Changes since last dev build 20260518 3.7.5.r4604 (pre 3.7.6)
Full change list: https://avisynthplus.readthedocs.io/en/latest/avisynthdoc/changelist376.html
20260622 3.7.5.r4626 (pre 3.7.6)
- Fix (20260518-dev) regression: Subtitle raised an error on grayscale (Y-only) clips
(regression from recent gdi/placement refactoring). - Make Overlay's internal 444 conversion to give bit-identical results for C and SIMD, add AVX2 remove ISSE path
- Make Overlay's "multiply" to give bit-identical results for C and SIMD
20260620 3.7.5.rXXXX (pre 3.7.6)
- Fix AviSynth#485: AVX2 horizontal resampler for PlanarRGB(A) to packed RGB32/RGB64
missed processing the right-edge tail pixels in each row.
20260611 3.7.5.rXXXX (pre 3.7.6)
- Resampler AVX512: cleanup — remove obsolete/duplicate kernel variants superseded
by the pretransposed-coefficient approach introduced on 20260610.
20260610 3.7.5.rXXXX (pre 3.7.6)
- Resizers: AVX512 H-resizer precomputed coefficient table (contribution: DTL2020).
resize_prepare_coeffs_AVX512_H builds a transposed/unpacked coefficient table
(pixel_coefficient_AVX512_H) before the row loop, so permute indices and
coefficients are no longer re-derived per-row inside the kernel. - Resizers: new AVX512 H-resizer "_pretransposed_coeffs" kernels for uint8
(mpz_ks4/ks8/ks16, mpz_2s32_ks64, 2s32_ks8) and uint16
(mp_ks8, 2s32_ks4/ks8, 4s16_ks8, 4s16_ks48); each templated on UseVNNI so
one implementation covers both VNNI/VBMI ("fast") and BASE AVX512 paths.
2s32_ks8 uint8 kernel reworked into this common VNNI/BASE-friendly form.
New kernels wired into the AVX512 H-resizer dispatcher. - Resizers: fix permutex2var_epi8 simulation — rework and correct castsi
zero-extension usage; fix coefficient table layout in resize_prepare_coeffs_AVX512_H. - Resizers: AVX512/AVX2 float H-resizer dispatch updated for the pretransposed-coeff
path; dispatching logic and comments revised.
20260529 3.7.5.rXXXX (pre 3.7.6)
- Fix AviSynth#482: YV411 (4:1:1) to higher-bit-depth YUV conversion: perform chroma
subsampling first, then bit-depth conversion, instead of the reverse order that
produced incorrect output for all non-8-bit targets. - Layer: fix regression in PlanarRGB(A) "add" mode — opacity was inadvertently
ignored after the recent packed/planar RGB refactoring.