Docs / API Reference

API Reference

Complete reference for the NMC Python SDK (import neurmorph), the CLI, the C runtime API (nrm_runtime.h), and the HAL shim interface.

Python SDK

Install with pip install nmc-cli. The SDK exposes compilation, calibration, inspection, and device management functions.

Functionneurmorph.compile(model_path, target, calibrate=True, cal_file=None)

Compile a TorchScript or ONNX model to a .snn binary. Returns a CompileResult object with output path, accuracy delta, and memory stats.

Parameter	Type	Description
model_path	str / Path	Path to .pt or .onnx export
target	str	"nt3000" \| "brainpulse2" \| "syncore-v"
calibrate	bool	Run threshold calibration pass (recommended)
cal_file	str \| None	Path to pre-computed .cal file (skips calibration pass if provided)

Functionneurmorph.inspect(snn_path)

Parse a .snn binary and return a ModelInfo object with neuron count, layer topology, SRAM requirements, and calibration checksum.

Functionneurmorph.deploy(snn_path, device)

Flash a compiled .snn binary to a connected eval kit device. Device string format: nt3000:usb0.

CLI Reference

The nmc CLI is installed alongside the Python package. All commands support --json for machine-readable output and --verbose for pass-level diagnostics.

Command	Description
`nmc compile <model> --target <chip>`	Compile model to .snn binary
`nmc calibrate <model> --dataset <dir>`	Run threshold calibration, save .cal file
`nmc inspect <binary.snn>`	Print binary metadata and memory layout
`nmc deploy <binary.snn> --device <dev>`	Flash binary to connected device
`nmc infer --device <dev> --sample <npy>`	Run one inference and print result + latency
`nmc doctor`	Check environment: Python, drivers, device detection
`nmc --version`	Print NMC version

C Runtime API

Include nrm_runtime.h in your embedded project. Requires a HAL shim implementation for your target chip.

Function	Description
`nrm_load(model, data, arena)`	Load .snn binary data into static arena. Returns 0 on success.
`nrm_step(model)`	Execute one timestep window. Blocks until complete.
`nrm_reset(model)`	Reset neuron state to post-load initial values.
`nrm_arena_init(arena, base, size)`	Initialize static memory arena from a base pointer and byte size.
`nrm_get_output(model, buf, len)`	Copy output spike vector to caller-allocated buffer.

HAL Shim API

Port the runtime to a new chip by implementing the 8-function HAL shim. Provide your implementations in a C source file and link against the runtime library.

Symbol	Signature	Description
`hal_spike_read`	`int(spike_t *buf, size_t n)`	Read up to n spikes from input bus into buf
`hal_spike_write`	`int(const spike_t *buf, size_t n)`	Write n spikes to output bus
`hal_mem_read`	`void(void *dst, addr_t src, size_t n)`	Read n bytes from chip address space
`hal_mem_write`	`void(addr_t dst, const void *src, size_t n)`	Write n bytes to chip address space
`hal_timer_us`	`uint32_t(void)`	Return monotonic microsecond counter
`hal_sleep_us`	`void(uint32_t us)`	Sleep for at least us microseconds
`hal_power_mode`	`void(nrm_power_t mode)`	Set chip power state (ACTIVE \| SLEEP \| DEEP_SLEEP)
`hal_log`	`void(const char *msg)`	Optional: output diagnostic string (may be a no-op)