Adding a NIR node type¶

Why extend the parsers?¶

The frontend only knows the NIR node types it has parsers for — everything else raises a clear "unsupported type" error. Extending coverage is how the project grows: new neuron models, the convolutional/pooling family, or any node your framework emits. We actively want these contributions, and the design keeps them small and self-contained.

The most valuable additions right now are the convolutional and pooling nodes (to lift the fully-connected-only limitation) and any neuron variant your hardware needs. If you're unsure where to start, get in touch and we'll point you at the right node.

How it's structured¶

NODE_PARSERS is the single registry mapping NIR node types to handler functions. All other per-node behavior — quantization, MLIR emission, classification traits — lives on the NodeInfo subclass itself, so adding a new NIR node type is three steps.

1. Create a `NodeInfo` subclass¶

from snn_mlir.nodes import NodeInfo
from dataclasses import dataclass

@dataclass
class MyNodeInfo(NodeInfo):
    name: str
    size: int

    # Classification traits are read-only properties on NodeInfo; override
    # the ones that apply (they default to False on the base class).
    @property
    def is_neuron(self) -> bool:
        return True

    # Override quantize() if the node has quantizable parameters (no-op by
    # default). Called once per layer before MLIR emission in quantized mode.
    def quantize(self) -> None:
        ...

    def emit_mlir(self, input_var, is_last, quantize):
        # Return (list_of_mlir_lines, output_var_name)
        ...

The is_synapse / is_neuron traits are what let graph-level logic (such as automatic snn.rescale insertion) work without isinstance checks — so set them correctly and new node types are handled by existing machinery automatically.

2. Write a parser function¶

import nir
def parse_mynode(node: nir.MyNode, name: str) -> MyNodeInfo:
    return MyNodeInfo(name=name, size=node.output_shape[0])

This is also where you discretize any continuous NIR parameters (see NIR mapping) and validate the dialect's assumptions.

3. Register it¶

from snn_mlir.nodes import NODE_PARSERS
NODE_PARSERS[nir.MyNode] = parse_mynode

Add a roundtrip test in test/Dialect/SNN/ for the new op and a Python unit test for the parser, then open a PR — see Contributing.