Release Release v0.8.1 · imazen/fast-ssim2

v0.8.1 — stripwise SSIMULACRA2 + perf

compute_ssimulacra2_strip and Ssimulacra2Reference::compare_strip — strip-wise SSIMULACRA2 with bounded peak memory for very large images. Processes the image in horizontal strips (default 32-aligned, halo of 96 rows for IIR Gaussian convergence) and accumulates per-scale SSIM and edge-diff sums across strips. Scores match the full-image path to within ~1e-5 on the 0..100 scale at 1024² and larger; identical-image inputs round-trip to 100 in both modes. Bounds dist-side peak memory to ~24 * strip_h * width * 4 B instead of ~24 * height * width * 4 B.
Ssimulacra2StripConfig — configurable halo size and underlying SIMD backend for strip processing. Defaults are tuned for atomic-tolerance parity with the full path.
HALO_ROWS_DEFAULT (96) and MIN_STRIP_HEIGHT (8) public constants.
Hidden Ssimulacra2Reference::scale_planes accessor returning a ScalePlanesView; required by the strip walker.

CompareContext and Ssimulacra2Reference::compare_with(&mut ctx, distorted) — zero-allocation batch-comparison API. Pair with reference.compare_context(); subsequent calls reuse the working buffers (mul, mu2, sigma2_sq, sigma12, img2_planar, blur state) instead of allocating ~13 image-sized Vec<f32> planes per call. Measured 1.10–1.25× faster than compare() on the precompute benchmark at 256x256 / 512x512 / 1024x1024 / 1920x1080.
LinearRgbImage::try_new fallible constructor returning LinearRgbImageError for invalid dimensions or data length.
Ssimulacra2Error::ImageTooLarge variant and public MAX_IMAGE_PIXELS constant (16384*16384).

Skip per-channel SSIM and edge-difference work whose final-score weight is zero. Bit-identical to the prior path (the dropped contributions multiplied by zero downstream); reference-parity test passes across the C++ corpus including 64x64 cases where scales_n < NUM_SCALES makes score()'s linear WEIGHT walk shift in the layout. Lossless variant of Technique 2 from Kanetaka et al. IWAIT 2026, DOI 10.1117/12.3100969.
Hoist the 6 IIR-state vectors used by the SIMD vertical blur pass out of the per-call inner function and onto SimdGaussian, eliminating ~180 small Vec<f32> allocations per ssim2 frame.

LinearRgbImage::new now validates dimensions and data length at runtime (was debug_assert_eq! only) so release-mode misuse no longer constructs malformed images that panic deep in From<LinearRgbImage> for yuvxyb::LinearRgb.
SimdGaussian::new no longer eagerly allocates max_width * 4096 floats; the temp buffer grows on demand. Also guards against usize overflow on 32-bit targets when width * height would wrap.

Full Changelog: v0.8.0...v0.8.1