2.1 KiB
rknpu2 — Rockchip's NPU runtime, fetched rather than committed
The controller talks to the RK3576's NPU through two files from Rockchip:
| File | Used by |
|---|---|
include/rknn_api.h |
build.rs, on every build with the device feature: src/rknn_layout_probe.c includes it so the Rust mirror of rknn_tensor_attr is checked against the real layout on any host |
lib/aarch64/librknnrt.so |
build.rs, under --features rknn only: the board binary links against it |
Neither is in the repository. Both are Rockchip's, distributed under the RKNN SDK License, which is proprietary: it permits use for developing applications for Rockchip products and cannot be re-granted under this repository's Apache-2.0. Rockchip publishes both files in airockchip/rknn-toolkit2, so the build takes them from there.
bash deploy/fsm/vendor/rknpu2/fetch.sh
docker-build.sh runs it before every build, so the cross-build and
docker-build.sh test need nothing extra. A native cargo build or cargo test
with the default device feature needs it run once by hand; build.rs stops
with a message pointing here when the header is missing. The builds that
scripts/deploy.py runs on the host (--no-default-features, for the browser and
play --app) do not include the header and do not need it.
The pin
fetch.sh takes release v2.3.2 and checks both files' sha256. The version is
not free to float:
deploy/convert/requirements.txtpinsrknn-toolkit2==2.3.2, and a.rknndoes not load under alibrknnrtolder than the toolkit that produced it;- the board carries
librknnrt2.3.2, and.claude/skills/deploy/scripts/check_board.pychecks it.
Moving to another release means changing all three together, and the two sha256
values in fetch.sh with them.
The files fetch.sh installs are byte-identical to the copies that were
committed here before the repository was opened (checked 2026-09-29), so a build
from a fresh clone links what the board was verified against.