docs: ASCII typography and style normalization across all repo text
Replace every em dash with real punctuation (rewrites, not hyphen swaps) in docs, code comments, scripts, configs, and the port records; convert en dashes, curly quotes, ellipsis glyphs, arrows, and section signs to ASCII; drop machine-writing tell phrases from living docs. ADR titles now use a colon. The M2 bring-up DTS model string carried an em dash into the patch series and its record echoes; fixed at both, and the full series re-verified to apply cleanly onto pristine 6.18.46. One comment in freshness.h deliberately names the em dash glyph the UI renders as the unknown mark; that is data, kept as prose naming it. Verified: cargo tests (sim, config-lint, rs485-bridge), shellcheck, both driver MC/DC harnesses, patches-apply.
This commit is contained in:
BFE Engineering
@@ -1,18 +1,18 @@
|
||||
# Open NPU to 100% — plan, feasibility, and the honest ceiling
|
||||
# Open NPU to 100%: plan, feasibility, and the honest ceiling
|
||||
|
||||
**Goal as stated:** take the RV1106 NPU to "100% — open source, verified on
|
||||
**Goal as stated:** take the RV1106 NPU to "100%, open source, verified on
|
||||
hardware." **Bottom line up front:** two very different things live under that
|
||||
sentence, and only one of them is reachable soon.
|
||||
|
||||
1. **Open, on-hardware-verified _kernel driver_** (`/dev/dri/cardN` binds, answers
|
||||
a version-query ioctl on our self-built Linux 6.18.46). **Achievable now**,
|
||||
low-single-digit engineer-days, license-clean GPL forward-port. This is the
|
||||
`PORT-PLAN.md` work in this directory — a real, bounded milestone (M6).
|
||||
`PORT-PLAN.md` work in this directory, a real, bounded milestone (M6).
|
||||
2. **Open _userspace_ that runs a real model** (an open compiler/encoder emitting a
|
||||
valid `regcmd` stream this driver can submit, no closed `librknnrt`, no `.rknn`
|
||||
blob). **Not achievable on any near horizon.** It is a from-scratch,
|
||||
per-SoC register reverse-engineering project with **zero existing public prior
|
||||
art for RV1106/RV1103** — the least-covered tier of the entire RKNPU family.
|
||||
art for RV1106/RV1103**, the least-covered tier of the entire RKNPU family.
|
||||
|
||||
This document scopes both, states the most ambitious end state that is actually
|
||||
achievable, and defines the concrete first milestone toward open compute.
|
||||
@@ -23,7 +23,7 @@ document is the strategic wrapper and the item-(2) reality check. Read both.
|
||||
|
||||
## 1. Brutally honest feasibility
|
||||
|
||||
### 1.1 "Port the GPL kernel driver" — TRACTABLE (do it)
|
||||
### 1.1 "Port the GPL kernel driver": TRACTABLE (do it)
|
||||
|
||||
- The vendor driver (`flare-edge/sdk/sysdrv/source/kernel/drivers/rknpu/`, v0.9.2,
|
||||
`DRIVER_DATE 20230825`) is **SPDX GPL-2.0 at the file level**, `MODULE_LICENSE("GPL v2")`
|
||||
@@ -35,26 +35,26 @@ document is the strategic wrapper and the item-(2) reality check. Read both.
|
||||
`dma_mask=32`, `pc_data_amount_scale=2`, `pc_task_number_bits=16`,
|
||||
`pc_task_status_offset=0x3c`, single-core `irqs`/`resets` arrays,
|
||||
`nbuf_phyaddr=0/nbuf_size=0`).
|
||||
- The DT node is **already fully specified** in the base tree, only disabled —
|
||||
- The DT node is **already fully specified** in the base tree, only disabled:
|
||||
`npu@ff660000`, `reg=<0xff660000 0x10000>`, `GIC_SPI 109`, clocks
|
||||
`ACLK_RKNN`/`HCLK_RKNN`, `assigned-clock-rates=<420000000>`, resets
|
||||
`SRST_A_RKNN`/`SRST_H_RKNN`, `status="disabled"` (`rv1106.dtsi:1127-1138`). One
|
||||
board-DTS override (`&npu { status="okay"; }`) enables it.
|
||||
- **No IOMMU** on this board (`"rknpu iommu device-tree entry not found!, using
|
||||
non-iommu mode"` — CMA/contiguous DMA only), **no power-domains** (single-rail),
|
||||
non-iommu mode"`, CMA/contiguous DMA only), **no power-domains** (single-rail),
|
||||
**no OPP table** (fixed clock). This _shrinks_ the port: the multi-domain/DVFS/
|
||||
thermal code is all provably dead for RV1106's DT.
|
||||
- The one real build blocker is four vendor-only downstream headers
|
||||
(`soc/rockchip/rockchip_{iommu,opp_select,system_monitor,ipa}.h`) absent from
|
||||
mainline 6.18 — all four back **dead-code call sites** for this SoC, fixable with
|
||||
mainline 6.18: all four back **dead-code call sites** for this SoC, fixable with
|
||||
small local compat-shim stubs (same pattern already used for `clk-rv1106.c`'s
|
||||
`panic_notifier_list` move). Full delta in `PORT-PLAN.md §2.4`.
|
||||
`panic_notifier_list` move). Full delta in `PORT-PLAN.md section 2.4`.
|
||||
|
||||
**Verdict: real, bounded, worth doing.** Same class as the RGA port. Ends at a
|
||||
clean probe + `RKNPU_GET_DRV_VERSION`/`RKNPU_GET_HW_VERSION` ioctl round-trip on
|
||||
hardware — **and stops there**, because of §1.2.
|
||||
hardware, **and stops there**, because of section 1.2.
|
||||
|
||||
### 1.2 "Open userspace that runs a model" — HARD → effectively BLOCKED for RV1106
|
||||
### 1.2 "Open userspace that runs a model": HARD -> effectively BLOCKED for RV1106
|
||||
|
||||
The kernel driver's entire hardware contract is: power/clock/reset, DMA/GEM buffer
|
||||
management, then it drops a **pointer + length to a userspace-authored `regcmd`
|
||||
@@ -68,20 +68,20 @@ block, from `rknpu_ioctl.h:21-39` used via `REG_WRITE` in `rknpu_job.c`):
|
||||
| Offset | Name | Meaning |
|
||||
|---|---|---|
|
||||
| `0x00`/`0x04` | VERSION / VERSION_NUM | read-only HW version |
|
||||
| `0x08` | PC_OP_EN | pulse 1→0 kicks the job; `0x1`=slave mode before commit |
|
||||
| `0x08` | PC_OP_EN | pulse 1->0 kicks the job; `0x1`=slave mode before commit |
|
||||
| `0x10` | PC_DATA_ADDR | device address of the regcmd buffer |
|
||||
| `0x14` | PC_DATA_AMOUNT | `(regcfg_amount + 4 + scale-1)/scale - 1` |
|
||||
| `0x20`/`0x24`/`0x28`/`0x2c` | INT_MASK / INT_CLEAR / INT_STATUS / INT_RAW_STATUS | interrupt handshake |
|
||||
| `0x30` | PC_TASK_CONTROL | `((0x6\|pingpong)<<pc_task_number_bits)\|task_number` (the `0x6` is an undecoded fixed mode field) |
|
||||
| `0x34` | PC_DMA_BASE_ADDR | `task_base_addr` |
|
||||
|
||||
**This is the entire openly-documented register map** — the wrapper/sequencer
|
||||
**This is the entire openly-documented register map**: the wrapper/sequencer
|
||||
only. **None of it describes the compute engine** (convolution/pooling/activation
|
||||
config, feature/weight DMA descriptors). Those live _inside_ the opaque regcmd
|
||||
buffer. The op-code semantics of that stream:
|
||||
|
||||
- **Have no public register-level documentation anywhere.** TRM Part 2 does not
|
||||
exist publicly — confirmed by exhausting Rockchip doc mirrors and confirmed
|
||||
exist publicly, confirmed by exhausting Rockchip doc mirrors and confirmed
|
||||
directly by a Luckfox engineer that **Luckfox itself was never given one.** Even
|
||||
Rockchip's own 90-page ROCKIVA SDK guide is pure userspace-API reference with
|
||||
zero register offsets. This is a deliberate architectural gate, not a
|
||||
@@ -89,21 +89,21 @@ buffer. The op-code semantics of that stream:
|
||||
- **Only the closed RKNN-Toolkit2 compiler emits valid regcmd streams.** The
|
||||
on-device closed `librknnrt` parses `.rknn` blobs and issues the same
|
||||
`RKNPU_MEM_*`/`RKNPU_SUBMIT` ioctls. `DRM_IOCTL_RKNPU_SUBMIT` technically accepts
|
||||
a raw command buffer directly — nothing _requires_ `librknnrt` — but since the
|
||||
a raw command buffer directly (nothing _requires_ `librknnrt`), but since the
|
||||
buffer format is closed, that ioctl is **not** a "write your own kernel" door.
|
||||
|
||||
### 1.3 Why there is no shortcut through the existing open efforts
|
||||
|
||||
| Open effort | Chip(s) | Reaches RV1106? |
|
||||
|---|---|---|
|
||||
| Mainline `accel/rocket` (merged 2025-07, in our `research/linux-6.18.46/`) | RK3588 only, as shipped | **No.** `Kconfig`: `depends on (ARCH_ROCKCHIP && ARM64)`. RV1106 is 32-bit `arch/arm` — excluded before the register question. `of_match` is hardcoded `"rockchip,rk3588-rknn-core"`. |
|
||||
| Mesa **Teflon** (TFLite delegate) | Only etnaviv or rocket backends | **No.** Entirely downstream of Rocket — its Rockchip reach _is_ Rocket's reach: RK3588 only. |
|
||||
| `gahingwoo/linux-rk3576-npu` + `charsiu` (active, days-old as of 2026-08) | RK3576 | **No — wrong generation.** MobileNetV1 end-to-end + a Llama-3.2-1B INT4 open runtime (`charsiu`, opens `/dev/accel/accel0`, emits its own register stream) reportedly work byte-exact — but on RK3576's NPU, a **newer generation** than RV1106's. Unmerged (LKML "[PATCH v9] accel/rocket: RK3576" in review; unverifiable by second source — Anubis-walled). |
|
||||
| Mainline `accel/rocket` (merged 2025-07, in our `research/linux-6.18.46/`) | RK3588 only, as shipped | **No.** `Kconfig`: `depends on (ARCH_ROCKCHIP && ARM64)`. RV1106 is 32-bit `arch/arm`: excluded before the register question. `of_match` is hardcoded `"rockchip,rk3588-rknn-core"`. |
|
||||
| Mesa **Teflon** (TFLite delegate) | Only etnaviv or rocket backends | **No.** Entirely downstream of Rocket, its Rockchip reach _is_ Rocket's reach: RK3588 only. |
|
||||
| `gahingwoo/linux-rk3576-npu` + `charsiu` (active, days-old as of 2026-08) | RK3576 | **No, wrong generation.** MobileNetV1 end-to-end + a Llama-3.2-1B INT4 open runtime (`charsiu`, opens `/dev/accel/accel0`, emits its own register stream) reportedly work byte-exact, but on RK3576's NPU, a **newer generation** than RV1106's. Unmerged (LKML "[PATCH v9] accel/rocket: RK3576" in review; unverifiable by second source, Anubis-walled). |
|
||||
| RK3568/RK3566 Armbian fork | RK3568 | **No.** Built by byte-level diff against captured vendor command streams; multi-task still imperfect; unmerged. |
|
||||
| `phhusson/rknpu-reverse-engineering` | RK3588 only, dormant since 2024-03, no license | **No.** Exploratory `strings`/`strace`/GEM-dump notes; **no compiler, no encoder, no runtime.** ~1 person-year got RK3588 to "structures identified." |
|
||||
| **RV1106/RV1103** | — | **Zero hits of any kind.** No driver, no fork, no RE writeup, no captured traces. Least-covered tier in the family. |
|
||||
| **RV1106/RV1103** | n/a | **Zero hits of any kind.** No driver, no fork, no RE writeup, no captured traces. Least-covered tier in the family. |
|
||||
|
||||
RV1106 sits in the RKNPU2 **"v1" family** (RK3566/68, RV1109/26, RV1103/06 —
|
||||
RV1106 sits in the RKNPU2 **"v1" family** (RK3566/68, RV1109/26, RV1103/06,
|
||||
INT8-only, `quantize=8` mandatory). RK3576/RK3588 are a **different, newer**
|
||||
generation. Precedent (RK3568 needing byte-level RE off RK3588) shows this is a
|
||||
**bespoke per-SoC RE project even between adjacent generations, not a recompile.**
|
||||
@@ -112,36 +112,36 @@ generation. Precedent (RK3568 needing byte-level RE off RK3588) shows this is a
|
||||
|
||||
Ranked by realism:
|
||||
|
||||
- **Tier A — ACHIEVABLE NOW (commit to this):** open GPL kernel driver, statically
|
||||
- **Tier A, ACHIEVABLE NOW (commit to this):** open GPL kernel driver, statically
|
||||
built into our 6.18.46, `/dev/dri/cardN` binds, `/proc/rknpu/load` live for the
|
||||
Monitor page, `RKNPU_GET_HW_VERSION`/`RKNPU_GET_DRV_VERSION` verified on
|
||||
hardware. **No blob shipped.** This is item (1); it is the honest "100% open" for
|
||||
the _driver_.
|
||||
- **Tier B — ACHIEVABLE AS A RESEARCH SPIKE, bounded, low-value:** a **single
|
||||
- **Tier B, ACHIEVABLE AS A RESEARCH SPIKE, bounded, low-value:** a **single
|
||||
hand-written op** (one conv or one matmul) executed via a `regcmd` **captured**
|
||||
from the closed stack and **replayed** through our open `DRM_IOCTL_RKNPU_SUBMIT`,
|
||||
byte-for-byte, with an open host tool — proving the open submit path can drive
|
||||
byte-for-byte, with an open host tool: proving the open submit path can drive
|
||||
real compute. This does **not** require understanding the regcmd ISA; it requires
|
||||
capturing one. It is the concrete first milestone toward open compute (§3, M-NPU-2)
|
||||
capturing one. It is the concrete first milestone toward open compute (section 3, M-NPU-2)
|
||||
and the only way to get an on-hardware compute number without the RE project.
|
||||
- **Tier C — NOT ACHIEVABLE on any near horizon:** an open **compiler/encoder** that
|
||||
- **Tier C, NOT ACHIEVABLE on any near horizon:** an open **compiler/encoder** that
|
||||
emits novel regcmd streams for arbitrary models. This is the ~person-year+
|
||||
from-scratch RE project (§3, M-NPU-3+). We should **scope it, not staff it**,
|
||||
from-scratch RE project (section 3, M-NPU-3+). We should **scope it, not staff it**,
|
||||
unless a strategic reason to own open NPU compute emerges.
|
||||
|
||||
**Honest recommendation:** ship Tier A. Do Tier B only as a time-boxed spike _if_ a
|
||||
real non-visual inference use case (audio/RS-485/touch-gesture classifier — the
|
||||
real non-visual inference use case (audio/RS-485/touch-gesture classifier, the
|
||||
only sensible NPU use here; no camera exists on this board) is actually wanted.
|
||||
Treat Tier C as closed-vendor-runtime-only for the foreseeable future. This matches
|
||||
the standing `PROVENANCE.md` policy: **we do not ship the closed runtime blob**;
|
||||
the open path is a from-scratch regcmd encoder tracked separately, never a vendored
|
||||
binary — and until that encoder exists, open NPU _compute_ does not exist for us.
|
||||
binary, and until that encoder exists, open NPU _compute_ does not exist for us.
|
||||
|
||||
---
|
||||
|
||||
## 2. Key technical facts (for the implementer)
|
||||
|
||||
### Register / memory map (TRM Part1 — Part2 does not exist publicly)
|
||||
### Register / memory map (TRM Part1: Part2 does not exist publicly)
|
||||
|
||||
| Block | Base | Size |
|
||||
|---|---|---|
|
||||
@@ -154,13 +154,13 @@ binary — and until that encoder exists, open NPU _compute_ does not exist for
|
||||
IRQ `GIC_SPI 109 LEVEL_HIGH`, node `ff660000.npu`. Clocks `ACLK_RKNN`/`HCLK_RKNN`
|
||||
off the CRU matrix; only measured clock point anywhere is **NPU 500MHz** (Rockchip
|
||||
internal power-test xlsx, "typical IPC workload" corner). A commented-out board-DTS
|
||||
override would set 700MHz — never applied.
|
||||
override would set 700MHz, never applied.
|
||||
|
||||
### UAPI (the seam an open userspace must target)
|
||||
|
||||
Two memory-manager paths, Kconfig-selected. **RV1106's shipping defconfig picks
|
||||
`CONFIG_ROCKCHIP_RKNPU_DMA_HEAP=y`** (`/dev/rknpu` misc device, bare `IOCTL_RKNPU_*`
|
||||
codes, magic `'r'`) — but `PORT-PLAN.md` deliberately chooses **DRM_GEM** for the
|
||||
codes, magic `'r'`), but `PORT-PLAN.md` deliberately chooses **DRM_GEM** for the
|
||||
6.18 port (`/dev/dri/cardN`, `DRM_IOCTL_RKNPU_*`, `DRM_RENDER_ALLOW`), because M4
|
||||
already pulls DRM in for the panel and DRM_GEM is the vendor default. Same struct
|
||||
payloads either way. Six ioctls: `RKNPU_ACTION`, `RKNPU_SUBMIT`,
|
||||
@@ -182,16 +182,16 @@ struct rknpu_submit { // rknpu_ioctl.h:260-274
|
||||
};
|
||||
```
|
||||
|
||||
An open userspace must produce: **(a) the `struct rknpu_task[]` list — fully
|
||||
An open userspace must produce: **(a) the `struct rknpu_task[]` list, fully
|
||||
inferable from the kernel driver alone, trivial;** and **(b) the regcmd byte stream
|
||||
each task points to — NOT inferable from the driver or any published TRM.** (b) is
|
||||
each task points to, NOT inferable from the driver or any published TRM.** (b) is
|
||||
the whole problem.
|
||||
|
||||
### `/proc/rknpu` — safety (hardware-confirmed)
|
||||
### `/proc/rknpu`: safety (hardware-confirmed)
|
||||
|
||||
- `/proc/rknpu/load` — **safe** (50 clean reads verified); the Monitor page polls it.
|
||||
- `/proc/rknpu/freq`, `/power` — exposed, unverified.
|
||||
- `/proc/rknpu/volt` — **CONFIRMED SIGSEGV** (null deref — no regulator wired). Never poll.
|
||||
- `/proc/rknpu/load`: **safe** (50 clean reads verified); the Monitor page polls it.
|
||||
- `/proc/rknpu/freq`, `/power`: exposed, unverified.
|
||||
- `/proc/rknpu/volt`: **CONFIRMED SIGSEGV** (null deref, no regulator wired). Never poll.
|
||||
|
||||
### Provenance / policy
|
||||
|
||||
@@ -215,26 +215,26 @@ open encoder exists.
|
||||
|
||||
## 3. Milestones
|
||||
|
||||
### M-NPU-1 — Open kernel driver, verified on hardware (ACHIEVABLE — the real deliverable)
|
||||
### M-NPU-1: Open kernel driver, verified on hardware (ACHIEVABLE, the real deliverable)
|
||||
|
||||
Follows `PORT-PLAN.md` in full. Summary of the gates:
|
||||
|
||||
1. **Build**: `CONFIG_ROCKCHIP_RKNPU=y` + `_DRM_GEM=y` + `_DEBUG_FS`/`_PROC_FS`;
|
||||
compile `rknpu_{drv,job,gem,reset,iommu,debugger}.o` clean into `built-in.a`.
|
||||
Prove the four §2.4 compat shims (`compat/soc/rockchip/rockchip_{iommu,opp_select,
|
||||
system_monitor,ipa}.h`) against real 6.18 headers — this is where they get
|
||||
Prove the four section 2.4 compat shims (`compat/soc/rockchip/rockchip_{iommu,opp_select,
|
||||
system_monitor,ipa}.h`) against real 6.18 headers: this is where they get
|
||||
verified, not just inspected.
|
||||
2. **DT**: board override `&npu { status="okay"; }`; `dtc -W` clean; re-check the
|
||||
"place the fdt high" DTB-placement fix (bigger `built-in.a`).
|
||||
3. **Boot** via the `warden-c8a3` `_b`-slot one-shot loop (never touches `_a`,
|
||||
auto-reverts on hang): `dmesg | grep -i rknpu` shows clean probe — clock/reset/
|
||||
auto-reverts on hang): `dmesg | grep -i rknpu` shows clean probe, clock/reset/
|
||||
IRQ acquired, no panic, no `-EPROBE_DEFER` stall.
|
||||
4. **Node**: `/dev/dri/cardN` (or `renderD1xx`) appears — classic DRM node, **not**
|
||||
4. **Node**: `/dev/dri/cardN` (or `renderD1xx`) appears, classic DRM node, **not**
|
||||
`/dev/accel/`.
|
||||
5. **Verify ioctl**: host-buildable C program opens the node, issues
|
||||
`DRM_IOCTL_RKNPU_ACTION {.flags=RKNPU_GET_DRV_VERSION}`, checks `value` decodes
|
||||
to `0.9.2`; `RKNPU_GET_HW_VERSION` returns something plausible. Exercises full
|
||||
dispatch → power-get/put → clock/reset with no regcmd dependency.
|
||||
dispatch -> power-get/put -> clock/reset with no regcmd dependency.
|
||||
6. **Capture the durable delta** as a patch series in this dir (compat shims + DT
|
||||
fragment + config fragment); vendor source stays in `research/linux-6.18.46/`.
|
||||
|
||||
@@ -242,32 +242,32 @@ Follows `PORT-PLAN.md` in full. Summary of the gates:
|
||||
not surfaced by the read-through; DTB placement. **This is "NPU driver = 100% open,
|
||||
verified."** Done here for the driver half.
|
||||
|
||||
### M-NPU-2 — First open compute: capture-and-replay one op (SPIKE, only if wanted)
|
||||
### M-NPU-2, First open compute: capture-and-replay one op (SPIKE, only if wanted)
|
||||
|
||||
The concrete first milestone toward open _compute_, and the honest Tier-B ceiling.
|
||||
It does **not** require decoding the ISA.
|
||||
|
||||
1. On a unit with the closed stack available (build host / a dev panel with the
|
||||
closed `librknnrt` + a trivial single-conv or single-matmul `.rknn`), capture
|
||||
the regcmd buffer(s) the runtime DMAs in — via `strace` of the `RKNPU_MEM_*`/
|
||||
the regcmd buffer(s) the runtime DMAs in: via `strace` of the `RKNPU_MEM_*`/
|
||||
`RKNPU_SUBMIT` ioctls + `GEM_FLINK`/`GEM_OPEN` buffer dumps (phhusson's method),
|
||||
plus the input/output tensor buffers.
|
||||
2. Write an **open host tool** that: allocates the same GEM buffers via
|
||||
`RKNPU_MEM_CREATE`, writes the captured regcmd bytes + captured input tensor,
|
||||
builds the `struct rknpu_task[]` (fully open, §2), and submits via
|
||||
builds the `struct rknpu_task[]` (fully open, section 2), and submits via
|
||||
`DRM_IOCTL_RKNPU_SUBMIT` on our M-NPU-1 driver.
|
||||
3. **Verify**: output tensor matches the closed runtime's output **byte-for-byte**
|
||||
on hardware, and matches a CPU/NEON reference for the same op. Record the
|
||||
**measured job-submission latency** (ioctl+DMA+blocking-IRQ) — a number that
|
||||
**measured job-submission latency** (ioctl+DMA+blocking-IRQ), a number that
|
||||
exists nowhere today and settles the "NPU vs NEON dispatch overhead" argument.
|
||||
|
||||
**Deliverable:** proof the open submit path drives real compute + first latency
|
||||
number. **Not** a general runtime — the regcmd is a fixed captured blob for one op
|
||||
number. **Not** a general runtime: the regcmd is a fixed captured blob for one op
|
||||
shape. **Value:** low; do only as a time-boxed research spike behind a real
|
||||
use case. **Never ship a captured Rockchip regcmd as a product artifact** (it is
|
||||
their compiler's output) — this stays in `research/`.
|
||||
their compiler's output), this stays in `research/`.
|
||||
|
||||
### M-NPU-3+ — Open regcmd encoder (SCOPE ONLY — do not staff without a strategic reason)
|
||||
### M-NPU-3+: Open regcmd encoder (SCOPE ONLY, do not staff without a strategic reason)
|
||||
|
||||
The Tier-C from-scratch RE project. Method (the only demonstrated one):
|
||||
differentially decode captured regcmd streams across many hand-built minimal ONNX
|
||||
@@ -286,14 +286,14 @@ ever begun; otherwise this is the documented reason open NPU compute is deferred
|
||||
jobs; it delivers **no open way to produce a valid job.** Guard this in every
|
||||
status line.
|
||||
- Additional vendor-only symbols beyond the four headers may surface only at
|
||||
compile — M-NPU-1 gate 1 is the real test.
|
||||
compile: M-NPU-1 gate 1 is the real test.
|
||||
- **SKU unconfirmed** (G2 128MB / 0.5-TOPS vs G3 256MB / 1.0-TOPS; datasheet Rev 2.0
|
||||
retracted the split to "both 1.0 TOPS"). Determines the compute ceiling if
|
||||
M-NPU-2/3 are ever pursued; needs an on-hardware chip-ID/TOPS register read.
|
||||
- **No open regcmd project for RV1106 exists** (re-confirm periodically — community
|
||||
- **No open regcmd project for RV1106 exists** (re-confirm periodically, community
|
||||
moves fast, esp. the `gahingwoo` RK3576 work; but it is the wrong generation).
|
||||
- **No NPU use case is currently established for this board** (no camera; RGA2 +
|
||||
RKIVE already cover the 2D/CNN-shaped image tasks better — see
|
||||
RKIVE already cover the 2D/CNN-shaped image tasks better, see
|
||||
`npu-graphics-feasibility.md`). Absent a concrete non-visual classifier need,
|
||||
M-NPU-2 and beyond have no pull, and M-NPU-1 alone (driver binds, load graph
|
||||
works, no blob) is the correct stopping point.
|
||||
@@ -301,4 +301,4 @@ ever begun; otherwise this is the documented reason open NPU compute is deferred
|
||||
---
|
||||
_Cross-refs: `PORT-PLAN.md` (authoritative file-by-file kernel port),
|
||||
`../../docs/npu-graphics-feasibility.md`, `../CAPABILITIES-AUDIT.md`,
|
||||
`../PROVENANCE.md`, `../DRIVER-PARITY.md:41`, `../REMAINING-PORTS.md §6`._
|
||||
`../PROVENANCE.md`, `../DRIVER-PARITY.md:41`, `../REMAINING-PORTS.md section 6`._
|
||||
|
||||
@@ -1,29 +1,29 @@
|
||||
# RKNPU kernel driver → 6.18 — port plan (M6 class)
|
||||
# RKNPU kernel driver -> 6.18: port plan (M6 class)
|
||||
|
||||
Scope: port the **kernel driver only** (`rknpu.ko`'s source, statically built into
|
||||
our tree) so `/dev/dri/cardN` (or `renderD1xx`) binds on the RV1106 NPU and answers
|
||||
a version-query ioctl. This does **not** run a model — see §3 for why that's a
|
||||
a version-query ioctl. This does **not** run a model. See section 3 for why that's a
|
||||
separate, much bigger, and largely closed problem. Written against the same
|
||||
target as the rest of this port: **Linux 6.18.46 vanilla**
|
||||
(`flare-edge/research/linux-6.18.46/`), forward-ported from vendor 5.10.160, built
|
||||
with our `arm-rockchip830-...-gcc 8.3` toolchain — see `../PORT-STATUS.md` and
|
||||
with our `arm-rockchip830-...-gcc 8.3` toolchain. See `../PORT-STATUS.md` and
|
||||
`../../docs/bringup.md` for the method and milestones this slots into (M6, listed
|
||||
in `../DRIVER-PARITY.md` as "NPU (rknpu, ff660000) | out-of-tree | [ ] M6").
|
||||
|
||||
Builds on `warden-sdk/docs/npu-graphics-feasibility.md`, which already read this
|
||||
same driver source to answer a narrower question (can the NPU do graphics — no).
|
||||
same driver source to answer a narrower question (can the NPU do graphics: no).
|
||||
This document answers the porting question that doc explicitly deferred.
|
||||
|
||||
---
|
||||
|
||||
## 1. What we're forward-porting (vendor SDK source)
|
||||
|
||||
Source: `flare-edge/sdk/sysdrv/source/kernel/drivers/rknpu/` — vendor version
|
||||
Source: `flare-edge/sdk/sysdrv/source/kernel/drivers/rknpu/`, vendor version
|
||||
**0.9.2** (`DRIVER_MAJOR/MINOR/PATCHLEVEL` in `include/rknpu_drv.h:34-36`,
|
||||
`DRIVER_DATE "20230825"`), currently loaded on the shipping 5.10.160 kernel as
|
||||
`rknpu.ko`, IRQ `ff660000.npu`.
|
||||
|
||||
### The driver already targets RV1106 natively — this is not a from-scratch adaptation
|
||||
### The driver already targets RV1106 natively: this is not a from-scratch adaptation
|
||||
|
||||
The vendor driver is a **single multi-SoC codebase**, not something written for
|
||||
RK3588 that we'd have to adapt. `rknpu_drv.c`'s `of_match` table already carries a
|
||||
@@ -49,7 +49,7 @@ static const struct rknpu_config rv1106_rknpu_config = {
|
||||
```
|
||||
|
||||
And the devicetree side is **already complete** in `rv1106.dtsi` (base tree,
|
||||
`status = "disabled"`) — nothing to author, only enable:
|
||||
`status = "disabled"`), nothing to author, only enable:
|
||||
|
||||
```dts
|
||||
// arch/arm/boot/dts/rv1106.dtsi:1127-1138
|
||||
@@ -67,17 +67,17 @@ npu: npu@ff660000 {
|
||||
};
|
||||
```
|
||||
|
||||
No `power-domains` property (RV1106's NPU is single-core, single-rail — unlike
|
||||
No `power-domains` property (RV1106's NPU is single-core, single-rail, unlike
|
||||
RK3588's 3-core NPU which needs `genpd_dev_npu0/1/2`) and no `iommus` property
|
||||
(matches the 5.10 boot-log finding already on record: `"rknpu iommu device-tree
|
||||
entry not found!, using non-iommu mode"`, `npu-graphics-feasibility.md:123-126`).
|
||||
Both facts materially shrink the v1 port's scope — see §2.3 and §2.4.
|
||||
Both facts materially shrink the v1 port's scope. See section 2.3 and section 2.4.
|
||||
|
||||
### The uapi/job model (for context — already characterized in the feasibility doc)
|
||||
### The uapi/job model (for context, already characterized in the feasibility doc)
|
||||
|
||||
`include/rknpu_ioctl.h`: `struct rknpu_task` (`regcfg_amount`, `regcfg_offset`,
|
||||
`regcmd_addr`) + `struct rknpu_submit` (`task_obj_addr`, `regcfg_obj_addr`,
|
||||
`core_mask`, `fence_fd`) — a **register command-list (regcmd) task-queue model**,
|
||||
`core_mask`, `fence_fd`), a **register command-list (regcmd) task-queue model**,
|
||||
submitted via `DRM_IOCTL_RKNPU_SUBMIT` and executed to completion with the driver
|
||||
blocking on a hardware IRQ (`rknpu_job.c`, `wait_event_timeout`). Six ioctls total:
|
||||
`RKNPU_ACTION`, `RKNPU_SUBMIT`, `RKNPU_MEM_{CREATE,MAP,DESTROY,SYNC}`
|
||||
@@ -88,19 +88,19 @@ verify step uses: `RKNPU_GET_HW_VERSION = 0`, `RKNPU_GET_DRV_VERSION = 1`
|
||||
### Memory manager choice: DRM GEM (not DMA-heap)
|
||||
|
||||
`Kconfig` offers a mutually-exclusive choice: `ROCKCHIP_RKNPU_DRM_GEM` (default)
|
||||
vs `ROCKCHIP_RKNPU_DMA_HEAP`. **Pick DRM_GEM** — M4 (VOP2 display) already pulls
|
||||
vs `ROCKCHIP_RKNPU_DMA_HEAP`. **Pick DRM_GEM**: M4 (VOP2 display) already pulls
|
||||
the DRM core into this kernel for the panel, so there's no new subsystem cost, and
|
||||
DRM_GEM is the vendor's default/most-tested path. This also means the node that
|
||||
appears is a classic DRM char device (`/dev/dri/cardN` / `renderD1xx` via
|
||||
`drm_dev_alloc`/`drm_dev_register`, `rknpu_drv.c:725,730`) — **not** the newer
|
||||
`/dev/accel/` framework mainline's own `rocket` driver uses (§3). Don't confuse
|
||||
`drm_dev_alloc`/`drm_dev_register`, `rknpu_drv.c:725,730`), **not** the newer
|
||||
`/dev/accel/` framework mainline's own `rocket` driver uses (section 3). Don't confuse
|
||||
the two node namespaces when verifying.
|
||||
|
||||
---
|
||||
|
||||
## 2. Files, config, and the concrete 5.10→6.18 API deltas
|
||||
## 2. Files, config, and the concrete 5.10->6.18 API deltas
|
||||
|
||||
### 2.1 File list — v1 minimal port
|
||||
### 2.1 File list: v1 minimal port
|
||||
|
||||
| File | Bring in? | Why |
|
||||
|---|---|---|
|
||||
@@ -108,15 +108,15 @@ the two node namespaces when verifying.
|
||||
| `rknpu_job.c` / `include/rknpu_job.h` | **Yes** | job submit, IRQ handler, PC task-list execution |
|
||||
| `rknpu_gem.c` / `include/rknpu_gem.h` | **Yes** | GEM memory manager (DRM_GEM path) |
|
||||
| `rknpu_reset.c` / `include/rknpu_reset.h` | **Yes** | `SRST_A_RKNN`/`SRST_H_RKNN` reset control |
|
||||
| `rknpu_iommu.c` / `include/rknpu_iommu.h` | Yes, but dead code path | 61 lines, self-contained, already version-gated to 6.1; harmless to carry even though `iommu_en` stays false on our non-IOMMU DT (§2.4) |
|
||||
| `rknpu_debugger.c` / `include/rknpu_debugger.h` | Yes (optional) | `/proc/rknpu/load` — the Monitor page already polls this on 5.10 (`npu-graphics-feasibility.md:151-158`); keep for continuity even though wiring the UI back up is out of scope here |
|
||||
| `rknpu_mem.c` | **No** | only for `ROCKCHIP_RKNPU_DMA_HEAP` — we're not using that memory manager |
|
||||
| `rknpu_mm.c` / `include/rknpu_mm.h` | **No** | SRAM/NBUF allocator (`ROCKCHIP_RKNPU_SRAM`, needs `NO_GKI`); `rv1106_rknpu_config` has `nbuf_phyaddr=0, nbuf_size=0` — dead weight on this SoC |
|
||||
| `rknpu_fence.c` / `include/rknpu_fence.h` | **No (v1)** | `ROCKCHIP_RKNPU_FENCE`/`SYNC_FILE` — dma-fence cross-driver sync, not needed to prove basic binding; revisit if a real workload needs fenced submission later |
|
||||
| `Kconfig`, `Makefile` | **Yes, trimmed** | drop the `rknpu_mem.o`/`rknpu_mm.o`/`rknpu_fence.o` conditional lines' configs (leave the `Makefile` structure as-is — it's already `obj-$(CONFIG_...)`-gated per file, so simply not enabling those Kconfig symbols is sufficient; no Makefile edit required) |
|
||||
| `rknpu_iommu.c` / `include/rknpu_iommu.h` | Yes, but dead code path | 61 lines, self-contained, already version-gated to 6.1; harmless to carry even though `iommu_en` stays false on our non-IOMMU DT (section 2.4) |
|
||||
| `rknpu_debugger.c` / `include/rknpu_debugger.h` | Yes (optional) | `/proc/rknpu/load`: the Monitor page already polls this on 5.10 (`npu-graphics-feasibility.md:151-158`); keep for continuity even though wiring the UI back up is out of scope here |
|
||||
| `rknpu_mem.c` | **No** | only for `ROCKCHIP_RKNPU_DMA_HEAP`: we're not using that memory manager |
|
||||
| `rknpu_mm.c` / `include/rknpu_mm.h` | **No** | SRAM/NBUF allocator (`ROCKCHIP_RKNPU_SRAM`, needs `NO_GKI`); `rv1106_rknpu_config` has `nbuf_phyaddr=0, nbuf_size=0`, dead weight on this SoC |
|
||||
| `rknpu_fence.c` / `include/rknpu_fence.h` | **No (v1)** | `ROCKCHIP_RKNPU_FENCE`/`SYNC_FILE`: dma-fence cross-driver sync, not needed to prove basic binding; revisit if a real workload needs fenced submission later |
|
||||
| `Kconfig`, `Makefile` | **Yes, trimmed** | drop the `rknpu_mem.o`/`rknpu_mm.o`/`rknpu_fence.o` conditional lines' configs (leave the `Makefile` structure as-is; it's already `obj-$(CONFIG_...)`-gated per file, so simply not enabling those Kconfig symbols is sufficient; no Makefile edit required) |
|
||||
|
||||
### 2.2 Config symbols (built-in, matching this port's established pattern —
|
||||
M1–M3 build everything statically to avoid the vermagic class of bug that
|
||||
### 2.2 Config symbols (built-in, matching this port's established pattern:
|
||||
M1-M3 build everything statically to avoid the vermagic class of bug that
|
||||
blocks `aic8800.ko` today, `PORT-STATUS.md:114-115`)
|
||||
|
||||
```
|
||||
@@ -131,7 +131,7 @@ CONFIG_DRM=y # already required by M4 (VOP2)
|
||||
### 2.3 DT change
|
||||
|
||||
One-line status flip on the board DT (the same pattern M2/M3 used for
|
||||
`grf-clock-controller` and the eMMC node — override in the board file, don't
|
||||
`grf-clock-controller` and the eMMC node: override in the board file, don't
|
||||
touch the base `rv1106.dtsi`):
|
||||
|
||||
```dts
|
||||
@@ -140,19 +140,19 @@ touch the base `rv1106.dtsi`):
|
||||
};
|
||||
```
|
||||
|
||||
No new properties needed — `compatible`/`reg`/`interrupts`/`clocks`/`resets` are
|
||||
already correct and match the driver's own `rv1106_rknpu_config` exactly (§1).
|
||||
**Do not add an `iommus =` property for v1** — see §2.4.
|
||||
No new properties needed: `compatible`/`reg`/`interrupts`/`clocks`/`resets` are
|
||||
already correct and match the driver's own `rv1106_rknpu_config` exactly (section 1).
|
||||
**Do not add an `iommus =` property for v1**. See section 2.4.
|
||||
|
||||
### 2.4 The real build blocker: four vendor-only `soc/rockchip/*.h` headers
|
||||
|
||||
This is the one item in this plan that isn't "already handled by the vendor's own
|
||||
version gates" — verified directly against our `flare-edge/research/linux-6.18.46/`
|
||||
version gates", verified directly against our `flare-edge/research/linux-6.18.46/`
|
||||
tree, not assumed:
|
||||
|
||||
```c
|
||||
// rknpu_drv.c:37-40 (inside #ifndef FPGA_PLATFORM, which is never defined for
|
||||
// our build — grep of Makefile/Kconfig shows no FPGA_PLATFORM define anywhere)
|
||||
// our build: grep of Makefile/Kconfig shows no FPGA_PLATFORM define anywhere)
|
||||
#include <soc/rockchip/rockchip_iommu.h>
|
||||
#include <soc/rockchip/rockchip_opp_select.h>
|
||||
#include <soc/rockchip/rockchip_system_monitor.h>
|
||||
@@ -160,33 +160,33 @@ tree, not assumed:
|
||||
```
|
||||
|
||||
`find flare-edge/research/linux-6.18.46/include -iname 'rockchip_{iommu,opp_select,
|
||||
system_monitor,ipa}.h'` returns **nothing** — all four are Rockchip downstream-BSP
|
||||
system_monitor,ipa}.h'` returns **nothing**. All four are Rockchip downstream-BSP
|
||||
convenience headers (DVFS/OPP-table selection, thermal/system-monitor
|
||||
registration, IPA power-model, and a vendor wrapper around the IOMMU-core API)
|
||||
that were never upstreamed. `rknpu_drv.h:21-23` pulls in `rockchip_opp_select.h`
|
||||
unconditionally too, gated only by `KERNEL_VERSION(5,10,0) <= LINUX_VERSION_CODE`
|
||||
— true for 6.18, so it's compiled by default, not something a Kconfig toggle
|
||||
unconditionally too, gated only by `KERNEL_VERSION(5,10,0) <= LINUX_VERSION_CODE`,
|
||||
true for 6.18, so it's compiled by default, not something a Kconfig toggle
|
||||
avoids.
|
||||
|
||||
**All four are genuinely dead code for RV1106 at runtime**, which is what makes
|
||||
this a small, well-scoped fix rather than a real feature to build:
|
||||
- `rockchip_iommu_is_enabled()` (`rknpu_drv.c:902`) is only called inside
|
||||
`if (rknpu_dev->multiple_domains)` — true only for RK3588's 3-core NPU; RV1106
|
||||
never sets it (no `power-domains` property, §1).
|
||||
`if (rknpu_dev->multiple_domains)`: true only for RK3588's 3-core NPU; RV1106
|
||||
never sets it (no `power-domains` property, section 1).
|
||||
- The OPP/system-monitor/IPA calls drive dynamic frequency/voltage scaling and
|
||||
thermal cooling-device registration against an OPP table — RV1106's DT pins a
|
||||
thermal cooling-device registration against an OPP table. RV1106's DT pins a
|
||||
single fixed clock rate (`assigned-clock-rates = <420000000>`) and has no
|
||||
`operating-points-v2` table; none of this is exercised today either.
|
||||
|
||||
**Fix**: add small local compat shim headers (in this port's own include path,
|
||||
ahead of the vendor source's include search path) providing just the symbols
|
||||
these call sites reference — `rockchip_iommu_is_enabled()` returning `false`, and
|
||||
these call sites reference, `rockchip_iommu_is_enabled()` returning `false`, and
|
||||
no-op/`-ENOTSUPP` stand-ins for the opp/monitor/ipa registration calls actually
|
||||
referenced in `rknpu_drv.c`. This is the same "compat shim for a header that moved
|
||||
or doesn't exist upstream" pattern already used for `clk-rv1106.c`'s
|
||||
`panic_notifier_list` move (`../PORT-STATUS.md:18-20`) — same class of fix,
|
||||
`panic_notifier_list` move (`../PORT-STATUS.md:18-20`), same class of fix,
|
||||
same low risk, because the code behind it is provably dead for this SoC's DT.
|
||||
**Do not** reach for `#define FPGA_PLATFORM` as a shortcut — that macro also
|
||||
**Do not** reach for `#define FPGA_PLATFORM` as a shortcut: that macro also
|
||||
guards the reset-control logic in `rknpu_reset.c` (nearly the whole file is
|
||||
`#ifndef FPGA_PLATFORM`), which we need live; it's too blunt an instrument here.
|
||||
|
||||
@@ -199,73 +199,73 @@ actually still true at 6.18:
|
||||
|
||||
| Vendor gate (`rknpu_drv.c`) | 6.18 status (verified) |
|
||||
|---|---|
|
||||
| `#if KERNEL_VERSION(6,1,0) > LINUX_VERSION_CODE` around `.gem_free_object_unlocked` | Correctly **skipped** — that field is gone from `struct drm_driver` in 6.18's `include/drm/drm_drv.h` (grepped, zero hits), and the driver's `#else` branch already uses the modern `struct drm_gem_object_funcs` (`.free`, `.export`, `.get_sg_table`, `.vmap`, `.vunmap`, `.mmap` — `rknpu_gem.c:352-358`) |
|
||||
| `#if KERNEL_VERSION(6,1,0) > LINUX_VERSION_CODE` around `.gem_free_object_unlocked` | Correctly **skipped**: that field is gone from `struct drm_driver` in 6.18's `include/drm/drm_drv.h` (grepped, zero hits), and the driver's `#else` branch already uses the modern `struct drm_gem_object_funcs` (`.free`, `.export`, `.get_sg_table`, `.vmap`, `.vunmap`, `.mmap`, `rknpu_gem.c:352-358`) |
|
||||
| `DEFINE_DRM_GEM_FOPS(...)` (6.1+) vs hand-rolled `file_operations` | 6.1+ branch applies; macro is a standard DRM-core helper, present in 6.18 |
|
||||
| `.gem_prime_mmap = drm_gem_prime_mmap` (6.1+) vs a custom `rknpu_gem_prime_mmap` | 6.1+ branch applies |
|
||||
| `struct drm_driver` fields the vendor initializer sets (`major`, `minor`, `patchlevel`, `driver_features`, `dumb_create`, `dumb_map_offset`) | All still present in 6.18's `drm_drv.h` (line-checked) |
|
||||
| `DRM_IOCTL_DEF_DRV(...)` macro (ioctl table) | Still defined in 6.18's `include/drm/drm_ioctl.h:151` |
|
||||
| `iommu_map()` / `iommu_unmap()` / `iommu_get_domain_for_dev()` / `iommu_attach_device()` / `iommu_detach_device()` (`rknpu_gem.c`, `rknpu_reset.c`) | `iommu_map()`'s extern signature is unchanged in 6.18's `include/linux/iommu.h:914` (mainline did add a newer `iommu_map_nosync()` alongside it, but didn't remove the classic call) — moot anyway since this path is dead on our non-IOMMU DT (§2.4) |
|
||||
| `devm_reset_control_get`, `clk_bulk_data`, `pm_runtime_get_sync`/`put_sync`/`resume_and_get` | Stable mainline APIs across the whole 5.10→6.18 span; no gate needed |
|
||||
| `iommu_map()` / `iommu_unmap()` / `iommu_get_domain_for_dev()` / `iommu_attach_device()` / `iommu_detach_device()` (`rknpu_gem.c`, `rknpu_reset.c`) | `iommu_map()`'s extern signature is unchanged in 6.18's `include/linux/iommu.h:914` (mainline did add a newer `iommu_map_nosync()` alongside it, but didn't remove the classic call), moot anyway since this path is dead on our non-IOMMU DT (section 2.4) |
|
||||
| `devm_reset_control_get`, `clk_bulk_data`, `pm_runtime_get_sync`/`put_sync`/`resume_and_get` | Stable mainline APIs across the whole 5.10->6.18 span; no gate needed |
|
||||
|
||||
Net: outside the four-header fix in §2.4, this is expected to be a **build-fix-build
|
||||
pass**, not a rewrite — confirm by actually compiling into the tree (the checks
|
||||
Net: outside the four-header fix in section 2.4, this is expected to be a **build-fix-build
|
||||
pass**, not a rewrite. Confirm by actually compiling into the tree (the checks
|
||||
above are header-presence/signature verification, not a build).
|
||||
|
||||
### 2.6 Explicitly deferred (not required to prove the driver binds)
|
||||
|
||||
- **IOMMU enablement.** Stays off — matches current 5.10 runtime behavior and
|
||||
- **IOMMU enablement.** Stays off: matches current 5.10 runtime behavior and
|
||||
avoids the newer-kernel IOMMU-core churn entirely (mainline replaced
|
||||
`iommu_domain_alloc(bus)` with device-based `iommu_paging_domain_alloc(dev)`
|
||||
somewhere in the 6.x series — confirmed by grepping 6.18.46's `iommu.h`, which
|
||||
somewhere in the 6.x series, confirmed by grepping 6.18.46's `iommu.h`, which
|
||||
has the new call and no bus-based `iommu_domain_alloc`). Since our DT carries no
|
||||
`iommus=` property, this churn never gets compiled against in the first place.
|
||||
- **dma-fence / `ROCKCHIP_RKNPU_FENCE`.** Cross-driver sync primitive, not needed
|
||||
to answer a version-query ioctl.
|
||||
- **DVFS / thermal cooling / multi-power-domain.** RV1106-inapplicable per §2.4;
|
||||
- **DVFS / thermal cooling / multi-power-domain.** RV1106-inapplicable per section 2.4;
|
||||
stubbed out, not implemented.
|
||||
- **SRAM/NBUF allocator.** Dead weight on this SoC's config table (§2.1).
|
||||
- **SRAM/NBUF allocator.** Dead weight on this SoC's config table (section 2.1).
|
||||
|
||||
---
|
||||
|
||||
## 3. Verify steps (driver binding only — no model, no RKNN runtime)
|
||||
## 3. Verify steps (driver binding only, no model, no RKNN runtime)
|
||||
|
||||
Uses the same proven safe-test loop as M2/M3: the A/B `_b`-slot one-shot boot on
|
||||
`warden-c8a3` (`../../docs/m2-boot-on-c8a3.md`) — never touches the working `_a`
|
||||
`warden-c8a3` (`../../docs/m2-boot-on-c8a3.md`), never touches the working `_a`
|
||||
slot, auto-reverts on hang.
|
||||
|
||||
1. **Build**: `CONFIG_ROCKCHIP_RKNPU=y` (+ the symbols in §2.2) added to the
|
||||
1. **Build**: `CONFIG_ROCKCHIP_RKNPU=y` (+ the symbols in section 2.2) added to the
|
||||
defconfig fragment; `rknpu_drv.o`/`rknpu_job.o`/`rknpu_gem.o`/`rknpu_reset.o`/
|
||||
`rknpu_iommu.o`/`rknpu_debugger.o` compile clean into `built-in.a` — this is
|
||||
where the §2.4 shim headers get proven, not just inspected.
|
||||
`rknpu_iommu.o`/`rknpu_debugger.o` compile clean into `built-in.a`; this is
|
||||
where the section 2.4 shim headers get proven, not just inspected.
|
||||
2. **DT**: `npu@ff660000` flipped to `okay`; `dtc -W` clean, no warnings, no
|
||||
overrun of the existing DTB-placement fix from M2 (`PORT-STATUS.md`'s "place
|
||||
the fdt high" note — a bigger built-in.a makes this worth re-checking).
|
||||
the fdt high" note; a bigger built-in.a makes this worth re-checking).
|
||||
3. **Boot** (via the `_b`-slot loop): `dmesg | grep -i rknpu` shows the
|
||||
`platform_driver` probing without error — clock/reset/IRQ acquired, no panic,
|
||||
`platform_driver` probing without error, clock/reset/IRQ acquired, no panic,
|
||||
no `-EPROBE_DEFER` stall. Compare against the 5.10 baseline probe log for the
|
||||
same node if available.
|
||||
4. **Node appears**: `ls -la /dev/dri/` shows a new `cardN`/`renderD1xx` for the
|
||||
npu — classic DRM char device (not `/dev/accel/`, see §1).
|
||||
npu, classic DRM char device (not `/dev/accel/`, see section 1).
|
||||
5. **Trivial ioctl, not a model**: a small host-buildable C program opens the DRM
|
||||
node and issues `DRM_IOCTL_RKNPU_ACTION` with `{.flags = RKNPU_GET_DRV_VERSION}`
|
||||
(`rknpu_ioctl.h:114`), checks the returned `value` decodes to `0.9.2`
|
||||
(`RKNPU_GET_DRV_VERSION_MAJOR/MINOR/PATCHLEVEL` macros,
|
||||
`rknpu_ioctl.h:52-54`) — and/or `RKNPU_GET_HW_VERSION` returns something
|
||||
plausible. This exercises the full ioctl-dispatch → power-get/put →
|
||||
`rknpu_ioctl.h:52-54`), and/or `RKNPU_GET_HW_VERSION` returns something
|
||||
plausible. This exercises the full ioctl-dispatch -> power-get/put ->
|
||||
clock/reset path with zero dependency on a regcmd buffer or the RKNN runtime.
|
||||
6. **Explicitly not required for "done" here**: `DRM_IOCTL_RKNPU_SUBMIT`, any
|
||||
`.rknn` model, `librknnrt`. That's the userspace question — §4.
|
||||
`.rknn` model, `librknnrt`. That's the userspace question: section 4.
|
||||
|
||||
### Effort estimate
|
||||
|
||||
Smaller than M1 (clk/pinctrl — required inferring an unknown CPU-clock mux from a
|
||||
Smaller than M1 (clk/pinctrl: required inferring an unknown CPU-clock mux from a
|
||||
sibling diff) and smaller than the display work ahead in M4 (register-map
|
||||
guesswork against RV1126/RK3568 siblings). This one is closer in shape to M3 ("the
|
||||
eMMC node was all M3 needed" — `PORT-STATUS.md:99`): the DT is already fully
|
||||
eMMC node was all M3 needed", `PORT-STATUS.md:99`): the DT is already fully
|
||||
specified upstream, the driver's C source already has a dedicated, tested RV1106
|
||||
config table and of_match entry (§1), and the GEM/DRM surface is already correctly
|
||||
version-gated past 6.1 (§2.5). The concentrated risk is (a) actually compiling
|
||||
the §2.4 shims against real 6.18 headers rather than trusting the header-presence
|
||||
config table and of_match entry (section 1), and the GEM/DRM surface is already correctly
|
||||
version-gated past 6.1 (section 2.5). The concentrated risk is (a) actually compiling
|
||||
the section 2.4 shims against real 6.18 headers rather than trusting the header-presence
|
||||
check above, and (b) the possibility of additional vendor-only symbols not
|
||||
surfaced by this read-through. Realistic order of magnitude: low-single-digit
|
||||
engineer-days to a clean probe + version-ioctl round trip on hardware, assuming
|
||||
@@ -273,29 +273,29 @@ no surprise blocks the way M2's boot-image format did.
|
||||
|
||||
---
|
||||
|
||||
## 4. The userspace-runtime reality: what porting the kernel driver does — and does NOT — unlock
|
||||
## 4. The userspace-runtime reality: what porting the kernel driver does (and does NOT) unlock
|
||||
|
||||
The instruction that motivated this document was explicit: don't let "the driver
|
||||
is open source" imply the NPU becomes open-source-usable. It doesn't. This section
|
||||
updates `npu-graphics-feasibility.md §4`'s conclusion with the current
|
||||
updates `npu-graphics-feasibility.md section 4`'s conclusion with the current
|
||||
(2026-08-24) state of every open effort found.
|
||||
|
||||
### 4.1 The kernel driver itself: genuinely open, and this is a real forward-port of Rockchip's own code
|
||||
|
||||
`rknpu_drv.c` et al. are SPDX `GPL-2.0`, authored by Rockchip
|
||||
(`Felix Zeng <felix.zeng@rock-chips.com>`), and are the actual vendor driver — the
|
||||
(`Felix Zeng <felix.zeng@rock-chips.com>`), and are the actual vendor driver: the
|
||||
same driver every Rockchip Linux SDK ships, mirrored at
|
||||
[`github.com/airockchip/rknpu`](https://github.com/airockchip/rknpu) /
|
||||
[`github.com/rockchip-linux/rknpu`](https://github.com/rockchip-linux/rknpu).
|
||||
Porting it forward is legitimate, license-clean work, not a workaround. **This is
|
||||
a different codebase from mainline's own driver** (§4.2) — don't conflate "port
|
||||
a different codebase from mainline's own driver** (section 4.2): don't conflate "port
|
||||
the vendor driver" with "adopt mainline's `accel/rocket`"; they are unrelated
|
||||
implementations of the same hardware class, and only one of them (the vendor one)
|
||||
covers RV1106 at all.
|
||||
|
||||
### 4.2 Mainline `accel/rocket`: real, merged, and does not reach RV1106
|
||||
|
||||
- Merged into mainline Linux and Mesa in **2025-07** — Tomeu Vizoso, ["Rockchip NPU
|
||||
- Merged into mainline Linux and Mesa in **2025-07**, Tomeu Vizoso, ["Rockchip NPU
|
||||
update 6: We are in mainline!"](https://blog.tomeuvizoso.net/2025/07/rockchip-npu-update-6-we-are-in-mainline.html),
|
||||
following the LKML series
|
||||
["[PATCH v2 0/7] New DRM accel driver for Rockchip's RKNN NPU"](https://lkml.iu.edu/hypermail/linux/kernel/2502.3/02497.html).
|
||||
@@ -304,33 +304,33 @@ covers RV1106 at all.
|
||||
question**: `depends on (ARCH_ROCKCHIP && ARM64) || COMPILE_TEST`
|
||||
(`drivers/accel/rocket/Kconfig` in our 6.18.46 tree). RV1106 is a 32-bit
|
||||
Cortex-A7 (`arch/arm`, confirmed on our own hardware:
|
||||
`PORT-STATUS.md`'s boot log — `CPU: ARMv7 Processor`). This is a Kconfig
|
||||
dependency, not necessarily an unfixable technical wall on its own — but it
|
||||
`PORT-STATUS.md`'s boot log, `CPU: ARMv7 Processor`). This is a Kconfig
|
||||
dependency, not necessarily an unfixable technical wall on its own, but it
|
||||
signals no one has done the 32-bit validation work, on top of the register-level
|
||||
work below.
|
||||
- **Hardware coverage, per the driver's own docs**
|
||||
(`Documentation/accel/rocket/index.rst` in our tree): *"Hardware currently
|
||||
supported: * RK3588."* Nothing else, as shipped in 6.18.46.
|
||||
- **RK3576** — active, but incomplete, and not in our tree. A 2026-07-15
|
||||
- **RK3576**: active, but incomplete, and not in our tree. A 2026-07-15
|
||||
reverse-engineering effort (
|
||||
[CNX Software](https://www.cnx-software.com/2026/07/15/reverse-engineering-brings-rk3576-npu-support-to-open-source-rocket-driver-for-mainline-linux/),
|
||||
code at [`gahingwoo/linux-rk3576-npu`](https://github.com/gahingwoo/linux-rk3576-npu))
|
||||
got single-task inference working end-to-end on a Radxa ROCK 4D running Linux
|
||||
7.1-rc5 — but **multi-task chained inference (any real multi-layer network)
|
||||
7.1-rc5, but **multi-task chained inference (any real multi-layer network)
|
||||
fails: only the first task per NPU power session actually computes.** Not
|
||||
merged into the kernel we're building against. Cited here because it's the
|
||||
closest active precedent to "port Rocket to a new RKNPU generation," and even
|
||||
that isn't production-usable yet.
|
||||
- **RK3568/RK3566** — an out-of-tree community fork exists (Armbian forum,
|
||||
- **RK3568/RK3566**: an out-of-tree community fork exists (Armbian forum,
|
||||
["ODROID-M1: RK3568 NPU on the open stack"](https://forum.armbian.com/topic/61651-odroid-m1-rk3568-npu-on-the-open-stack-rocket-kernel-driver-mesa-teflon/)),
|
||||
built on `accel/rocket` "with local fixes" atop the RK3588 Mesa merge request,
|
||||
requiring **byte-level comparison against captured vendor command streams** to
|
||||
work out weight-layout and CBUF differences from RK3588. Not merged upstream.
|
||||
Confirms the general pattern: **porting Rocket to a new RKNPU generation is a
|
||||
bespoke reverse-engineering project per SoC, not a recompile** — the same
|
||||
bespoke reverse-engineering project per SoC, not a recompile**, the same
|
||||
conclusion `npu-graphics-feasibility.md` already reached, now with two more
|
||||
data points (RK3576, RK3568) supporting it.
|
||||
- **RV1106/RV1103** — zero hits in this research. No mainline coverage, no known
|
||||
- **RV1106/RV1103**: zero hits in this research. No mainline coverage, no known
|
||||
public fork, no known RE project targeting it specifically (unlike RK3568 and
|
||||
RK3576, which both have named, in-progress efforts). This is the least-covered
|
||||
tier of the RKNPU family in the open-source world today.
|
||||
@@ -341,7 +341,7 @@ Merged into Mesa 24.1 ([Phoronix](https://www.phoronix.com/news/Gallium3D-Teflon
|
||||
[docs.mesa3d.org/teflon.html](https://docs.mesa3d.org/teflon.html)) as a Gallium3D
|
||||
frontend for TensorFlow Lite. Per its own docs: **"Teflon only works with etnaviv
|
||||
or rocket gallium drivers."** There is no Teflon path independent of a working
|
||||
Rocket kernel driver underneath it — so Teflon's real-world Rockchip coverage is
|
||||
Rocket kernel driver underneath it, so Teflon's real-world Rockchip coverage is
|
||||
exactly Rocket's: solid on RK3588, experimental/WIP on RK3576 and (unofficially)
|
||||
RK3568, absent for RV1106.
|
||||
|
||||
@@ -353,7 +353,7 @@ exploratory/documentation stage (structures like `regcfg_amount`/`regcmd_addr`
|
||||
identified, DRM device enumerated) with **no compiled tool, compiler, or runtime
|
||||
output**, and does not touch RV1106. This matches
|
||||
`npu-graphics-feasibility.md`'s existing finding that the regcmd ISA is
|
||||
undocumented outside Rockchip and reasoned-about only, not published — no new
|
||||
undocumented outside Rockchip and reasoned-about only, not published. No new
|
||||
project has changed that for any SoC generation, let alone this one.
|
||||
|
||||
### 4.5 Net conclusion (updated, still holds)
|
||||
@@ -366,8 +366,8 @@ RK3588 (has mainline Rocket + Teflon) or even RK3576/RK3568 (have active,
|
||||
imperfect RE efforts), **there is no open compiler, no open runtime, and no known
|
||||
public reverse-engineering project of any kind.** Every real inference workload on
|
||||
this NPU has to go through the closed pipeline
|
||||
(`npu-graphics-feasibility.md §1`: RKNN-Toolkit2 on a PC → `.rknn` blob →
|
||||
`librknnrt`/RKNN C API on-device) for the foreseeable future — porting the kernel
|
||||
(`npu-graphics-feasibility.md section 1`: RKNN-Toolkit2 on a PC -> `.rknn` blob ->
|
||||
`librknnrt`/RKNN C API on-device) for the foreseeable future; porting the kernel
|
||||
driver is worth doing (it's real, bounded, evidence-backed work, same class as
|
||||
RGA), but it does not change that reality, and shouldn't be scoped or sold as if
|
||||
it does.
|
||||
@@ -379,14 +379,14 @@ it does.
|
||||
```
|
||||
npu/
|
||||
PORT-PLAN.md this document
|
||||
compat/ (to add) local stub headers for §2.4:
|
||||
compat/ (to add) local stub headers for section 2.4:
|
||||
soc/rockchip/rockchip_iommu.h
|
||||
soc/rockchip/rockchip_opp_select.h
|
||||
soc/rockchip/rockchip_system_monitor.h
|
||||
soc/rockchip/rockchip_ipa.h
|
||||
```
|
||||
The vendor driver source itself is not duplicated here (same convention as
|
||||
`clk/`, `pinctrl/`, `mach/` — vendor source stays forward-ported in
|
||||
`flare-edge/research/linux-6.18.46/` as a scratch tree; only the durable delta —
|
||||
compat shims, DT fragment, config fragment — belongs in this repo, captured as a
|
||||
`clk/`, `pinctrl/`, `mach/`: vendor source stays forward-ported in
|
||||
`flare-edge/research/linux-6.18.46/` as a scratch tree; only the durable delta (
|
||||
compat shims, DT fragment, config fragment) belongs in this repo, captured as a
|
||||
patch series once M6 actually lands).
|
||||
|
||||
@@ -3,10 +3,10 @@
|
||||
Status: **zImage + rockchip/rv1106-warden.dtb build cleanly, 0 errors, 0 warnings**,
|
||||
with `CONFIG_ROCKCHIP_RKNPU=y` (+ `_DRM_GEM`/`_DEBUG_FS`/`_PROC_FS=y`) built in, 99
|
||||
`rknpu`-prefixed symbols linked into `System.map`, and `&npu { status = "okay"; }`
|
||||
in the board dtb. **Not flashed or probed on hardware** — per this task's explicit
|
||||
in the board dtb. **Not flashed or probed on hardware**, per this task's explicit
|
||||
boundary, this is a build-only port; on-target verification (dmesg probe log,
|
||||
`/dev/dri/cardN`, the version-query ioctl round trip) is deferred to the parent
|
||||
session. GPL source port only — no closed blob, same class of work as the RGA and
|
||||
session. GPL source port only: no closed blob, same class of work as the RGA and
|
||||
audio ports in this series (`../rga/PORT-DONE.md`, `../audio/PORT-PROGRESS.md`).
|
||||
|
||||
Trees involved:
|
||||
@@ -16,27 +16,27 @@ Trees involved:
|
||||
(v0.9.2, `DRIVER_DATE "20230825"`)
|
||||
|
||||
Scope and strategic framing are in `PORT-PLAN.md` (file-by-file plan, authoritative)
|
||||
and `OPEN-NPU-PLAN.md` (the "open driver, closed userspace" reality — this port
|
||||
and `OPEN-NPU-PLAN.md` (the "open driver, closed userspace" reality, this port
|
||||
delivers exactly Tier A there: an open, on-hardware-verifiable *kernel* driver, and
|
||||
nothing more; it does not and cannot unlock running a model).
|
||||
|
||||
## File set copied
|
||||
|
||||
Per `PORT-PLAN.md` §2.1's "Yes" column, copied verbatim from the vendor tree into
|
||||
`drivers/rknpu/` (patched in place afterward — see API-delta table below):
|
||||
Per `PORT-PLAN.md` section 2.1's "Yes" column, copied verbatim from the vendor tree into
|
||||
`drivers/rknpu/` (patched in place afterward, see API-delta table below):
|
||||
|
||||
| File | Bring in | Notes |
|
||||
|---|---|---|
|
||||
| `rknpu_drv.c` / `include/rknpu_drv.h` | yes | probe/remove, of_match table (incl. the RV1106 entry + `rv1106_rknpu_config`), DRM driver registration, power get/put |
|
||||
| `rknpu_job.c` / `include/rknpu_job.h` | yes | job submit, IRQ handler, PC task-list execution, `rknpu_get_hw_version()`/`rknpu_get_drv_version()` |
|
||||
| `rknpu_gem.c` / `include/rknpu_gem.h` | yes | GEM memory manager (DRM_GEM path) |
|
||||
| `rknpu_reset.c` / `include/rknpu_reset.h` | yes | `SRST_A_RKNN`/`SRST_H_RKNN` reset control — compiled unmodified, zero API drift |
|
||||
| `rknpu_iommu.c` / `include/rknpu_iommu.h` | yes (dead code path) | 61 lines; `iommu_en` stays false on our non-IOMMU DT — compiled unmodified |
|
||||
| `rknpu_debugger.c` / `include/rknpu_debugger.h` | yes | `/proc/rknpu/load` continuity target for the Monitor page — compiled unmodified |
|
||||
| `rknpu_mem.c`, `rknpu_mm.c`, `rknpu_fence.c` | **no** (headers only) | DMA_HEAP / SRAM / FENCE variants — not enabled for this port (§2.2). Their headers (`rknpu_mem.h`, `rknpu_mm.h`, `rknpu_fence.h`) *are* copied because `rknpu_drv.h`/`rknpu_gem.h` include them unconditionally for struct/prototype declarations, but zero object code results: every call site into these three is either `#ifdef CONFIG_ROCKCHIP_RKNPU_{DMA_HEAP,SRAM,FENCE}` (compiled out, all three configs left off) or `if (IS_ENABLED(CONFIG_ROCKCHIP_RKNPU_SRAM) && ...)` (compile-time-constant-folded to dead code and dropped by the optimizer — confirmed: none of `rknpu_mm_*`/`rknpu_mem_*` appear in `System.map`) |
|
||||
| `rknpu_reset.c` / `include/rknpu_reset.h` | yes | `SRST_A_RKNN`/`SRST_H_RKNN` reset control: compiled unmodified, zero API drift |
|
||||
| `rknpu_iommu.c` / `include/rknpu_iommu.h` | yes (dead code path) | 61 lines; `iommu_en` stays false on our non-IOMMU DT: compiled unmodified |
|
||||
| `rknpu_debugger.c` / `include/rknpu_debugger.h` | yes | `/proc/rknpu/load` continuity target for the Monitor page: compiled unmodified |
|
||||
| `rknpu_mem.c`, `rknpu_mm.c`, `rknpu_fence.c` | **no** (headers only) | DMA_HEAP / SRAM / FENCE variants: not enabled for this port (section 2.2). Their headers (`rknpu_mem.h`, `rknpu_mm.h`, `rknpu_fence.h`) *are* copied because `rknpu_drv.h`/`rknpu_gem.h` include them unconditionally for struct/prototype declarations, but zero object code results: every call site into these three is either `#ifdef CONFIG_ROCKCHIP_RKNPU_{DMA_HEAP,SRAM,FENCE}` (compiled out, all three configs left off) or `if (IS_ENABLED(CONFIG_ROCKCHIP_RKNPU_SRAM) && ...)` (compile-time-constant-folded to dead code and dropped by the optimizer, confirmed: none of `rknpu_mm_*`/`rknpu_mem_*` appear in `System.map`) |
|
||||
|
||||
UAPI header: `rknpu_ioctl.h` was split the way `drivers/accel/rocket/` (the mainline
|
||||
Rocket driver, already vendored in this tree) does it — the real content lives at
|
||||
Rocket driver, already vendored in this tree) does it; the real content lives at
|
||||
`include/uapi/drm/rknpu_ioctl.h` (mainline convention for DRM ioctl UAPI headers),
|
||||
and `drivers/rknpu/include/rknpu_ioctl.h` is a one-line redirect (`#include
|
||||
<drm/rknpu_ioctl.h>`) so the vendor source's unmodified `#include "rknpu_ioctl.h"`
|
||||
@@ -45,33 +45,33 @@ keeps working. This also makes the header directly usable by
|
||||
|
||||
## Kconfig / Makefile wiring
|
||||
|
||||
`drivers/rknpu/Kconfig` (new) — trimmed from the vendor's own `Kconfig`: drops the
|
||||
`drivers/rknpu/Kconfig` (new), trimmed from the vendor's own `Kconfig`: drops the
|
||||
`ROCKCHIP_RKNPU_SRAM`/`_FENCE` options and the `DRM_GEM` vs `DMA_HEAP` `choice`
|
||||
block entirely (DRM_GEM is the only memory manager this port wires up — see
|
||||
`PORT-PLAN.md` §1 "Memory manager choice"). Sourced from `drivers/Kconfig` right
|
||||
block entirely (DRM_GEM is the only memory manager this port wires up, see
|
||||
`PORT-PLAN.md` section 1 "Memory manager choice"). Sourced from `drivers/Kconfig` right
|
||||
after `source "drivers/accel/Kconfig"` (RKNPU is a classic DRM device, not
|
||||
`drivers/accel/` — see §1 of `PORT-PLAN.md`, "not `/dev/accel/`").
|
||||
`drivers/accel/`, see section 1 of `PORT-PLAN.md`, "not `/dev/accel/`").
|
||||
|
||||
`drivers/rknpu/Makefile` (new) — trimmed from the vendor's own `Makefile`: only the
|
||||
`drivers/rknpu/Makefile` (new), trimmed from the vendor's own `Makefile`: only the
|
||||
five always-on objects plus GEM (`rknpu_drv.o`, `rknpu_reset.o`, `rknpu_job.o`,
|
||||
`rknpu_debugger.o`, `rknpu_iommu.o`, `rknpu_gem.o` gated on
|
||||
`CONFIG_ROCKCHIP_RKNPU_DRM_GEM`); no edit needed for the never-built
|
||||
`rknpu_mem.o`/`rknpu_mm.o`/`rknpu_fence.o` lines because they're simply omitted
|
||||
(matches `PORT-PLAN.md` §2.1's note that no Makefile edit is required, just not
|
||||
(matches `PORT-PLAN.md` section 2.1's note that no Makefile edit is required, just not
|
||||
setting their Kconfig symbols). `ccflags-y` adds `compat/` to the include path
|
||||
**ahead of** `include/` — see the compat-shim section below for why the ordering
|
||||
**ahead of** `include/`. See the compat-shim section below for why the ordering
|
||||
matters.
|
||||
|
||||
Wired into `drivers/Makefile`: `obj-$(CONFIG_ROCKCHIP_RKNPU) += rknpu/` placed
|
||||
immediately after `obj-y += gpu/` (rknpu registers a classic DRM device, so — like
|
||||
`gpu/` itself — it must come after `char/` and `iommu/` per the existing comments
|
||||
immediately after `obj-y += gpu/` (rknpu registers a classic DRM device, so (like
|
||||
`gpu/` itself) it must come after `char/` and `iommu/` per the existing comments
|
||||
there, and building after `gpu/` specifically ensures the DRM core is ready first).
|
||||
|
||||
## Config symbols set
|
||||
|
||||
Via `./scripts/config --enable <SYMBOL>` then `make ARCH=arm CROSS_COMPILE=...
|
||||
olddefconfig`, confirmed `=y` in `.config` afterward (no silent Kconfig-dependency
|
||||
drop — all four symbols round-tripped through `olddefconfig` unchanged):
|
||||
drop, all four symbols round-tripped through `olddefconfig` unchanged):
|
||||
|
||||
```
|
||||
CONFIG_ROCKCHIP_RKNPU=y
|
||||
@@ -81,52 +81,52 @@ CONFIG_ROCKCHIP_RKNPU_PROC_FS=y
|
||||
```
|
||||
|
||||
Prerequisites already satisfied pre-port: `CONFIG_DRM=y`, `CONFIG_ARCH_ROCKCHIP=y`
|
||||
(both already on from the M4 display work). Left off per `PORT-PLAN.md` §2.2/§2.6:
|
||||
`ROCKCHIP_RKNPU_DMA_HEAP`, `ROCKCHIP_RKNPU_SRAM`, `ROCKCHIP_RKNPU_FENCE` — RV1106
|
||||
(both already on from the M4 display work). Left off per `PORT-PLAN.md` section 2.2/section 2.6:
|
||||
`ROCKCHIP_RKNPU_DMA_HEAP`, `ROCKCHIP_RKNPU_SRAM`, `ROCKCHIP_RKNPU_FENCE`; RV1106
|
||||
needs none of them (single-core NPU, no NBUF, no OPP table, no multi-domain
|
||||
fencing).
|
||||
|
||||
## The four compat-shim headers (`drivers/rknpu/compat/soc/rockchip/`)
|
||||
|
||||
All four vendor-only headers from `PORT-PLAN.md` §2.4 (absent from mainline) are
|
||||
addressed, but **not uniformly** — one of the four needed a genuinely fresh shim
|
||||
All four vendor-only headers from `PORT-PLAN.md` section 2.4 (absent from mainline) are
|
||||
addressed, but **not uniformly**, one of the four needed a genuinely fresh shim
|
||||
rather than a vendor-verbatim copy, for a reason the plan didn't anticipate:
|
||||
|
||||
### `rockchip_iommu.h` — written fresh, NOT a copy of the vendor header
|
||||
### `rockchip_iommu.h`: written fresh, NOT a copy of the vendor header
|
||||
|
||||
This is the one real surprise of the port. The vendor header itself already
|
||||
carries a working `#if IS_ENABLED(CONFIG_ROCKCHIP_IOMMU) ... #else <stub> #endif`
|
||||
split, so naively copying it looked safe. **It is not**, in this specific tree:
|
||||
`CONFIG_ROCKCHIP_IOMMU` is not a vacant symbol here — it's `=y` in our `.config`
|
||||
`CONFIG_ROCKCHIP_IOMMU` is not a vacant symbol here: it's `=y` in our `.config`
|
||||
already, for the *real*, unrelated mainline IOMMU driver
|
||||
(`drivers/iommu/rockchip-iommu.c`, pulled in by the M4 display/VOP2 work). That
|
||||
driver implements the standard `struct iommu_ops` and does **not** export a
|
||||
function called `rockchip_iommu_is_enabled()` — confirmed by grepping
|
||||
function called `rockchip_iommu_is_enabled()`, confirmed by grepping
|
||||
`drivers/iommu/rockchip-iommu.c` for the symbol (zero hits). Had the vendor
|
||||
header been copied as-is, `IS_ENABLED(CONFIG_ROCKCHIP_IOMMU)` would have evaluated
|
||||
true (not false, as the plan assumed) and selected the `extern bool
|
||||
rockchip_iommu_is_enabled(struct device *dev);` declaration with **no definition
|
||||
anywhere in the tree** — a link failure that would only show up at the very end of
|
||||
anywhere in the tree**: a link failure that would only show up at the very end of
|
||||
a full kernel build, not at this file's own compile step.
|
||||
|
||||
Fix: `drivers/rknpu/compat/soc/rockchip/rockchip_iommu.h` is a fresh 25-line file,
|
||||
unconditional (no `#if IS_ENABLED(...)` at all), providing only the one symbol
|
||||
`rknpu_drv.c` actually calls: `rockchip_iommu_is_enabled()` (used once, in
|
||||
`rknpu_power_off()`, inside `if (rknpu_dev->multiple_domains)` — true only for
|
||||
`rknpu_power_off()`, inside `if (rknpu_dev->multiple_domains)`; true only for
|
||||
RK3588's 3-core NPU; RV1106's `rv1106_rknpu_config` never sets it, and the board DT
|
||||
carries no `iommus=` property, so this is genuinely dead code, exactly as
|
||||
`PORT-PLAN.md` §2.4 predicted — the fix just couldn't be "copy the vendor header,"
|
||||
`PORT-PLAN.md` section 2.4 predicted: the fix just couldn't be "copy the vendor header,"
|
||||
it had to be "write an unconditional one that doesn't shadow-collide with this
|
||||
tree's real `CONFIG_ROCKCHIP_IOMMU`."
|
||||
|
||||
### `rockchip_opp_select.h`, `rockchip_system_monitor.h`, `rockchip_ipa.h` — copied verbatim, safe as-is
|
||||
### `rockchip_opp_select.h`, `rockchip_system_monitor.h`, `rockchip_ipa.h`: copied verbatim, safe as-is
|
||||
|
||||
Unlike `ROCKCHIP_IOMMU`, none of `CONFIG_ROCKCHIP_OPP`, `CONFIG_ROCKCHIP_SYSTEM_MONITOR`,
|
||||
`CONFIG_ROCKCHIP_IPA` exist anywhere in mainline (grepped every `Kconfig` in the
|
||||
tree and the `.config` — zero hits for all three), so these three vendor headers'
|
||||
tree and the `.config`, zero hits for all three), so these three vendor headers'
|
||||
own `#if IS_ENABLED(...)`/`#if IS_REACHABLE(...)` gates always evaluate false and
|
||||
their static-inline stub branches (`-EOPNOTSUPP`/`ERR_PTR(-ENOTSUPP)`/no-ops) are
|
||||
always selected — no collision risk, copied byte-for-byte from
|
||||
always selected, no collision risk, copied byte-for-byte from
|
||||
`sdk/sysdrv/source/kernel/include/soc/rockchip/`, with only an explanatory header
|
||||
comment added to each. The full struct definitions (`struct rockchip_opp_info`,
|
||||
`struct monitor_dev_info`, `struct ipa_power_model_data`) are still required
|
||||
@@ -137,7 +137,7 @@ pointers dereferenced in the (dead-for-RV1106, but still compiled) DVFS code pat
|
||||
|
||||
`compat/` is listed **first** in `ccflags-y` (ahead of `include/`), specifically so
|
||||
these four shadow anything a future in-tree `soc/rockchip/` addition might
|
||||
introduce — `rockchip_iommu.h` in particular must never resolve to a different,
|
||||
introduce. `rockchip_iommu.h` in particular must never resolve to a different,
|
||||
unrelated header of the same name.
|
||||
|
||||
### Devfreq/OPP/monitor/IPA code: compiles, but is entirely unreferenced for our build
|
||||
@@ -149,18 +149,18 @@ different implementations, one per era). For 6.18 the `#else` (>= 6.1) branch is
|
||||
what's textually compiled, and it references file-scope statics (`npu_mdevp`,
|
||||
`npu_devfreq_profile`, `npu_cooling_power`) that a naive manual read of the
|
||||
surrounding `#if` nesting suggested might only be declared under the sibling
|
||||
`< 6.1` branch. **This did not manifest as a build error** — `rknpu_drv.o` compiled
|
||||
`< 6.1` branch. **This did not manifest as a build error**: `rknpu_drv.o` compiled
|
||||
clean on the first fixed pass, which is the authoritative answer (per this task's
|
||||
own "prove by compiling, not by inspection" instruction — a manual preprocessor
|
||||
own "prove by compiling, not by inspection" instruction, a manual preprocessor
|
||||
trace over ~700 lines of nested version gates is exactly the kind of thing to
|
||||
distrust vs. the compiler). Confirmed after the fact: **none** of
|
||||
`rknpu_devfreq_init`, `npu_devfreq_target`, `npu_devfreq_profile`, `npu_mdevp`, or
|
||||
`npu_cooling_power` appear in `System.map` — the call site into
|
||||
`npu_cooling_power` appear in `System.map`; the call site into
|
||||
`rknpu_devfreq_init()` (in `rknpu_probe()`) is itself gated by the same `#if
|
||||
KERNEL_VERSION(6,1,0) > LINUX_VERSION_CODE`, false for 6.18, so the whole devfreq
|
||||
init path is unreachable and GCC drops the unused `static` functions entirely.
|
||||
Net effect matches `PORT-PLAN.md`'s framing exactly — RV1106 has no OPP table and
|
||||
no DVFS — just via straightforward dead-code elimination rather than anything this
|
||||
Net effect matches `PORT-PLAN.md`'s framing exactly: RV1106 has no OPP table and
|
||||
no DVFS, just via straightforward dead-code elimination rather than anything this
|
||||
port had to force.
|
||||
|
||||
## 5.10 -> 6.18 API-delta fixes (all mechanical, all found by iterating single-object builds)
|
||||
@@ -169,34 +169,34 @@ port had to force.
|
||||
|
||||
| # | Change |
|
||||
|---|---|
|
||||
| 1 | `struct drm_driver` has no `.gem_prime_mmap` member any more (the legacy driver-level `gem_prime_*` fallback vtable was removed). Dropped the `.gem_prime_mmap = drm_gem_prime_mmap,` initializer for the `KERNEL_VERSION(6,1,0) <= LINUX_VERSION_CODE` branch — not a functionality loss, because per-object mmap is already wired via `drm_gem_object_funcs.mmap = rknpu_gem_mmap_obj` in `rknpu_gem.c`'s `rknpu_gem_object_funcs` (the mechanism this field used to be a fallback *from*, per `PORT-PLAN.md` §2.5). |
|
||||
| 1 | `struct drm_driver` has no `.gem_prime_mmap` member any more (the legacy driver-level `gem_prime_*` fallback vtable was removed). Dropped the `.gem_prime_mmap = drm_gem_prime_mmap,` initializer for the `KERNEL_VERSION(6,1,0) <= LINUX_VERSION_CODE` branch: not a functionality loss, because per-object mmap is already wired via `drm_gem_object_funcs.mmap = rknpu_gem_mmap_obj` in `rknpu_gem.c`'s `rknpu_gem_object_funcs` (the mechanism this field used to be a fallback *from*, per `PORT-PLAN.md` section 2.5). |
|
||||
| 2 | `struct drm_driver` has no `.date` member any more (dropped from mainline DRM). Removed the `.date = DRIVER_DATE,` initializer; `.major`/`.minor`/`.patchlevel` already carry the version. |
|
||||
| 3 | `hrtimer_init(&t, clock, mode)` + separate `t.function = fn` assignment -> combined `hrtimer_setup(&t, fn, clock, mode)` (same fix class as the RGA port's hrtimer change, `../rga/PORT-DONE.md` item 2). |
|
||||
| 4 | `platform_driver.remove`: `int (*)(struct platform_device *)` -> `void (*)(struct platform_device *)`. Changed `rknpu_remove()` from `static int ... { ...; return 0; }` to `static void ...` (dropped the trailing `return 0;`) — same fix as the RGA and audio ports. |
|
||||
| 4 | `platform_driver.remove`: `int (*)(struct platform_device *)` -> `void (*)(struct platform_device *)`. Changed `rknpu_remove()` from `static int ... { ...; return 0; }` to `static void ...` (dropped the trailing `return 0;`), same fix as the RGA and audio ports. |
|
||||
| 5 | `MODULE_IMPORT_NS(DMA_BUF)` -> `MODULE_IMPORT_NS("DMA_BUF")` (quoted-string form; same fix as the RGA port item 6). |
|
||||
|
||||
### `rknpu_gem.c`
|
||||
|
||||
| # | Change |
|
||||
|---|---|
|
||||
| 1 | `<linux/pfn_t.h>` and the `pfn_t` wrapper type (`__pfn_to_pfn_t()`, `PFN_DEV`) were removed entirely from mainline. Dropped the include; `vmf_insert_mixed()` (the only call site reachable at `KERNEL_VERSION(4,15,0) <= LINUX_VERSION_CODE`, which is our branch) now takes a plain `unsigned long pfn` directly — `pfn = page_to_pfn(...)` was already computing that raw value, so the fix is just passing `pfn` instead of `__pfn_to_pfn_t(pfn, PFN_DEV)`. |
|
||||
| 1 | `<linux/pfn_t.h>` and the `pfn_t` wrapper type (`__pfn_to_pfn_t()`, `PFN_DEV`) were removed entirely from mainline. Dropped the include; `vmf_insert_mixed()` (the only call site reachable at `KERNEL_VERSION(4,15,0) <= LINUX_VERSION_CODE`, which is our branch) now takes a plain `unsigned long pfn` directly: `pfn = page_to_pfn(...)` was already computing that raw value, so the fix is just passing `pfn` instead of `__pfn_to_pfn_t(pfn, PFN_DEV)`. |
|
||||
| 2 | `vmap()`/`vunmap()`/`VM_MAP` used to be pulled in transitively; 6.18 needs `<linux/vmalloc.h>` included explicitly. Added it. |
|
||||
| 3 | `%zu` format specifier for `rknpu_obj->size` (`-Werror=format=`): the field is declared `unsigned long`, not `size_t`, on this target — changed to `%lu`. (Pure `-Wformat` pickiness, not a real 5.10-vs-6.18 delta; the vendor's own type just doesn't match `%zu` on this ABI and newer GCC/kernel `-Werror` catches it.) |
|
||||
| 3 | `%zu` format specifier for `rknpu_obj->size` (`-Werror=format=`): the field is declared `unsigned long`, not `size_t`, on this target, changed to `%lu`. (Pure `-Wformat` pickiness, not a real 5.10-vs-6.18 delta; the vendor's own type just doesn't match `%zu` on this ABI and newer GCC/kernel `-Werror` catches it.) |
|
||||
| 4 | `iommu_map()` gained a trailing `gfp_t gfp` argument (`iommu_map(domain, iova, paddr, size, prot)` -> `iommu_map(domain, iova, paddr, size, prot, gfp)`). Added `GFP_KERNEL` at both call sites (cache-buffer path and the per-sg-entry DDR path). |
|
||||
| 5 | `vma->vm_flags` is a read-only field now (direct assignment is a compile error, not just deprecated) — replaced all four sites with `vm_flags_set()`/`vm_flags_clear()`: `rknpu_gem_mmap_pages()` (`VM_MIXEDMAP`), `rknpu_gem_mmap_cache()` (`VM_MIXEDMAP`), `rknpu_gem_mmap_buffer()` (`VM_DONTCOPY\|VM_DONTEXPAND\|VM_DONTDUMP\|VM_IO` set, `VM_PFNMAP` cleared). The three read-only accesses (`vm_get_page_prot(vma->vm_flags)` in `rknpu_gem_mmap`) needed no change — only assignment is blocked. |
|
||||
| 5 | `vma->vm_flags` is a read-only field now (direct assignment is a compile error, not just deprecated), replaced all four sites with `vm_flags_set()`/`vm_flags_clear()`: `rknpu_gem_mmap_pages()` (`VM_MIXEDMAP`), `rknpu_gem_mmap_cache()` (`VM_MIXEDMAP`), `rknpu_gem_mmap_buffer()` (`VM_DONTCOPY\|VM_DONTEXPAND\|VM_DONTDUMP\|VM_IO` set, `VM_PFNMAP` cleared). The three read-only accesses (`vm_get_page_prot(vma->vm_flags)` in `rknpu_gem_mmap`) needed no change: only assignment is blocked. |
|
||||
|
||||
`rknpu_reset.c`, `rknpu_iommu.c`, `rknpu_debugger.c`, `rknpu_job.c` needed **zero**
|
||||
changes — compiled clean against 6.18 unmodified, confirming `PORT-PLAN.md` §2.5's
|
||||
changes: compiled clean against 6.18 unmodified, confirming `PORT-PLAN.md` section 2.5's
|
||||
assessment that most of the surface was already correctly version-gated by the
|
||||
vendor.
|
||||
|
||||
## Device tree change
|
||||
|
||||
One board-DTS override appended to `arch/arm/boot/dts/rockchip/rv1106-warden.dts`
|
||||
(base `rv1106.dtsi`'s `npu@ff660000` node — `compatible`, `reg`, `interrupts`
|
||||
(base `rv1106.dtsi`'s `npu@ff660000` node: `compatible`, `reg`, `interrupts`
|
||||
(`GIC_SPI 109`), `clocks` (`ACLK_RKNN`/`HCLK_RKNN`), `assigned-clock-rates
|
||||
= <420000000>`, `resets` (`SRST_A_RKNN`/`SRST_H_RKNN`) — left untouched, per
|
||||
instructions, exactly as `PORT-PLAN.md` §2.3 specified):
|
||||
= <420000000>`, `resets` (`SRST_A_RKNN`/`SRST_H_RKNN`), left untouched, per
|
||||
instructions, exactly as `PORT-PLAN.md` section 2.3 specified):
|
||||
|
||||
```dts
|
||||
&npu {
|
||||
@@ -225,7 +225,7 @@ make ARCH=arm CROSS_COMPILE=$CROSS_COMPILE zImage rockchip/rv1106-warden.dtb -j"
|
||||
```
|
||||
|
||||
Result: **exit 0**. `grep -iE "error|warn"` over the full build log returns exactly
|
||||
one line — the pre-existing, NPU-unrelated dtc warning noted above. Zero errors,
|
||||
one line: the pre-existing, NPU-unrelated dtc warning noted above. Zero errors,
|
||||
zero rknpu-related warnings. `arch/arm/boot/zImage` (8.7 MB) and
|
||||
`arch/arm/boot/dts/rockchip/rv1106-warden.dtb` (36.9 KB) both produced. XZ kernel
|
||||
compression and the existing console/earlycon config were left untouched, per
|
||||
@@ -235,7 +235,7 @@ etc.) was reverted or altered.
|
||||
## Verification: rknpu is genuinely linked in (not silently dropped to a module)
|
||||
|
||||
Per the task's explicit instruction, checked with **host** `grep` on `System.map`
|
||||
(ground truth for built-in linkage — cross-`nm` mis-lists symbols on this
|
||||
(ground truth for built-in linkage, cross-`nm` mis-lists symbols on this
|
||||
toolchain, same caveat as every other port in this series):
|
||||
|
||||
```
|
||||
@@ -259,7 +259,7 @@ symbol presence the task asked to confirm.
|
||||
## Hardware test program: `rknpu_version_test.c`
|
||||
|
||||
Written to `warden-sdk/kernel/rv1106-enablement/npu/rknpu_version_test.c` per the
|
||||
task's spec. Dependency-free beyond the kernel tree's own UAPI headers — no
|
||||
task's spec. Dependency-free beyond the kernel tree's own UAPI headers, no
|
||||
libdrm, no target sysroot headers:
|
||||
|
||||
```sh
|
||||
@@ -271,11 +271,11 @@ arm-rockchip830-linux-uclibcgnueabihf-gcc \
|
||||
<warden-sdk>/kernel/rv1106-enablement/npu/rknpu_version_test.c
|
||||
```
|
||||
|
||||
**This exact command was run in this session** (build-only — the resulting binary
|
||||
**This exact command was run in this session** (build-only, the resulting binary
|
||||
was not copied to or executed on any target) and produced a clean ARM EABI5 static
|
||||
ELF binary with exit code 0. One expected, harmless warning appears:
|
||||
`#warning "Attempt to use kernel headers from user space"` from
|
||||
`include/uapi/linux/types.h` — the standard notice every raw-kernel-uapi-header
|
||||
`include/uapi/linux/types.h`, the standard notice every raw-kernel-uapi-header
|
||||
userspace build gets; it does not affect correctness (`__u32` etc. are still
|
||||
correctly defined with `__KERNEL__` undefined).
|
||||
|
||||
@@ -285,13 +285,13 @@ then again with `.flags = RKNPU_GET_HW_VERSION`, and prints both. Traced against
|
||||
driver source to get the exact semantics right:
|
||||
- `RKNPU_GET_DRV_VERSION` returns `RKNPU_GET_DRV_VERSION_CODE(DRIVER_MAJOR,
|
||||
DRIVER_MINOR, DRIVER_PATCHLEVEL)` = `MAJOR*10000 + MINOR*100 + PATCHLEVEL`
|
||||
(`rknpu_drv.c:rknpu_get_drv_version()`) — for this port's unmodified
|
||||
(`rknpu_drv.c:rknpu_get_drv_version()`), for this port's unmodified
|
||||
`DRIVER_MAJOR/MINOR/PATCHLEVEL = 0/9/2`, that's raw code `902`, which the test
|
||||
program decodes back to `"0.9.2"` via the UAPI header's own
|
||||
`RKNPU_GET_DRV_VERSION_{MAJOR,MINOR,PATCHLEVEL}()` macros.
|
||||
- `RKNPU_GET_HW_VERSION` returns a raw value read directly off the NPU core's
|
||||
`VERSION`/`VERSION_NUM` registers (`rknpu_job.c:rknpu_get_hw_version()`) — no
|
||||
published decode table exists for it (per `PORT-PLAN.md` §3 step 5, "checks...
|
||||
`VERSION`/`VERSION_NUM` registers (`rknpu_job.c:rknpu_get_hw_version()`): no
|
||||
published decode table exists for it (per `PORT-PLAN.md` section 3 step 5, "checks...
|
||||
returns something plausible"); the test program just prints it in hex.
|
||||
|
||||
**Expected output on a successful hardware run** (parent session):
|
||||
@@ -305,7 +305,7 @@ and clock/reset all exercised.
|
||||
```
|
||||
|
||||
A driver version that decodes to anything other than `0.9.2` would indicate a stale
|
||||
build or a stub/mock intercepting the ioctl, not a real driver response — that's
|
||||
build or a stub/mock intercepting the ioctl, not a real driver response: that's
|
||||
the value of checking the decoded string, not just the ioctl return code.
|
||||
|
||||
## Explicitly deferred to the parent session (not done here, per this task's boundary)
|
||||
@@ -314,23 +314,23 @@ the value of checking the decoded string, not just the ioctl return code.
|
||||
- Boot-time probe verification: `dmesg | grep -i rknpu` should show clean
|
||||
clock/reset/IRQ acquisition, no panic, no permanent `-EPROBE_DEFER` (a single
|
||||
deferral early at boot, before other clock/reset providers are up, would be
|
||||
normal — same caveat class as the audio port's acodec probe-order note).
|
||||
Compare against the 5.10 baseline probe log if available (`PORT-PLAN.md` §3
|
||||
normal, same caveat class as the audio port's acodec probe-order note).
|
||||
Compare against the 5.10 baseline probe log if available (`PORT-PLAN.md` section 3
|
||||
step 3).
|
||||
- `ls -la /dev/dri/` should show a new `cardN`/`renderD1xx` — classic DRM node
|
||||
- `ls -la /dev/dri/` should show a new `cardN`/`renderD1xx`: classic DRM node
|
||||
(this port intentionally does **not** produce a `/dev/accel/` node; see
|
||||
`PORT-PLAN.md` §1 and `OPEN-NPU-PLAN.md` §1.3 for why mainline's own
|
||||
`drivers/accel/rocket/` driver — RK3588/ARM64-only — is a different codebase
|
||||
`PORT-PLAN.md` section 1 and `OPEN-NPU-PLAN.md` section 1.3 for why mainline's own
|
||||
`drivers/accel/rocket/` driver (RK3588/ARM64-only) is a different codebase
|
||||
that doesn't reach RV1106 at all).
|
||||
- Run `rknpu_version_test` (built above) against the real node; confirm the
|
||||
decoded driver version prints `0.9.2` and the hw version is non-zero/plausible.
|
||||
- **Not attempted, not required for this milestone**: `DRM_IOCTL_RKNPU_SUBMIT`,
|
||||
any `.rknn` model, `librknnrt` — that's the closed-userspace question
|
||||
`OPEN-NPU-PLAN.md` §1.2–1.4 covers; out of scope here by design (no blob is
|
||||
any `.rknn` model, `librknnrt`, that's the closed-userspace question
|
||||
`OPEN-NPU-PLAN.md` section 1.2-1.4 covers; out of scope here by design (no blob is
|
||||
shipped by this port, and none is needed to prove the kernel driver itself).
|
||||
|
||||
## Also updated this session
|
||||
|
||||
`../DRIVER-PARITY.md`'s NPU row: `[ ] M6 — plan: npu/PORT-PLAN.md` -> `[wip] M6 built,
|
||||
`../DRIVER-PARITY.md`'s NPU row: `[ ] M6, plan: npu/PORT-PLAN.md` -> `[wip] M6 built,
|
||||
0 errors/0 warnings, 99 rknpu-prefixed symbols in System.map, &npu
|
||||
{status="okay"} in the dtb — not yet flashed/probed on hardware`.
|
||||
{status="okay"} in the dtb, not yet flashed/probed on hardware`.
|
||||
|
||||
@@ -1,4 +1,4 @@
|
||||
# NPU (rknpu) open kernel driver — VERIFIED on warden-c8a3 (2026-08-25)
|
||||
# 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
|
||||
@@ -17,7 +17,7 @@ $ rknpu_version_test
|
||||
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
|
||||
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
|
||||
@@ -25,16 +25,16 @@ 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`.)
|
||||
(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
|
||||
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
|
||||
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
|
||||
|
||||
@@ -9,7 +9,7 @@
|
||||
* RKNPU_GET_HW_VERSION, and prints the decoded results. This exercises the
|
||||
* full ioctl-dispatch -> power-get/put -> clock/reset path with zero
|
||||
* dependency on a regcmd buffer or the (closed) RKNN runtime -- see
|
||||
* OPEN-NPU-PLAN.md §1.4 Tier A and PORT-PLAN.md §3 step 5.
|
||||
* OPEN-NPU-PLAN.md section 1.4 Tier A and PORT-PLAN.md section 3 step 5.
|
||||
*
|
||||
* Dependency-free beyond the UAPI header: build with
|
||||
* -I<kernel-tree>/include/uapi
|
||||
@@ -41,7 +41,7 @@
|
||||
* value read directly off the NPU core's VERSION/VERSION_NUM registers
|
||||
* (rknpu_job.c:rknpu_get_hw_version()) -- any non-zero, non-0xffffffff
|
||||
* value is a plausible "the register block is alive" signal; there is no
|
||||
* published decode table for it beyond that (PORT-PLAN.md §3 step 5).
|
||||
* published decode table for it beyond that (PORT-PLAN.md section 3 step 5).
|
||||
*/
|
||||
|
||||
#include <errno.h>
|
||||
|
||||
Reference in New Issue
Block a user