?

Streamnets represent a subset of circuits and architectures where information flows between components in a directional, causal way.

What is the execution model of this network?

The execution model is essentially a flavor of dataflow programming:

A streamnet object has a method step that puts new inputs into the input ports of the network. The step instruction is broadcasted to all elements in the network.

Each element is either its own Streamnet object or an arbitrary object with a similar interface. Elements can be stateful and/or have side effects outside the streamnet, such as logging.

When a step instruction is broadcasted inside a streamnet, all elements pulls the existing input states and carry out the computation. This is consistent with a synchronous execution model, though elements can implement delays and queues. They can also pass information probabilistically to incorporate asynchronous behavior.

Wait, is this a new type of object?

There are many dataflow or flow programming languages that probably implement similar abstractions. Streamnet is a variation on netlists in the sense that it is directed and input pins must have an indegree of one. It is also a generalization of flow networks.

The purpose here is to serve as a unifying model to tie in circuits, digital architectures, and spiking neural networks.

Streamnet implementaation

This repository contains a Python implementation of a Streamnet. Obviously the goal is to recreate this model in a way that takes advantage of parallelism.

Team

Streamnet is being developed by Angel Yanguas-Gil as part of Threadwork, a project funded through DOE.