ADR-0645: Thread integer ADM p-norm through SIMD callbacks¶

Status: Accepted
Date: 2026-05-20
Deciders: Lusoris maintainers
Tags: simd, feature-extractor, testing

Context¶

integer_adm exposes adm_p_norm so operators can tune the fixed-point contrast-measure finalisation exponent with the same option shape as float_adm. ADR-0623 deliberately landed the option before the full SIMD plumbing and left an open backlog row: the AVX2 / AVX-512 contrast-measure callbacks still accepted no p-norm parameter, so CPU SIMD dispatch continued to use the default 3.0 exponent even when the caller set adm_p_norm=2.0.

The ADM SIMD paths are bit-exact twins of the scalar fixed-point kernel. Changing their callback ABI is lower-risk than adding a parallel fallback path because the default 3.0 arithmetic stays the same and the option value is carried to the final p-norm exponent in one place.

Decision¶

Thread adm_p_norm through the adm_cm and i4_adm_cm callback signatures used by scalar, AVX2, and AVX-512 integer ADM, and compute the final powf(..., 1.0f / adm_p_norm) terms from that parameter. Keep the default 3.0 path unchanged for upstream/Netflix-compatible scoring and add an AVX2 unit test that proves both scale-0 and scale-1 CM callbacks respond to a non-default p-norm.

Alternatives considered¶

Option	Pros	Cons	Why not chosen
Leave SIMD fixed at `3.0` and document the limitation	No code risk in an old upstream-mirror hot path	User-visible option remains misleading on common x86 builds	This preserves a known correctness gap instead of closing it
Disable SIMD when `adm_p_norm != 3.0`	Simple semantic guarantee	Surprising performance cliff; splits scalar/SIMD behavior by option value	The SIMD callbacks can carry the parameter directly
Add separate p-norm-specific callback variants	Keeps old ABI untouched	More dispatch state and duplicated code for no benefit	The existing callback signature is internal and easier to extend

Consequences¶

Positive: adm_p_norm now reaches the x86 SIMD contrast-measure paths instead of being silently ignored.
Positive: The default adm_p_norm=3.0 path keeps the same expression shape, preserving Netflix-compatible default scoring.
Negative: The internal ADM callback ABI changes, so scalar and both x86 headers must move in lockstep on future rebases.
Neutral / follow-ups: GPU integer-ADM backends still expose their own smaller option tables and keep their existing fixed exponent until a GPU parity PR explicitly adds adm_p_norm there.

References¶

ADR-0623
docs/state.md — T-INTEGER-ADM-P-NORM-SIMD-GAP
Source: req — user asked to continue with the next backlog.