vela Directory Reference

Directory dependency graph for vela:

Directories
	generated
	src

Detailed Description

Arm Ethos-U Vela is the offline compiler that turns a quantized .tflite model into an Ethos-U-optimized one: the operators the NPU supports are fused into a single ethos-u custom operator wrapping a Vela-generated command stream, and the rest fall back to CPU kernels in the on-device runtime.

A build-time HOST tool, not firmware

Nothing from Vela is linked into firmware. It runs on the developer or CI host; the on-device runtime is the vendored TFLite-micro, which executes the command stream through the first-party ra8_npu driver. It is a version-pinned build dependency rather than vendored source SOUP because it is a Python tool run at build time – but it still carries a qualification record (docs/SOUP/vela.md) and a row in THIRD_PARTY_LICENSES.md, because a build tool that shapes the bits in the image is part of the supply chain.

The pin exists so the emitted command stream is reproducible and so the operator protocol Vela targets stays matched to the vendored TFLite-micro ethosu operator. Moving either one is a re-qualification against the other, not a version bump.

The .npub container and its loader (#227)

A bare-metal target does not parse a TFLite flatbuffer at run time. An offline build step distils a model into a lean, linkable container – an Ethos-U55 command stream plus its tensor region layout, defined by ra8_npu_blob.h – baked as a C byte array the firmware links. The on-target loader validates that container, resolves every region base (baked regions into the blob, runtime regions into a caller-supplied arena) and fills an ra8_npu_job_t for ra8_npu_submit().

The committed golden header is regenerated in memory and diffed against the tree, which needs no Vela toolchain at all, so CI guards the container format without ever provisioning Vela. That golden's command stream is the documented stand-in convention (ra8_npu_fake_cmd.h) – the only Ethos-U55 stream this repo can produce deterministically without a Vela install. Distilling a real Vela output into a .npub is a follow-up on the RA8P1 NPU epic: it needs both a validated Vela run and an RA8P1 board to confirm on silicon.