perf(cuda): motion — 8-frame SAD batching to reduce per-launch overhead (ADR-0845)

[lusoris]

Summary

• Replaces the per-frame cuStreamSynchronize pattern in integer_motion_cuda.c with an 8-frame batch fence (MOTION_BATCH_DEPTH=8), reducing synchronization calls from once-per-frame to once-per-8-frames.
• Research-0760 (2026-05-29) shows the CUDA motion kernel is dispatch-bottlenecked at all resolutions below 4K: GPU busy fraction <0.1% at 576p, CUDA/CPU ratio 0.22x. Kernel execution (7 µs at 576p) is irrelevant — the bottleneck is ~12.7 ms/frame of driver round-trip overhead.
• Expected improvement: 576p from ~79 fps to ~800 fps (CUDA/CPU from 0.22x to >2x). 4K performance unchanged.

Design

• sad[MOTION_BATCH_DEPTH] — ring of 8 independent device SAD buffers (one per frame slot).
• submit() zeros slot index % 8, launches kernel, chains event to s->str. No DtoH.
• Batch-boundary collect() (every 8th frame): cuStreamSynchronize(s->str), queue 8 DtoH copies, sync again, emit all 8 scores via emit_batch_scores().
• flush() handles the final partial batch.
• emit_batch_scores() saves/restores s->frame_index per frame to preserve ADR-0219 motion3 moving-average semantics.
• motion_cuda is removed from the drain_batch engine optimization (ADR-0242) — it no longer registers finished events.

Correctness gate

ADR-0214 places=4 cross-backend correctness check required before promoting to READY. The motion3_postprocess_cuda frame_index save/restore pattern is designed to produce bit-identical scores to the pre-patch path.

Test plan

• Build with -Denable_cuda=true in vmaf-dev-mcp:cuda13.3 container
• Correctness: vmaf --feature motion --backend cuda on Netflix 576x324 fixture, compare to CPU at places=4
• A/B fps measurement at 576p, 1080p, 4K vs baseline from Research-0760
• Full VMAF pipeline correctness (all features, not just motion)
• Confirm MOTION_BATCH_DEPTH=1 degrades back to per-frame behavior

Reproducer

# Build
docker exec vmaf-dev-mcp bash -c "cd /workspace/core && ninja -C build"

# Correctness (576p, places=4)
docker exec vmaf-dev-mcp /build/vmaf/core/build/tools/vmaf \
  --reference /yuv/src01_hrc00_576x324.yuv \
  --distorted /yuv/src01_hrc01_576x324.yuv \
  --width 576 --height 324 --pixel_format 420 --bitdepth 8 \
  --feature motion --backend cuda --output /tmp/cuda_motion.json

# Compare to CPU reference
docker exec vmaf-dev-mcp /build/vmaf/core/build/tools/vmaf \
  --reference /yuv/src01_hrc00_576x324.yuv \
  --distorted /yuv/src01_hrc01_576x324.yuv \
  --width 576 --height 324 --pixel_format 420 --bitdepth 8 \
  --feature motion --output /tmp/cpu_motion.json

Checklist

• ADR filed: ADR-0845
• Research digest — no digest needed: design analysis in ADR-0845; perf baseline is Research-0760 (pre-existing on scaffold branch)
• Decision matrix — ADR-0845 ## Alternatives considered complete
• AGENTS.md invariant note — Motion SAD batch fencing: MOTION_BATCH_DEPTH invariants (ADR-0845)
• Reproducer / smoke-test command — see Reproducer section above
• CHANGELOG fragment — changelog.d/perf/cuda-motion-batch-launch-adr0845.md
• Rebase note — docs/rebase-notes.md entry added
• docs/state.md row added
• docs/backends/cuda/overview.md drain_batch participation list updated
• DRAFT — A/B measurement + places=4 correctness check pending

no rebase impact: this PR only touches integer_motion_cuda.c (fork-local, no upstream counterpart), AGENTS.md, and docs. No upstream Netflix/vmaf C source is modified.

🤖 Generated with Claude Code

[lusoris]

Superseded by master after 143-PR merge marathon 2026-05-31; diff-extract empty.