chore: release master

[github-actions]

🤖 I have created a release beep boop

4.0.0-lusoris.0

4.0.0-lusoris.0 (2026-04-20)

⚠ BREAKING CHANGES

• cli: %.6f default + unref skipped frames so Netflix golden gate passes (#55)
• cli: default CLI score precision changed from %.6f to %.17g. Parsers that string-equal-compare VMAF outputs will see additional digits. Pass --precision=legacy to restore old behavior.

Features

• ai: Full vmaf-train package — fr/nr/filter families + CLI + registry (91b5558)
• ai: Model validators + MCP eval + variance head + INT8 PTQ (4/5) (#8) (81f63a4)
• batch-a: Port four OPEN Netflix upstream PRs (ADR-0131/0132/0134/0135) (#72) (df622ea)
• ci: Nightly bisect-model-quality + sticky tracker (closes #4) (#41) (6cd4fb0)
• cli: --precision flag for IEEE-754 round-trip lossless scores (c989fbd)
• cli: Wire --tiny-model / --tiny-device / --no-reference flags (6381819)
• dev-llm: Ollama-backed review / commitmsg / docgen helpers (80ead16)
• DNN session runtime + SYCL list-devices + windows CI (1/5) (#5) (3cfa85b)
• dnn: Enforce VMAF_TINY_MODEL_DIR path jail on model load (#31) (22ba0e5)
• dnn: LPIPS-SqueezeNet FR extractor + ONNX + registry entry (#23) (0267dfb)
• dnn: Multi-input session API + ImageNet RGB tensor helper (#22) (d47481b)
• dnn: Ordered EP selection + fp16_io host-side cast (#34) (fad2b13)
• dnn: Runtime op-allowlist walk + tiny-model registry v0 (#21) (e98a220)
• dnn: Scaffold tiny-AI training, ONNX export, and DNN C seam (d122b72)
• dnn: Tests + ffmpeg patches + ci dnn job (1e5336d)
• libvmaf/feature: Port upstream ADM updates (Netflix 966be8d) (#44) (d06dd6c)
• libvmaf/feature: Port upstream motion updates (Netflix PR #1486) (#45) (9371a0a)
• libvmaf: Compile picture_pool unconditionally, drop VMAF_PICTURE_POOL gate (#32) (65460e3)
• libvmaf: DNN runtime — public dnn.h + ORT backend + sidecar loader (9b98594)
• mcp: Add vmaf-mcp server exposing VMAF to LLM tooling (0f233b8)
• ms-ssim: NEON bit-identical decimate (ADR-0125) (4531913)
• ms-ssim: Separable decimate + AVX2/AVX-512 SIMD (ADR-0125) (3283a61)
• ssimulacra2: Scalar port with libjxl FastGaussian IIR blur (#68) (410cf7b)
• sycl: Implement USM-backed picture pre-allocation pool (#33) (3f9a134)
• sycl: Implement vmaf_sycl_import_d3d11_surface (closes #27) (#35) (8ea83f7)

Bug Fixes

• ai: Bump torch >=2.8 + lightning >=2.5 for CVE fixes (#11) (98da94d)
• ai: Regenerate bisect cache with pinned pandas 2.3.3 (#42) (cdb9b9e)
• ci: Coverage gate lcov→gcovr + ORT + lint upstream tests in-tree (#46) (652aa70)
• ci: Drop dead sycl trigger + consolidate windows.yml into libvmaf.yml (#50) (e01314e)
• ci: Exclude binary model payloads from mixed-line-ending hook (#36) (0f61607)
• ci: Post bisect sticky comment via stdin instead of -f @file (#43) (f453d1b)
• ci: Silence Coverage Gate annotations (upload-artifact v7 + gcovr filter) (#54) (fef9a86)
• ci: Skip pytest doctest collection of vmaf/resource/ data files (#51) (d23005c)
• cli: %.6f default + unref skipped frames so Netflix golden gate passes (#55) (aa08d84)
• cuda: Guard vmaf_picture_ref against NULL src->ref on device-only path (#62) (661a8ac)
• cuda: Unconditional sm_86/sm_89 cubin coverage + actionable init-failure logging (#60) (d3b6fad)
• gitleaks: Allowlist manifests/README.md false positive (#59) (5786a70)
• libvmaf/feature: Free VIF init base pointer on fail path (#47) (d8ab927)
• libvmaf: Gate -fsycl link arg on icpx CXX, allow gcc/clang host linker (#52) (4a8322d)
• SIMD bit-identical reductions + CI fixes (#18) (f082cfd)

Documentation

• Add community hygiene files (SECURITY, CoC, CODEOWNERS, templates) (c28dd78)
• Add per-distro install guides + BENCHMARKS + hygiene files (fe8c744)
• Add SYCL backend, SIMD, and GPU documentation (3a0ee4f)
• Add SYCL bundling guide for self-contained deployment (5fe843f)
• adr: Supersede 0025/0028/0036 with paraphrased bodies (#37) (4955dde)
• Consolidate + reorganise documentation tree (#12) (4bbd573)
• Migrate from Sphinx+Doxygen to MkDocs Material (#17) (40fe6f1)
• Project-wide doc-substance sweep (ADR-0100 batches 1-4) (#25) (4f3f992)
• readme: Rewrite for the Lusoris fork, preserve upstream credit (e68befa)
• tiny-ai: User docs + README row + LFS + release-please subpackages (66bd6bd)

Tests

• Add SYCL unit tests, GPU validation scores, and benchmark scripts (e704022)

Build System

• Add SYCL backend to meson build and fix pkgconfig for static linking (c66e478)
• CUDA 13 + oneAPI 2025.3 + clang-format 22 + black 26 (3/5) (#7) (a7be84c)
• Dev infra + scaffolding templates + MCP Docker (e1482c8)
• make: Add lint/format/sec/sbom/netflix-golden targets (a077407)
• Multi-distro dev setup scripts + extra Dockerfiles (d1cfc97)

CI / Infrastructure

• Add explicit permissions to workflow files (e6af077)
• Add lint configs, pre-commit, and GitHub Actions gates (9d72f75)
• ai: Add DNN-enabled matrix legs (gcc + clang + macOS) (#56) (75ad729)
• Coverage + assertion-density gates + VPL host upload (replaces #6) (#13) (ea7b524)
• Fill lint/CI gaps (nightly, docs, editorconfig, iwyu, gitleaks, codeql) (75a581b)
• lint: Scan push-event delta, not full tree (ADR-0133) (#71) (af65ada)
• Release-please automation + dependabot groupings (8573524)
• rule-enforcement: Automate ADR-0100/0105/0106/0108 checks (#63) (1aa45ec)
• rule-enforcement: Strip markdown emphasis before deliverables grep (ADR-0136) (#73) (95bea61)
• Update workflows for SYCL, CUDA, and static builds (eaad704)

Miscellaneous

• adr: Adopt ADR-0108 deep-dive deliverables rule + backfill rebase notes (#39) (d60e63a)
• adr: Migrate to Nygard one-file-per-decision + golusoris-alignment sweep (#24) (8e3cd22)
• ci: Finish Node 24 bump — scorecard artifact SHA + nightly-bisect setup-python (#49) (3eb9af7)
• ci: Rename workflows + Title Case display names (ADR-0116) (#53) (f4379c8)
• license: Bump Netflix copyright from 2016-2020 to 2016-2026 (c159761)
• license: Correct fork copyright year to 2026 (0e98c94)
• license: Re-attribute SYCL files to Lusoris and Claude (a185f8e)
• lint: Clang-tidy upstream cleanup rounds 2-4 (#2) (722d21f)
• Post-merge cleanup — CI fix + lint + supply-chain + scorecard + dependabot (#14) (798db39)
• repo: Add AI-agent scaffolding and engineering-principles docs (b799db5)
• upstream: Record ours-merge of Netflix 966be8d (already ported in d06dd6c) (fddc5ca)
• upstream: Record ours-merge of Netflix 966be8d (bookkeeping) (27ce439)
• vscode: Clangd-first workspace settings + debug launch configs (04d18f1)

Ports from upstream Netflix/vmaf

• libvmaf: Thread-local locale handling from Netflix/vmaf#1430 (ADR-0137) (#74) (e0e78db)

SYCL Backend

• Implement oneAPI/SYCL compute backend for ADM, VIF, and Motion (e899082)

CUDA Backend

• Double-buffer GPU frames and optimize kernels (ea9a9d4)
• Eliminate ADM decouple intermediate buffers Inline the decouple computation into the CSF and CM kernels, eliminating 6 intermediate buffers (decouple_r, decouple_a, csf_a for both int16 scale-0 and int32 scales 1-3 paths). CSF kernels now read ref/dis DWT2 directly from buffer, compute decouple inline, and write only csf_f. CM kernels inline both decouple_r and csf_a computation per-pixel, reading only csf_f from buffer (needed for 3x3 stencil). Shared device functions extracted to adm_decouple_inline.cuh. The adm_decouple.cu source is no longer compiled. Reduces GPU memory allocation from 30 to ~16.5 buf_sz_one units (~107 MB savings at 4K resolution). (787e338)
• VIF rd_stride optimization - reduce read buffer allocation Add rd_stride field to VifBufferCuda for texture-aligned read buffer access. Compute rd_stride with tex_alignment padding, reducing allocation from full frame_size to rd_size (rd_stride * ceil(h/2)). Update filter1d.cu kernels to use rd_stride for rd buffer indexing. (0d2a196)

SIMD

• Add ARM NEON implementations for feature extractors (e020421)
• Add x86 AVX2/AVX-512 implementations for feature extractors (81fcd42)
• Use double-precision accumulation in float ADM reductions Float ADM sum_cube and csf_den_scale functions accumulated cubed values in float32, causing ~8e-5 drift between scalar and SIMD paths due to different accumulation order. Fix by computing val^3 in float SIMD, converting to double via _mm256_cvtps_pd / _mm512_cvtps_pd before accumulating, and using double for outer accumulators in all paths (scalar, AVX2, AVX512). Update test expectations accordingly. (24c88a3)

0.2.0

0.2.0 (2026-04-20)

Features

• ai: Full vmaf-train package — fr/nr/filter families + CLI + registry (91b5558)
• ai: Model validators + MCP eval + variance head + INT8 PTQ (4/5) (#8) (81f63a4)
• ci: Nightly bisect-model-quality + sticky tracker (closes #4) (#41) (6cd4fb0)
• DNN session runtime + SYCL list-devices + windows CI (1/5) (#5) (3cfa85b)
• dnn: LPIPS-SqueezeNet FR extractor + ONNX + registry entry (#23) (0267dfb)
• dnn: Scaffold tiny-AI training, ONNX export, and DNN C seam (d122b72)

Bug Fixes

• ai: Bump torch >=2.8 + lightning >=2.5 for CVE fixes (#11) (98da94d)
• ai: Regenerate bisect cache with pinned pandas 2.3.3 (#42) (cdb9b9e)
• SIMD bit-identical reductions + CI fixes (#18) (f082cfd)

Build System

• CUDA 13 + oneAPI 2025.3 + clang-format 22 + black 26 (3/5) (#7) (a7be84c)

Miscellaneous

• adr: Migrate to Nygard one-file-per-decision + golusoris-alignment sweep (#24) (8e3cd22)
• lint: Clang-tidy upstream cleanup rounds 2-4 (#2) (722d21f)
• Post-merge cleanup — CI fix + lint + supply-chain + scorecard + dependabot (#14) (798db39)
• upstream: Record ours-merge of Netflix 966be8d (already ported in d06dd6c) (fddc5ca)
• upstream: Record ours-merge of Netflix 966be8d (bookkeeping) (27ce439)

0.2.0

0.2.0 (2026-04-20)

Features

• ai: Model validators + MCP eval + variance head + INT8 PTQ (4/5) (#8) (81f63a4)
• dev-llm: Ollama-backed review / commitmsg / docgen helpers (80ead16)

Miscellaneous

• Post-merge cleanup — CI fix + lint + supply-chain + scorecard + dependabot (#14) (798db39)
• upstream: Record ours-merge of Netflix 966be8d (already ported in d06dd6c) (fddc5ca)
• upstream: Record ours-merge of Netflix 966be8d (bookkeeping) (27ce439)

0.2.0

0.2.0 (2026-04-20)

Features

• ai: Model validators + MCP eval + variance head + INT8 PTQ (4/5) (#8) (81f63a4)
• mcp: Add vmaf-mcp server exposing VMAF to LLM tooling (0f233b8)

Miscellaneous

• Post-merge cleanup — CI fix + lint + supply-chain + scorecard + dependabot (#14) (798db39)
• upstream: Record ours-merge of Netflix 966be8d (already ported in d06dd6c) (fddc5ca)
• upstream: Record ours-merge of Netflix 966be8d (bookkeeping) (27ce439)

---

This PR was generated with Release Please. See documentation.