BFE EngineeringandClaude Fable 5
5155224594
Preparing the repo to go public (ADR-0007): - CI: kernel-build moves from the self-hosted runner to ubuntu-latest (installs its own cross toolchain + qemu, caches the pristine tarball). On a public repo a registered self-hosted runner is reachable from approved fork-PR workflows — i.e. arbitrary code on private infrastructure — and the build never actually needed the SDK host. ADR-0004 marked superseded-in-part; docs/ci-cd.md rewritten (site specifics now live only in the private deployment log). - Licensing: LICENSE gains the GPL-2.0 carve-out for patches/ and the kernel source excerpts (Linux derivatives; per-driver provenance was already tracked in PROVENANCE.md); patches/README.md states it too. - Scrubbed from the tip: bench-unit dev credentials and its gadget IP (m2-boot notes), the site AP SSID+BSSID and a neighboring AP's BSSID and the device WLAN MAC (wifi bring-up evidence — BSSIDs are geolocatable), the runner mesh IP. NOTE: these remain in git history; decision on a pre-publication history rewrite is separate. - Emoji cleanup across 21 tracked files (kernel port docs, review report, enforce-mcdc.sh) per repo text conventions: status marks became [x]/[wip]/[ ]/OK plain text. - "[maintainer]-gated" process phrasing normalized to "maintainer-gated" (attributions in dated evidence docs kept). Verified: zero emojis tracked; scrub grep clean; patches carry no internal references; ci.yml parses; shellcheck unchanged. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018HUayid7W5w7jBdb9Rrj1K
2.3 KiB
NPU (rknpu) open kernel driver — VERIFIED on warden-c8a3 (2026-08-25)
The open GPL rknpu kernel driver runs on our self-built Linux 6.18.46. This is the
achievable open end state (Tier A in OPEN-NPU-PLAN.md); open compute remains a
from-scratch RE project (see the ceiling note below).
Evidence (serial, _b slot = our 6.18)
[drm] Initialized rknpu 0.9.2 for ff660000.npu on minor 1
/dev/dri/ -> card0 (VOP display) card1 (rknpu) renderD128
$ rknpu_version_test
/dev/dri/card0: not an rknpu node (version ioctl: Invalid argument)
opened /dev/dri/card1 (fd=3)
driver version: 0.9.2 (raw code 902)
hw version: 0x54524548 (raw)
PASS: /dev/dri/card1 answered both version-query ioctls -- probe,
power-get/put, and clock/reset all exercised.
The RKNPU_GET_DRV_VERSION/RKNPU_GET_HW_VERSION ioctls exercise the full
dispatch → power-domain get/put → clock/reset path, so a clean answer confirms the
whole driver bring-up, not just registration.
The fix that took it from probe-fail to PASS
The base dtsi npu@ff660000 node declares its interrupt but has no
interrupt-names; the rknpu driver requests its IRQ by name ("npu_irq"), so
probe bailed error -ENXIO: IRQ npu_irq not found and never registered its DRM
device. Board DTS override adds interrupt-names = "npu_irq"; (+ status="okay").
(The rest of the port — GPL source, 4 compat-shim headers for dead-code vendor
headers, 10 mechanical 6.18 API deltas — is in PORT-PROGRESS.md.)
Also: the version test must iterate DRM cards and keep the one that ANSWERS the
ioctl — the display card (card0) opens fine but returns EINVAL. Fixed in
rknpu_version_test.c.
The honest ceiling (why "100% open NPU" stops at the driver)
The kernel driver only DMAs an opaque userspace-authored regcmd blob into the PC
registers and pulses go — it never inspects the compute stream. The compute-engine
register map (TRM Part 2) is not public for RV1106, there is zero open RE prior
art for this NPU generation, and only the closed RKNN-Toolkit2 compiler emits
valid regcmd. Mainline accel/rocket + Mesa Teflon are RK3588-only (64-bit). So an
open compiler/runtime is a from-scratch, ~person-year register-RE project, scoped
(not staffed) in OPEN-NPU-PLAN.md (M-NPU-2 spike / M-NPU-3). We ship the open
driver; we do NOT ship any closed blob.