NNIE backend for cv500 (CNN inference)

[widgetii]

cv500 has CNN inference on a separate NNIE block at 0x11100000 (IRQ SPI 45) — distinct from the classic IVE block at 0x11230000 that ive_neo drives. Today our cv500 dispatcher returns -EOPNOTSUPP for all XNN/CNN ops (gated by chip.has_xnn = false since the IVE block has no NEO unit). To run real CNN models on cv500 we need a new driver for the NNIE block.

This was explicitly deferred from the original cv500 plan (see plan section "Different IP layout").

Architectural choices

A. New kernel/nnie_neo/ module — separate driver bound to compatible="hisilicon,hisi-nnie", independent register window, independent IRQ. Cleanest separation but duplicates a lot of submit-chain infrastructure already in ive_neo.
B. Extend ive_neo with a second platform_driver entry for the NNIE compatible. Reuses ive_submit_chain and the chip-ops abstraction; cv500 grows a second register region.

(A) matches how the vendor ships it (hi3516cv500_nnie.ko separate from hi3516cv500_ive.ko). Probably the right call long-term.

What's involved

1. Wire format: cv500 libive.so exposes HI_MPI_IVE_CNN_LoadModel / _Predict / _GetResult / _UnloadModel. Find ioctl numbers + arg buffer layout (similar effort to PerspTrans/Hog RE in kernel/ive_neo + libraries/ive_neo: cv500 PerspTrans + Hog HW dispatch (default-on) #107).
2. HW task-node layout for NNIE — different from IVE's 208-byte node; need RE pass.
3. Memory arbitration [0x90] — the magic-bit sequence we skipped for cv500 IVE is required for NNIE (it's the Conv-unit DRAM-priority knob). Sequence needs to be dumped from vendor hi3516cv500_nnie.ko blob.
4. NNIE has its own queue/scheduler ABI — vendor blob probably uses the same pattern as IVE (ive_create_task / ive_schedule_task / handle tracking).
5. Userland: libive_neo already has stub HI_MPI_IVE_CNN_* returning HI_ERR_IVE_NOT_SUPPORT. Need real marshalling + ioctl path.

Scope estimate

Multi-day RE per major op (CNN_LoadModel + CNN_Predict are the load-bearing two). Similar shape to the PerspTrans/Hog work but with twice the surface (model loader is large; HW dispatch has more state to manage).

Definition of done

• New kernel module open_nnie_neo.ko bound to hisilicon,hisi-nnie on cv500.
• CNN model loader + forward pass produce non-trivial output on a small test model.
• Round-trip test in libraries/ive_neo/test/ that loads a tiny convolutional model and checks the output classification is plausible.
• Memory arbitration [0x90] sequence captured from the cv500 vendor blob and applied.

Related

• kaeru ive-neo-cv500-hw-init-reg-window-mismatch documents why [0x90] was skipped for IVE.
• kaeru hiive-cnn-chip-support-matrix for which HiSilicon SoCs actually support CNN HW.