Deprecated

This repository is deprecated. Please use vapoursynth-api4-wheels.

---

vs-dfttest2-api4

DFTTest re-implementation for VapourSynth API4.

This fork packages the CPU backend and, for CUDA-tagged builds, the NVRTC and cuFFT/CUDA backends. The HIP, HIPRTC, and GCC backend sources are kept in the repository, but they are not part of the default Windows package.

Installation

The Windows and Linux x86_64 VCS install paths are release-backed. Pick the package variant by installing from one of the repository tags:

pip install "vapoursynth-dfttest2 @ git+https://github.com/RyougiKukoc/vs-dfttest2-api4.git@cpu"
pip install "vapoursynth-dfttest2 @ git+https://github.com/RyougiKukoc/vs-dfttest2-api4.git@cu121"
pip install "vapoursynth-dfttest2 @ git+https://github.com/RyougiKukoc/vs-dfttest2-api4.git@cu129"

cpu installs only the CPU plugin. cu121 installs the CPU plugin plus CUDA 12.1 builds of the NVRTC and cuFFT backends. cu129 does the same with CUDA 12.9. Linux CUDA packages include the matching cuFFT/cudart runtime beside the plugins; the NVIDIA driver is supplied by the host. If you switch between tags, use --force-reinstall so pip replaces the already-installed wheel with the other variant.

On Linux x86_64, the hook downloads dfttest2-<variant>-linux-x86_64.zip from the matching cpu, cu121, or cu129 Release. All Linux wheels are tagged manylinux_2_27_x86_64. Package-wide ABI inspection finds a maximum of GLIBC 2.14 / GLIBCXX 3.4.18 for CPU, GLIBC 2.27 / GLIBCXX 3.4.22 for cu121, and GLIBC 2.27 / GLIBCXX 3.4.21 for cu129; each CUDA package includes libcufft.so.11 and libcudart.so.12, both at or below the same GLIBC floor. The static NVRTC archives do not add a runtime ELF payload. VapourSynth R79 itself requires a compatible glibc 2.27 environment. Set DFTTEST2_FORCE_BUILD=1 to use the local CMake fallback. It obtains API4 headers and vapoursynth.pc from the isolated build environment's vapoursynth/pkgconfig directory while retaining any existing PKG_CONFIG_PATH entries.

Usage

from dfttest2 import DFTTest
output = DFTTest(input)

See also VapourSynth-DFTTest.

Compilation

The default local CMake configuration builds CPU and NVRTC. Enable ENABLE_CUDA as well when building the cuFFT backend:

cmake -S . -B build -G Ninja `
  -D ENABLE_CPU=ON `
  -D ENABLE_NVRTC=ON `
  -D ENABLE_CUDA=OFF `
  -D ENABLE_GCC=OFF `
  -D ENABLE_HIP=OFF `
  -D USE_NVRTC_STATIC=ON `
  -D VS_INCLUDE_DIR=C:\path\to\vapoursynth\include

cmake --build build

For reproducible release packages, use the GitHub Actions workflow. It builds and smoke-tests the CPU-only, CUDA 12.1, and CUDA 12.9 variants separately, then uploads the Windows and Linux zip/wheel pair for the selected variant to the matching release tag. Linux CUDA release builds are compiled separately for CUDA 12.1 and CUDA 12.9; do not install both CUDA variants together.

Behavior fixes and regression checks

For ftype=0, non-default f0beta values now reach the native power calculation as an unscaled exponent. The CPU helper also forwards zmean=False, allowing the mean to be filtered. These corrections intentionally change affected outputs. Native CPU calls reject missing or incorrectly sized window/sigma arrays with a VapourSynth error, and GPU entrypoints reject variable formats and unsupported sample types before initializing CUDA. GPU kernels use 16-bit storage for 9–15-bit input.

Clang 20.1.8 with -ffast-math can miscompile VCL's AVX2 power calculation even with finite input. The CMake CPU frame targets add -fno-finite-math-only while retaining SIMD, FMA and the other optimizations. Windows packages continue to use the separate CPU, CUDA 12.1 and CUDA 12.9 builds with static NVRTC.

CI runs native argument checks, mathematical CPU output checks, and explicit tests of autoload isolation before packaging. Run the same suite against a built package with an R77 Python environment:

python tools/test_regressions.py --artifact-dir dist/windows-cu121 --vapoursynth-root _deps/vapoursynth-wheel-R77

On a compatible NVIDIA machine, add --gpu to exercise NVRTC and cuFFT with 8-bit, 10-bit, 16-bit and float input. Smoke helpers also verify the actual loaded DLL paths, so another installed copy cannot silently satisfy a check.