LASANA Circuit Plug-ins for SANA-FE

Copyright (c) 2026 — The University of Texas at Austin This work was produced under contract #2317831 to National Technology and Engineering Solutions of Sandia, LLC, which is under contract No. DE-NA0003525 with the U.S. Department of Energy.

---

Overview

This project provides analog circuit plug-ins for SANA-FE, a neuromorphic architecture simulator.

LASANA is a framework for training circuit surrogates that capture the energy, latency, and behavioural output of neuromorphic circuits as separate MLP or boosted tree models.

The plug-ins have been demonstrated on two circuits:

• An IMAC crossbar array (synapse + dendrite plug-in)
• An Indiveri adaptive integrate-and-fire neuron circuit (soma plug-in)

These surrogate models are adapted directly from the LASANA CodeOcean artifact, and allow analog circuits to be simulated extremely efficiently within large-scale architectural simulation.

---

Workflow

1. (Optional) Download circuit surrogate models

Relevant model files have been included in this repo. However, if you want to manually install the models, it is possible to download the sklearn MLP circuit surrogate models from the CodeOcean artifact linked above, for both the analog crossbar and neuron circuits.

2. (Optional) Convert models to ONNX format

python scripts/convert_joblist_to_onnx.py

3. (Optional) Compile ONNX models to object files with ONNX-MLIR

This step is optional but recommended for faster model execution. It requires ONNX-MLIR to be set up on your system.

onnx-mlir -O3 --EmitObj <onnx_file>

4. Build the plug-ins

mkdir build && cd build
cmake ..
make -j4

To build with ONNX-MLIR support instead of the default ONNX Runtime:

cmake -DUSE_ONNX_MLIR=ON ..

5. Use the plug-ins with SANA-FE

Refer to SANA-FE's documentation for details on integrating plug-ins into a simulation.