RK3568 NPU: what triggers the CNA weight-fetch DMA after the executer engages? (mainline rocket vs vendor rknpu)
Hi Chaoyi, hi all,
Following up on the RK3568 NPU work — thanks again for the AUTO_GATING bit-31 fix and the MMU-v2 clarification; both are in the driver now.
The mainline rocket accel driver reaches a good state on the RK3568
NPU: it probes, powers the domain, programs the pipeline, and submits a job cleanly. One functional problem remains, and I have narrowed it down to a single hardware step that I don't think I can resolve from outside Rockchip.
The test is one convolution (conv2d, 16 -> 128 channels, 5x5, stride 2) run through Teflon/mesa. On the vendor rknpu driver that conv produces correct output. On rocket the job engages the whole pipeline but never
computes: the output tensor comes back uniformly at the quantization zero-point (every element == 128), and the job times out with no completion interrupt.
I captured the working vendor rknpu and the stuck rocket running the SAME conv, on the SAME SoC, in the same boot, and drove both to the identical engaged hardware state. The per-unit status and the CNA configuration registers match exactly:
- both reach S_STATUS = 0x0c and S_POINTER = 0x1000e (executer bit 16
set) on all five sub-units (CNA / CMAC / ACCU / DPU / RDMA);
- PC TASKST = 0xf000 on both;
- the CNA config is byte-identical: CONV_CON*, DATA_SIZE*, WEIGHT_SIZE*, CBUF.
The divergence is entirely in the data-movement counters (the NPU perf registers). Read straight off the hardware:
counter (perf reg) vendor rknpu rocket what it counts ------------------- ------------ ------ ---------------------- wt_rd (0x803c) 0x1900 0 CNA weight-fetch DMA dt_wr (0x8034) 0x6400 0 DPU output writes dt_rd (0x8038) 0x32c9 0x8b feature / regcmd reads
The vendor's counters describe a full convolution: its CNA fetches weights (wt_rd = 0x1900) and its DPU writes the complete output (dt_wr = 0x6400). rocket's CNA never fetches weights (wt_rd = 0) and its DPU never writes (dt_wr = 0); rocket's dt_rd = 0x8b accounts for only the regcmd fetch, so it never even reads the feature data. With no weights the MAC array starves and the DPU writes nothing — which is exactly the uniform zero-point output we see.
(One aside for anyone who has chased this IP before: TASKST = 0xf000 is the NORMAL working value — the vendor produces full output at 0xf000. It is not a stall.)
I have matched every software-visible surface I can find:
- Command stream + data. Byte-exact replay of the vendor's captured conv2d regcmd, together with its weight BO and input BO, through the rocket UABI — still no compute. The content is not the gate.
- state_init. rk356x_rknpu_config.state_init is NULL on RK3568 (only
RK3576 has one). I tried the RK3576 ritual (0x1024 = 0x80000000,
S_POINTER 0 -> 1 -> 0x1e) on RK3568 anyway — no change. - Reset / power / clocks / IOMMU. rocket mirrors rknpu_soft_reset (srst assert/deassert, SCMI/PVTPLL two-step, NoC de-idle at PMU 0xfdd90000, IOMMU re-establish); clocks match the vendor (SCMI 1 GHz, CLK_NPU / aclk 200 MHz); AUTO_GATING uses your bit-31 fix.
- Non-NPU-block writes (captured live for both drivers in the same boot). The only raw non-NPU writes the vendor makes are the NoC/PMU de-idle (NOC_AUTO = 0xfffb, BUS_IDLE_ST = 0x1ea, PWR_STATUS = 0x7f); rocket makes them identically. There is no npu_grf read-margin and no memory-repair step on RK3568 (those exist only on RK3576 / RK3588).
With the executer engaged (S_STATUS = 0x0c, S_POINTER bit 16 set) and the CNA's regcmd config already landed, what actually triggers the CNA's weight-fetch DMA? The vendor rknpu reaches wt_rd = 0x1900 from a clean submit; rocket, on the same SoC with the same engage and the same regcmd, stays at wt_rd = 0.
My remaining hypotheses are all hardware-internal, and I would value your read on which (if any) is real:
1. an ordering/timing requirement between the engage and the weight-fetch trigger that a byte-identical but differently-sequenced submit does not satisfy;
2. an internal CNA/CBUF sequencer or credit/handshake that the vendor's power-on or first-job path leaves armed, and that is not a CPU-visible register;
3. an undocumented CNA precondition (a CBUF / DCOMP / weight-DMA enable) that must be set before the fetch will issue.
Even a pointer of the form "the weight fetch won't start unless X is set" would unblock the driver. The rest of the driver is upstream-ready; this is the only thing between "submits" and "computes", and I would be happy to run any targeted register capture on the board.
Thanks in advance.
Kind regards
Midgy