An Efficient Data Structure for Sparse Bit-Vectors with Applications in Neuromorphic Computing

Iterating over all the elements of a set is a very common problem in highly parallel systems. In hardware, this is typically realized by either storing the set-membership of each element in a fixed-length bit-vector or by storing just the indices of the members in a dynamically sized queue. However, the former solution is only efficient in terms of memory and runtime if each element is a member of the set with approximately 50% probability, whereas the latter is only efficient if the set is extremely sparse. We propose an alternative asynchronous, concurrent, distributed data structure based on a binary tree topology that is more efficient for sets with sparsity in between these two extremes. The proposed structure allows us to construct a set by adding individual elements one at a time in arbitrary order, and to iterate over all these elements exactly once, clearing the set in the process. We analyzed this data structure, simulated its behavior in CHP, synthesized it into an asynchronous digital circuit, optimized the circuits, and performed SPICE simulation to evaluate our design. The results confirm that our proposed structure offers a low-latency, low-power solution for moderately sparse data, and may thus prove useful for asynchronous and neuromorphic systems.