Vapoursynth-HIP

An easy to use vapoursynth plugin to compute SSIMU2 (SSIMULACRA2) from libjxl or Butteraugli from libjxl on GPU with the exception that it uses a real gaussian blur and not a recursive gaussian blur.

Butteraugli Source Code : https://github.com/libjxl/libjxl/tree/main/lib/jxl/butteraugli Ssimulacra2 Source Code : https://github.com/cloudinary/ssimulacra2

Usage:

To retrieve ssimu2 score between two video Nodes

result = clip1.vship.SSIMULACRA2(clip2)

res = [fr.props["_SSIMULACRA2"] for fr in result.frames()]

To retrieve Butteraugli score

result = clip1.vship.BUTTERAUGLI(clip2)

res = [[fr.props["_BUTTERAUGLI_2Norm"], fr.props["_BUTTERAUGLI_3Norm"], fr.props["_BUTTERAUGLI_INFNorm"]] for fr in result.frames()]

you can check "simpleExample.vpy" inside VshipUsageScript to get more details about how to use Vship

If you encounter VRAM issue, set a lower vs.core.num_threads (4 is usually the good compromise for speed)

the exact vram formula per active vsthread is: ssimu2: 24 (plane buffer) * 4 (size of float) * width * height * 4/3

butteraugli: 34 (plane buffer) * 4 (size of float) * width * height

to build:

Warning: it is compiled for the specific gpu used to compile by default

(same for windows and linux) you need

• make
• hipcc or nvcc
• vapoursynth

for nvidia cards: make buildcuda

for amd cards: make build

to install: either use the dll or: make install

Performance on my laptop on some clip:

special credits to dnjulek with the ZIG vapoursynth implementation