Projects

Featured · 2022 — present · ACT Lab, UCSD · Active

GeneSys

An open-source parameterizable NPU generator with a full-stack, multi-target compilation stack.

It lets you produce complete acceleration systems for emerging deep learning models — convolutional networks, transformers — with minimal human intervention.

My primary contribution has been the continued development of the compiler, originally written by Sean Kinzer. I packaged it with pip, implemented a greedy memory allocation strategy that cut the DRAM footprint substantially during execution, removed code bloat left over from stitching several projects together, and expanded layer support to accommodate more diverse models.

Project site → · GitHub → · Write-up →