docs(research): cross-backend 4K baseline + PR #79 adm_cm A/B at 4K (Research-0751)

[lusoris]

Summary

• Establishes the first measured 4K (3840×2160) CUDA throughput baseline on RTX 4090 (BigBuckBunny pair, 24 frames, vmaf_bench)
• A/B tests PR perf(cuda): ms_ssim_decimate smem tiling + adm_cm register reduction (ADR-0744) #79 adm_cm_line_kernel_8 __launch_bounds__(128,8) at 4K using the same container-patch methodology as Research-0749
• Extends the PR perf(cuda): VIF filter1d horizontal — __launch_bounds__ + __ldg, ADR-0743 #76 filter1d wave-count analysis from 576p (0.84 waves) and 1080p through to 4K saturation

Key findings

Feature	CUDA fps (4K, 24f median)	vs CPU
vif	147 fps	7.0×
adm	161 fps	2.3×
motion	176 fps	0.6× (launch overhead)

filter1d_8_horizontal (PR #76 target): 32,400 CTAs / 128 SMs = 253 waves at 4K, 69.7% active warps. The 0.84-wave launch-width limit from 576p is completely gone. PR #76 is fully expressed at production 4K.

adm_cm_line_kernel_8 (PR #79 __launch_bounds__): −0.3% kernel duration at 4K (noise) vs −9.3% at 1080p. The register-bound regime that the optimization targets is 8–32 waves; at 4K (32.2 waves) the scheduler is wave-saturated. The optimization remains beneficial and zero-cost at 4K — recommendation unchanged: ship the __launch_bounds__ change.

ms_ssim_decimate (scale 0 at 4K): 88.1% active warps, 126 waves. Smem-tiling revert from Research-0749 is confirmed correct — the kernel is already L1-resident at all scales (>99.5% hit rate).

Deliverables

• Research digest: docs/research/0751-cross-backend-4k-baseline-and-pr79-adm-cm-4k-measure.md
• Changelog fragment: changelog.d/changed/cross-backend-4k-baseline.md
• Rebase notes: sentinel added (no rebase impact)
• state.md updated
• research/README.md: 0751 entry added + pre-existing merge-conflict marker resolved
• no ADR needed: measurement digest, no architectural decision
• no AGENTS.md update needed: no rebase-sensitive invariants
• no per-surface docs needed: no user-discoverable surface changed

Reproducer

docker run --rm --gpus all --entrypoint bash \
  -v /path/to/.corpus:/corpus:ro \
  vmaf-dev-mcp:cuda13.3 -c '
    mkdir -p /tmp/vmaf_test && \
    ln -sf /corpus/netflix/ref/BigBuckBunny_25fps.yuv /tmp/vmaf_test/ref_3840x2160.yuv && \
    ln -sf /corpus/netflix/dis/BigBuckBunny_85_1080_3800.yuv /tmp/vmaf_test/dis_3840x2160.yuv && \
    /build/vmaf/core/build/tools/vmaf_bench --resolution 3840x2160 --frames 24
  '

Full ncu reproducer in Research-0751 §7.

🤖 Generated with Claude Code