v0.3.0 — GPU fast paths + v3 builds
GPU decode fast paths and CPU-tuned builds. Batteries included, as usual.
- CUDA fast paths for float output: f32 now takes the direct GPU→frame copy (+44% @ 4.6K), f16 decodes natively on GPU pipelines instead of f32 + CPU narrowing (+75% @ 4.6K, RTX 3080).
- Frame metadata is now harvested off the serial decode path (~+6% with
allmetaprops). - New
-v3zips (x86-64-v3: AVX2/F16C, CPUs from ~2013) with noticeably faster pixel-copy loops — prefer them on any recent machine. The plain x86_64 zips stay SSE2-compatible.
Each zip is drop-in: plugin + Blackmagic RAW runtime in the blackmagic_*_deps folder next to it. Verify against SHA256SUMS.txt.