Skip to content

ADR-0467: SSIMULACRA2 AVX-512 + NEON IIR Blur / picture_to_linear_rgb ULP Audit — Clean Close

  • Status: Accepted
  • Date: 2026-05-16
  • Deciders: lusoris
  • Tags: simd, ssimulacra2, audit

Context

BACKLOG item T3-9(b) (formerly T3-10) required a cross-host ULP audit of the SSIMULACRA2 AVX-512 and NEON IIR blur (ssimulacra2_blur_plane_avx512 / ssimulacra2_blur_plane_neon) and the picture_to_linear_rgb kernels (ssimulacra2_picture_to_linear_rgb_avx512 / _neon) after their merge in PRs #98–#100 (ADR-0161/0162/0163).

The concern was "post-merge cross-host snapshot drift": whether the SIMD paths produce non-zero ULP deltas vs the scalar reference that compound across scales and hosts, pushing the pooled SSIMULACRA2 score outside the places=4 (1e-4) gate.

Cross-host drift was the historical problem that motivated the deterministic vmaf_ss2_cbrtf (Newton–Raphson) and vmaf_ss2_srgb_eotf (LUT) helpers in ssimulacra2_math.h (ADR-0164). This audit verifies that both AVX-512 and NEON paths, which call those same helpers per-lane, inherit the host-independence guarantee at the kernel level.

Decision

The audit found zero ULP drift. No code change is required. T3-9(b) is closed as "clean — within contract."

Key evidence:

  1. C unit test (test_ssimulacra2_simd, 13 sub-tests) — all 13 sub-tests pass with memcmp byte-exactness on the development host (Zen 5 + AVX-512 BW/CD/DQ/VL/IFMA/VBMI). The dispatcher selects the AVX-512 path at runtime (VMAF_X86_CPU_FLAG_AVX512 set), confirming that every kernel exercised in the test is the AVX-512 variant, not the AVX2 fallback. Covered kernels: multiply_3plane, linear_rgb_to_xyb, downsample_2x2, ssim_map, edge_diff_map, blur_plane (IIR horizontal + vertical), and all five picture_to_linear_rgb format variants (420/8, 420/10, 444/8, 444/10, 422/8).

  2. Python snapshot gate (ssimulacra2_test.py, 2 tests) — both the primary 576x324 fixture (48 frames) and the 160x90 tail fixture pass at places=4 tolerance. Actual scores (x86-64): mean 80.551211, frame0 91.695977, frame47 77.992897 — all matching hardcoded baselines.

  3. IIR boundary analysis — hblur_16rows_avx512 and hblur_4rows_neon both initialise prev1_* / prev2_* vectors to zero, matching the scalar reference. The gather-index construction for inactive lanes in AVX-512 (lines 376–380 of ssimulacra2_avx512.c) points those lanes at row 0 offsets, but their scatter outputs are gated by for (unsigned i = 0; i < row_count; i++) — no output pollution.

  4. Transcendental isolation — both paths call vmaf_ss2_cbrtf and vmaf_ss2_srgb_eotf per-lane via scalar fallthrough in cbrtf_lane_avx512 / cbrtf_lane_neon and srgb_to_linear_lane_avx512 / _neon. These are the ADR-0164 deterministic helpers (no libc dependency), so host divergence at this callsite is structurally impossible.

Alternatives considered

Option Pros Cons Why not chosen
Accept as-is (clean close) No code churn; tests confirm correctness None Chosen: this is the correct outcome
Add per-lane double-accumulator crosscheck Would catch future regressions earlier Redundant with existing memcmp byte-exact gate Not needed — existing tests already enforce byte-exactness
Widen to cross-host CI (arm64 runner) Would close the "cross-host" concern definitively arm64 runner is already exercised by the existing CI matrix; NEON path tests run on that leg Already covered

Consequences

  • Positive: T3-9(b) is resolved; no residual ULP debt in the SSIMULACRA2 AVX-512/NEON blur and picture_to_linear_rgb kernels.
  • Positive: The audit confirmed that the ssimulacra2_math.h deterministic helpers fully insulate the SIMD paths from cross-host libc divergence by construction.
  • Neutral: No snapshot regeneration required (no numerical change).
  • Follow-up: T3-9(c) (iqa_convolve AVX-512 ceiling check) remains open per the BACKLOG; it is independent of this audit.

References

  • ADR-0161 (SSIMULACRA2 AVX2 phase 1 — pointwise kernels)
  • ADR-0162 (SSIMULACRA2 AVX2 phase 2 — IIR blur)
  • ADR-0163 (SSIMULACRA2 AVX2 phase 3 — picture_to_linear_rgb)
  • ADR-0164 (deterministic cbrtf + sRGB EOTF LUT)
  • ADR-0138/0139 (bit-exactness invariants, NOLINTNEXTLINE policy)
  • ADR-0214 (cross-backend ULP gate, places=4)
  • BACKLOG T3-9(b) (this item)
  • Source: per user direction to audit T3-9(b) post-merge ULP drift.