Research Digest: Netflix Upstream Feature Additions — CUDA Twin Gap Audit (2026-05-18)¶

Task: Identify which C-side sub-features consumed by the Netflix HDR VMAF model are present in the fork's CPU code but lack a CUDA twin, and determine the correct porting strategy for each.

Scope: Five candidate features reported by the user — aim, adm3, motion3, chroma_from_luma, cambi_eotf/effective_eotf.

Findings¶

1. aim — Anchored Integer Motion (adm sub-feature)¶

• CPU location: core/src/feature/float_adm.c and core/src/feature/adm.c.
• Mechanism: The compute_adm() function in adm.c runs a second adm_cm pass per scale with the CSF of decouple_r (not decouple_a) and with noise_weight = 0. The result is accumulated into aim_num / aim_den and finally emitted as VMAF_feature_aim_score from float_adm.c's extract().
• CUDA status before this PR: Missing. float_adm_cuda.c did not emit VMAF_feature_aim_score. --backend cuda silently fell back to CPU when the HDR model requested this feature.
• Fix: Two new kernel stages (2b and 3b) added to float_adm_score.cu. See ADR-0574.

2. adm3 — ADM Version 3 (adm sub-feature)¶

• CPU location: core/src/feature/float_adm.c.
• Mechanism: Derived in extract() from score (adm2) and score_aim (AIM) as either a harmonic mean (when adm_adm3_apply_hm=true) or a linear blend weighted by adm_dlm_weight. Emitted as VMAF_feature_adm3_score.
• CUDA status before this PR: Missing (same root cause as aim — no AIM accumulation in float_adm_cuda.c).
• Fix: Same two new kernel stages; host-side collect_fex_cuda derives and emits adm3_score from the AIM accumulation. ADR-0574.

3. motion3 — Motion Version 3 (motion sub-feature)¶

• CPU location: core/src/feature/integer_motion.c.
• Mechanism: Post-processing of motion2 via a moving average of blended scores; options motion_blend_factor, motion_blend_offset, motion_fps_weight, motion_moving_average, motion_max_val. Emitted as VMAF_integer_feature_motion3_score.
• CUDA status: Already implemented in integer_motion_cuda.c (motion3_postprocess_cuda(), ADR-0219). No porting needed.

4. chroma_from_luma — Chroma Prediction Correction¶

• Investigation: Searched for chroma_from_luma across core/src/feature/ and core/include/. Found references only in core/src/model.h and core/src/model.c as a field of the model's predictor configuration, not as a feature extractor.
• CUDA status: Not applicable — this is a model-level predictor attribute, not a GPU kernel. No feature extractor symbol vmaf_fex_*_chroma_from_luma exists. No porting needed.

5. cambi_eotf / effective_eotf — PQ/HLG EOTF Support in CAMBI¶

• CPU location: core/src/feature/cambi.c.
• Mechanism: eotf option selects PQ/HLG inverse EOTF pre-processing applied to the input luma plane before CAMBI computation. effective_eotf is the resolved EOTF name stored in feature output.
• CUDA status: Already implemented in integer_cambi_cuda.c with the same eotf and cambi_eotf options (HAVE_CUDA block). No porting needed.

Summary Table¶

Implementation approach chosen for aim / adm3¶

The AIM CM computation reuses the same DWT band buffers produced by stages 0 and 1. The key difference from the adm2 CM pass (stage 3) is:

• The CSF threshold is derived from decouple_r = k * o (the remodulated reference component), not from decouple_a = t - r (the anomaly).
• noise_weight = 0 — no noise constant is added to the CM numerator.

Rather than storing decouple_r in a scratch buffer (which would cost an extra device-side buffer the same size as csf_a), the two new kernels recompute k and r_val inline from the already-buffered ref_band / dis_band arrays. This trades a few arithmetic ops per pixel for zero extra bandwidth, which is favourable at all tested resolutions.

The accumulator slot layout extension (6→9) adds 50 % to the pinned D2H copy per scale but the absolute size is a few KB — negligible compared to the kernel execution time.