NIR mapping¶

What NIR is, and why we support it¶

The Neuromorphic Intermediate Representation (NIR) is a framework-neutral format for describing spiking and continuous neuron networks as a graph of well-defined node types (Linear, Affine, LIF, CubaLIF, …). It is the lingua franca that lets a model trained in one framework be read by another. By consuming NIR, snn-mlir becomes framework-agnostic for free: any of the supported simulators that export NIR can target the dialect, instead of us writing a bespoke importer per framework.

NIR is broader than MLIR's digital world

NIR is designed to describe both digital and analog neuron models, using continuous physical parameters — membrane and synaptic time constants (tau_mem, tau_syn), resistance (r), leak (v_leak), threshold (v_threshold), and so on. The snn dialect, by contrast, targets digital, discrete-time execution. The frontend therefore discretizes the continuous NIR parameters into the per-timestep decay factors the dialect uses.

Discretization¶

For each neuron node, the parser derives the discrete-time update factors from the NIR physical parameters. For a CubaLIF node, for example:

dt        = tau_mem / r
cur_decay = 1 − dt / tau_syn      # current (synaptic) leak per step
vol_decay = 1 − dt / tau_mem      # voltage (membrane) leak per step
threshold = v_threshold

Integrate-and-fire variants (CubaIF, IF) are simply the leaky case with the decay set to 1.0, which disables the exponential leak. The parser also enforces the dialect's assumptions — e.g. v_leak must be 0, and tau_syn, tau_mem, v_threshold must be uniform across the layer (see Limitations).

Node mapping¶

Each SNN op covers a family of NIR nodes:

NIR node	SNN op	Notes
`nir.Linear`	`snn.linear`	No bias
`nir.Affine`	`snn.linear`	Bias added as second operand
`nir.CubaLIF`	`snn.cubalif`	`cur_decay`, `vol_decay` < 1
`nir.CubaIF`	`snn.cubalif`	`cur_decay = vol_decay = 1.0` (no leak)
`nir.CubaLI`	`snn.cubali`	`cur_decay`, `vol_decay` < 1
`nir.CubaI`	`snn.cubali`	`cur_decay = vol_decay = 1.0` (no leak)
`nir.LIF`	`snn.lif`	`decay` < 1
`nir.IF`	`snn.lif`	`decay = 1.0` (no leak)
`nir.LI`	`snn.li`	`decay` < 1
`nir.I`	`snn.li`	`decay = 1.0` (no leak)
(internal)	`snn.rescale`	Inserted between `snn.linear` and neuron ops during quantized export; no NIR equivalent

Current NIR coverage & pending nodes¶

The supported set above covers feedforward, fully-connected networks. NIR node types that are not yet mapped include the convolutional and pooling family:

nir.Conv1d / nir.Conv2d
nir.AvgPool2d / nir.SumPool2d
nir.Flatten, and other spatial/structural nodes

Adding one is a contained task — see Adding a NIR node type. If the model you care about uses an unsupported node, we'd be glad to help; send us the NIR graph.