npu_infer Directory Reference

Directory dependency graph for npu_infer:

Directories
	src

Detailed Description

The inference-runtime companion to npu_smoke, which covers the bare command/queue driver. This app drives the pieces a real TensorFlow Lite for Microcontrollers Ethos-U runtime needs on top of that driver, and checks each end to end on the RA8P1:

• Quantization both ways – a float input arena is quantized to the UINT8 tensor the NPU reads, and the UINT8 result is dequantized back to float.
• IRQ-driven completion – the job is awaited on the NPU interrupt, cleared from the ISR, rather than busy-waited.
• The TFLite-Micro Ethos-U operator is really linked and registered, which makes this app the proof that the vendored TFLite-Micro runtime builds and links end to end for the RA8P1.

Scope

The NPU operator run in the emulator is the tiny deterministic add-constant of the documented stand-in convention, not a real Vela command stream – that needs the offline Vela compiler and silicon. The emulator's NPU model is honest about it and rejects a real Vela stream it cannot interpret. So the quantize, NPU op, dequantize pipeline and the interrupt completion path are checkable byte for byte here, while the model-driven interpreter path is proven to link and register and runs end to end on silicon later.

There is no RA8P1 HIL rig, so this app is discovered by the emulator suite and not by the bench runner.