libvmaf-fast 3.2.0-fast.1

The first release of libvmaf-fast, an unofficial fork of Netflix's libvmaf. Not affiliated with or endorsed by Netflix.

Built from fast at 3ddc3d84: libvmaf at cea2b4d8 (3.2.0) with 12 upstream pull requests merged, and the Vulkan engine under fast/. See fast/README.md.

In the archive (Windows x64)

• libvmaf/libvmaf.dll: libvmaf with CUDA, the public C API exported, built-in models, static C runtime. Reports version 3ddc3d84.
• vmaf_vulkan/vmaf_vulkan.dll: VMAF's features on any GPU with Vulkan. VIF, ADM and motion for VMAF v0.6.1 and NEG, bit-identical to libvmaf's CUDA code; ADM3 and motion3 for VMAF v1, bit-identical to libvmaf's CPU code.
• Licences for each, BUILD.txt, and SHA256SUMS of every file.

Both need only Windows and a GPU driver. Both builds are reproducible from the tagged commit with the scripts in fast/scripts.

Checked

On an RTX 5090 and an Intel integrated GPU, with these two DLLs:

• compare_vmaf_vulkan.py --matrix (45 cases) and compare_vmaf_v1.py --matrix (71 cases): all identical.
• diagnose_vmaf_vulkan.py: every sum and buffer matches the reference.
• 48 frames of real 4K 10-bit video: VMAF 85.104145, NEG 82.830715, VMAF v1 86.488461, identical to libvmaf.

The AMD Radeon 780M results in the README are from the same engine source, tested in VideoMetricsLab.

Speed (4K 10-bit, frames in memory)

• VMAF + NEG: 312 fps with Vulkan on an RTX 5090, against 21 fps for libvmaf on a 24-core CPU.
• VMAF v1: 124 fps with the RTX 5090 and the CPU together, against 51-67 fps on the CPU alone.

Not in this release

Windows x64 only. The Vulkan engine is a DLL beside libvmaf, not one of its feature extractors, so FFmpeg's libvmaf filter cannot use it yet.

Thanks

Most of the CUDA fixes are the work of StormBytePP, jmsether, lusoris and BardieJoensen, in their open pull requests to Netflix/vmaf (listed in the README).