Skip to content

SNN-MLIR

CI Documentation Status PyPI arXiv Collaboration Network

An out-of-tree MLIR dialect for Spiking Neural Networks (SNNs), compatible with the NIR (Neuromorphic Intermediate Representation) standard.

The dialect provides type-polymorphic operations that work with both f32 (float) and quantized (i8/i32) types, enabling a single IR to target both simulation and hardware-optimized deployments. A reference CPU lowering (SNNToLinalg) converts SNN ops to standard linalg/arith operations that any MLIR-based backend can consume.

A companion Python package (snn-mlir, available on PyPI) reads any NIR file and emits SNN dialect MLIR text, ready to feed into the snn-opt lowering toolchain. (The C runtime files used in the examples are generated separately by examples/_codegen.py, which is not part of the installable package.)

snn-mlir compilation flow: NIR → SNN dialect MLIR → LLVM IR → executable

Supported simulators

NIR is supported as an export target by a growing list of SNN frameworks. Any of them can in principle feed snn-mlir; the table below shows which we have tested end-to-end through the MLIR pipeline and which ship as a runnable example.

Framework Exports to NIR Tested through MLIR Example
LAVA / lava-dl ✅ ✅ SNN Oxford
snnTorch ✅ ✅ SNNTorch
hxtorch (BrainScaleS-2) ✅ — —
Nengo ✅ — —
Norse ✅ — —
Rockpool ✅ — —
Sinabs ✅ — —

Does your simulator export to NIR but isn't tested here?

We'd love to extend coverage. If your framework writes a NIR graph that snn-mlir doesn't yet handle, you are very welcome to send us the NIR graph and we'll test it and add support. See Contributing.

Which path is for you?

snn-mlir sits at the boundary between the neuromorphic and compiler worlds, so there are two natural entry points:

  • Coming from SNNs / neuromorphics? You'll most likely care about the Python package and the translation from NIR to MLIR — exporting your trained network and getting a portable, embedded-ready representation out.
  • Coming from compilers / embedded systems? You'll most likely care about the SNN MLIR dialect and lowering it to your hardware — bringing your own backend (CPU, FPGA, ASIC) under a shared IR.

Questions, doubts, or bugs?

We're happy to help. Reach out to the maintainers any time:

Citation

A companion paper describing snn-mlir is published on arXiv. If you use snn-mlir in your research, please cite the white paper directly:

@misc{gener2026snnmlirmlirdialectcompiling,
      title={SNN-MLIR: An MLIR Dialect for Compiling Neuromorphic SNNs from NIR to Bare-Metal C}, 
      author={Alejandro García Gener and Alvaro Rollón de Pinedo},
      year={2026},
      eprint={2606.09213},
      archivePrefix={arXiv},
      primaryClass={cs.PL},
      url={https://arxiv.org/abs/2606.09213}, 
}

License

Apache License 2.0 WITH LLVM-exception — see the LICENSE file in the repository.