perf(hip): run cambi and SpEED entirely on the device

[lusoris]

Summary

Closes T-HIP-CAMBI-HOST-RESIDUAL-2026-09-29 and T-HIP-SPEED-HOST-RESIDUAL-2026-09-29.
Built on #1636 (fix/hip-rc3-parity), which is merged; the base is now master.

Ports the SYCL device-resident architectures of ADR-1357 (CAMBI) and ADR-1358 (SpEED) to the HIP backend as ADR-1378 and ADR-1384. cambi_hip, speed_chroma_hip, and speed_temporal_hip no longer run any stage of the CPU extractors on the host or perform per-frame host-device round trips.

Each frame is one staged upload (vmaf_hip_picture_upload_staged()), the entire pipeline executes on the extractor's stream, followed by one compact readback (CambiHipResults 88 bytes, SpeedGpuFrameResult 24/40 bytes), with collect() being the only host wait.

Parity & Validation on ryzen-4090-arc (gfx1036, ROCm 7.2.4)

• CAMBI:
  • Netflix 576x324 pair: 48/48 frames identical to CPU (max diff 0.0).
  • BBB 3840x2160 (50 frames): 50/50 frames identical to CPU (max diff 0.0, bound 2.2e-15).
  • Short-frame domain (1920x64, 1920x128, 1920x160, 3840x128): runs clean and bit-identical to CPU reference.
  • Timings (ms/frame, median of 3 of (t(22) - t(2)) / 20, before -> after, CPU at 16 threads):
    • 576x324: 1.71 -> 1.80 ms/frame (CPU16: 0.14)
    • 3840x2160: 91.21 -> 130.55 ms/frame (CPU16: 10.48)
• SpEED:
  • With exact log2f (LD_PRELOAD=/tmp/crlog2f.so), speed_gpu_parity.py --backend hip exits 0, 100% bit-identical to CPU on all channels and frames.
  • On stock glibc 2.44, glibc log2f misrounding causes max diff 1.43e-6 on 1-2 frames of speed_chroma_v.
  • Short-frame domain: 1920x64 cleanly rejected with -234 (< 80 min dim); 1920x128/160 and 3840x128 run clean and stay in bounds.
  • Timings (ms/frame, median of 3, before -> after, CPU at 16 threads):
    • speed_chroma 576x324: 1.41 -> 0.70 ms/frame (CPU16: 0.07)
    • speed_chroma 3840x2160: 22.81 -> 5.55 ms/frame (CPU16: 5.37) — 4.1x speedup
    • speed_temporal 576x324: 1.25 -> 0.74 ms/frame (CPU16: 0.40)
    • speed_temporal 3840x2160: 46.54 -> 15.15 ms/frame (CPU16: 20.21) — 3.1x speedup

Re-run after the rebase (2026-10-01)

Rebased onto master 591d53449 (after #1636, #1637, #1630 and the CAMBI short-frame fix #1642). On the gfx1036, same host and ROCm:

• python3 scripts/dev/speed_gpu_parity.py --backend hip --feature cambi --no-timing: 48/48 and 50/50 frames identical (max diff 0.0), as above.
• cambi_hip against the CPU cambi with fix(cambi): keep the c-values walks inside short frames #1642's clipped window: identical on banded 1920x64, 1920x128, 1920x160, 1920x176, 2560x240 and 3840x128 clips, and on narrow 64x1920 and 96x1080 vertical ramps. This is the short- and narrow-frame check ADR-1393 asks of the device c-values kernel.
• SpEED with LD_PRELOAD=/tmp/crlog2f.so: every output identical on both clips, exit 0. On stock glibc: speed_chroma_v 47/48 and 49/50 frames identical (max 1.43e-6), speed_chroma_u 48/48 and 49/50 (9.5e-7), speed_chroma_uv 47/48 and 48/50 (9.5e-7), speed_temporal identical.
• Short 4:2:0 frames: speed_temporal_hip equals the CPU at 1920x128, 1920x160 and 3840x128; speed_chroma_hip equals it at 1920x160 and, like the CPU speed_chroma, refuses 1920x64, 1920x128 and 3840x128 with -EINVAL (exit 234), whose chroma planes are below the 80-pixel minimum.
• python3 scripts/ci/run_meson_test.py -- -C build-hip --suite gpu --num-processes 1: 52 OK, no Memory access fault; the fast suite passes on the HIP build (250 OK) and on a CPU-only build (194 OK, 1 skipped). One of three full device runs failed test_hip_upload_race on vif_hip (4 of 32 scores, the four scales of one frame). That is the gfx1036 command loss recorded as T-HIP-GFX1036-DROPPED-DISPATCHES-2026-10-01, in an extractor this PR does not touch; the test then passed 20 of 20 times on its own, on this build and on the fix(hip): bring motion_hip and the HIP option twins onto the CPU's arithmetic #1636 build.

The timings above were not measured again; the rebase changes no HIP source of this PR. Two commits were added: one aligns the CAMBI notes #1642 wrote (it described the HIP twin as running the c-values walk on the host) with the device-resident twin, and one regenerates the ADR citation map.

Deep-dive deliverables (ADR-0108)

• Research digest — docs/research/1378-hip-cambi-speed-device-resident.md.
• Decision matrix — ## Alternatives considered in ADR-1378 and ADR-1384.
• AGENTS.md invariant note — core/src/feature/hip/AGENTS.md (device-resident CAMBI and SpEED pipelines) and docs/development/rebase-sensitive-invariants.md; core/src/feature/AGENTS.md now says which twins still call the host c-values walk.
• Reproducer / smoke-test command — under "Reproducer" below.
• CHANGELOG fragment — changelog.d/changed/perf-hip-cambi-device-resident.md, changelog.d/changed/perf-hip-speed-device-resident.md.
• Rebase note — docs/rebase-notes.md, "ADR-1378 / ADR-1384 — HIP CAMBI and SpEED run entirely on the device".

Reproducer

On an AMD host (ryzen-4090-arc, gfx1036):

meson setup build-hip core -Denable_hip=true -Denable_hipcc=true -Dhip_gfx_targets=gfx1036 -Db_lto=false
ninja -C build-hip
python3 scripts/ci/run_meson_test.py -- -C build-hip --suite gpu --num-processes 1
python3 scripts/dev/speed_gpu_parity.py --backend hip --vmaf $PWD/build-hip/tools/vmaf --feature cambi --no-timing
python3 scripts/dev/speed_gpu_parity.py --backend hip --vmaf $PWD/build-hip/tools/vmaf --no-timing

Device-free: python3 scripts/ci/run_meson_test.py -- -C build-hip --suite=fast (test_hip_cambi_device_math, test_hip_speed_device_math, test_hip_device_resident_contract).

Bug-status hygiene (ADR-0165)

• docs/state.md updated in this PR: T-HIP-CAMBI-HOST-RESIDUAL-2026-09-29 and T-HIP-SPEED-HOST-RESIDUAL-2026-09-29 closed with the measured gfx1036 results.

Netflix golden-data gate (ADR-0024)

• I did not modify any assertAlmostEqual(...) score in the Netflix golden Python tests.