ampere-vp9-enablement/phase2_close.md

Phase 2 close — VP9 backend compiles clean

Date: 2026-05-17 ~02:55. Phase 2 (Implement rkvdec-vdpu381-vp9.c backend) closed at the "compiles clean against 7.0.0-rc3 ARM64 kernel" milestone.

Three commits on boltzmann:~/src/linux-rockchip branch vp9-enablement-iter1

Commit	Subject	LoC
47431635801d	media: rkvdec: Add VP9 register-struct definitions for VDPU381 variant	+238
da8482271938	media: rkvdec: Add shared VP9 codec-spec helpers in rkvdec-vp9-common.h	+511
71cc0d96d212	media: rkvdec: Add VP9 backend for VDPU381 variant (RK3588)	+513

Total: +1262 lines across 4 files. The vdpu381-hevc reference (Casanova v7.0) is 639 lines; Phase 1 estimated VP9 backend at 550-700; we landed at 513 backend + 511 common-header = within range.

Build verification

boltzmann$ cd ~/src/linux-rockchip
boltzmann$ PATH=/home/mfritsche/bin:$PATH make -j$(nproc) \
    ARCH=arm64 CROSS_COMPILE=aarch64-linux-gnu- \
    drivers/media/platform/rockchip/rkvdec/

  CC [M]  drivers/media/platform/rockchip/rkvdec/rkvdec-vdpu381-vp9.o
  LD [M]  drivers/media/platform/rockchip/rkvdec/rockchip-vdec.o

Zero errors, zero warnings on the new file after fixing two field-name typos caught by the compiler:

Bug	Fix
Used BSP-style reg128_rlc_base / reg130_decout_base / reg131_colmv_cur_base for struct rkvdec_vdpu381_regs_common_addr fields	Mainline naming is plain rlc_base / decout_base / colmv_cur_base; corrected
Unused variable i in vdpu381_config_vp9_regs	Removed

These both are pre-flash compiler-caught bugs of exactly the kind Janet predicted in v2 review re. namespace bridging — the legacy regs->common.reg02.dec_mode style is muscle-memory from reading the legacy file. Corrected without regression.

What's NOT yet correct (deferred TODOs in the backend file)

5 TODO Phase 2.1 markers in vdpu381_config_vp9_regs for fields that compile fine but aren't yet packed correctly. These will be needed for first-light decode but the skeleton is structurally correct:

1. RCB address setup (common_addr.rcb_base[10] — currently NULL; needs vdpu381_set_rcb_addrs equivalent to HEVC path)
2. Segmentation register packing per segid 0..7 (reg067_074_segid[] — currently zero-init; needs config_seg_registers port)
3. Loop-filter ref_deltas[0..3] bit-pack into reg094_ref_deltas_lastframe.ref_deltas_lastframe (28-bit field combining 4× 7-bit deltas)
4. Loop-filter mode_deltas[0..1] bit-pack into reg075_last_frame_info.mode_deltas_lastframe (14-bit field)
5. Probability storage rotation: reg160_delta_prob_base, reg162_last_prob_base, reg172_update_prob_wr_base (per BSP Vdpu382Vp9dCtx_t::prob_idx cycle logic)

Open questions surfaced during Phase 2 implementation

Q	Question	Status
C-1	reg103_frame_flags extra bits (prob_update_en, txfmmode_rfsh_en, etc.) — required for first decode?	Hypothesis: only the most-obviously-needed bits set (intra_only, refresh_en, allow_high_precision_mv, interp_filter_switch_en); others left zero — verify empirically
C-2	reg160_delta_prob_base and reg162_last_prob_base — new DMA allocations or alias existing?	Hypothesis: alias priv_tbl.probs regions; verify against BSP vdpu382_set_vp9d_buffer equivalent
C-3	reg028.vp9_rd_prob_idx/vp9_wr_prob_idx first-frame values?	Currently both = 0 in skeleton. BSP rotates; will need iteration
C-4	RCB write-out — does mainline rkvdec_setup_rcb already populate common_addr.rcb_base[], or does each backend need to wire it?	HEVC backend (rkvdec-vdpu381-hevc.c:362-374) writes RCB addrs explicitly. Mirror that pattern in VP9 backend
C-5	reg032_timeout_threshold — set to RKVDEC_TIMEOUT_4K (0x2cfffff) for 4K-max VP9?	Yes; add to register setup

Push status

git push gitea vp9-enablement-iter1 from boltzmann is hanging (ssh timeout). Commits are safely on boltzmann disk in branch vp9-enablement-iter1. Will retry push next session. Branch will eventually land at git.reauktion.de/marfrit/linux-rk3588-marfrit:vp9-enablement-iter1.

What Phase 3 needs (next-session entry point)

1. Retry push to gitea from boltzmann to back up the work off-disk.
2. Implement the 5 TODOs in vdpu381_config_vp9_regs. Each is a 5-15 line bit-packing task referencing the BSP hal_vp9d_vdpu382.c source. Expect ~100-200 LoC delta.
3. Build module + backup ampere's current rockchip-vdec.ko per feedback_backup_before_module_replace before install.
4. Install + modprobe cycle — scp boltzmann:~/src/linux-rockchip/drivers/media/platform/rockchip/rkvdec/rockchip-vdec.ko ampere:/lib/modules/7.0.0-rc3-devices+/kernel/drivers/media/platform/rockchip/rkvdec/, then ssh ampere "sudo modprobe -r rockchip-vdec && sudo modprobe rockchip-vdec".
5. Sanity enumeration: v4l2-ctl --device=/dev/video0 --list-formats-out should show VP9F alongside H264_SLICE and HEVC_SLICE.
6. Pre-flight diagnostic: instrument vdpu381_irq_handler with pr_warn("vp9 irq status=0x%x\n", status) for first 5 decodes (per Janet Amendment C) to verify IRQ status bit semantics before counting on VDPU381_STA_INT_DEC_RDY_STA = BIT(2).
7. First decode attempt: LIBVA_DRIVER_NAME=v4l2_request ffmpeg -hwaccel vaapi -i bbb_30s_720p.vp9.webm -vf hwdownload,format=nv12 -frames:v 5 -f rawvideo /tmp/hw-vp9.nv12. Watch dmesg for IRQ status. Most likely outcome on first attempt with TODO #1-5 unresolved: watchdog 2-second timeout (no decode complete IRQ).
8. Iterate on the TODOs until first-light. Then SW byte-compare on a non-fade-in frame index.

Persistence checkpoint

• Campaign repo: 4 commits at git.reauktion.de/claude-noether/ampere-vp9-enablement master (phase0, phase1, phase1_v2, phase1_v3_amendment, phase2_field_mapping, this doc).
• Kernel branch: 3 commits on boltzmann:~/src/linux-rockchip:vp9-enablement-iter1 (not yet pushed to gitea due to ssh timeout).
• Module artifact: boltzmann:~/src/linux-rockchip/drivers/media/platform/rockchip/rkvdec/rockchip-vdec.ko (compiled, NOT yet installed on ampere).
• No ampere changes: ampere is still on sibling-campaign close state (HEVC bit-perfect, iter3+iter4 kernel patches in modules).
• Rockchip BSP reference: boltzmann:~/src/rockchip-bsp/{mpp,rk-kernel-6.6} — clones retained for Phase 3 cross-reference of Vdpu382Vp9dCtx_t::prob_idx rotation and RCB setup.

Phase 2 closes here. Substantial code artifact in hand; remaining bit-level field translation is well-scoped (5 TODOs, all with BSP reference locations identified) for the next session.