Release Release v0.8.2 · imazen/fast-ssim2

Sub-8px inputs are reflect(mirror)-padded up to the metric's 8px pyramid floor instead of returning InvalidImageSize: compute_ssimulacra2 / compute_ssimulacra2_with_config now score images down to 1×1 (identical pairs still score 100) (480df7e)
Ssimulacra2Reference applies the same sub-8px reflect-padding, so the batch path accepts the same inputs as the one-shot path and produces identical scores; width()/height() report the caller-supplied (pre-padding) dimensions and mismatched pre-padding dimensions are still rejected. Strip APIs intentionally keep the ≥8×8 requirement (54df468)
Experimental hdr-pu feature: compute_ssimulacra2_pu_nits scores HDR content using the PU21 (banding_glare) encoding in place of the cube-root opsin nonlinearity; input is absolute-luminance linear RGB in cd/m². Validated on UPIQ HDR (380 pairs, SROCC 0.7044; see imazen/zenmetrics#25) (35f198a)
CI now lints (clippy -D warnings) and tests the non-default hdr-pu feature on the Linux clippy job, so the feature-gated path is exercised on every push (f987fc1)

Vectorised the recursive-Gaussian blur horizontal pass across rows (8 rows per SIMD lane group, one row per lane), mirroring the across-columns trick the vertical pass already uses. Bit-identical output to the scalar path — the SIMD IIR replicates the scalar op order exactly. Measured on Neoverse-N1 (where the scalar per-row recurrence was ~50% of the blur kernel): blur-only +9–12%, full SSIMULACRA2 +3.6–9.5% (ssimulacra2_1920x1080 540.9 → 503.0 ms). x86 (AVX2) unaffected — blur was already well-vectorised there (87e06d5)

What's Changed

perf(fast-ssim2): vectorise blur horizontal pass across rows (+5.8-12.4% on Neoverse-N1) by @lilith in #3

Full Changelog: v0.8.1...v0.8.2