vapoursynth-fftspectrum-rs

A faster FFT spectrum VapourSynth plugin.

Install

Via vsrepo:

vsrepo install fftspectrum_rs

Manually: download a release from the Releases page and unzip fftspectrum_rs.dll (Windows), libfftspectrum_rs.so (Linux), or libfftspectrum_rs.dylib (macOS) into a plugins directory. There are separate artifacts for Raptor Lake (*-raptorlake.zip) and AMD Zen 4 (*-znver4.zip) CPUs which may or may not have better performance than the plain x86_64 build.

API

fftspectrum_rs.FFTSpectrum(clip: vs.VideoNode) -> vs.VideoNode

• clip — Input video node. Must have 32-bit float format.

Only the first plane of clip will be processed. The output is a GRAYS video node.

Benchmark

On 1080p clips, this plugin is about 2.2 times faster than fftspectrum.FFTSpectrum(). This is mostly accomplished by minimizing the number of allocations, copies, and transposes done.

import numpy as np
from cv2 import DFT_COMPLEX_OUTPUT, dft
from vssource import BestSource
from vstools import core, depth, get_y, set_output, vs

core.set_affinity(range(0, 32, 2), 22000)

src = BestSource.source("/path/to/1080_clip.mkv", bits=0)
src_8 = depth(src, 8)
src_32 = depth(src, 32)

# fftspectrum
set_output(src_8.fftspectrum.FFTSpectrum(), "fftspectrum")

# numpy + ModifyFrame
def _to_polar(f: vs.VideoFrame, n: int) -> vs.VideoFrame:
    src = np.asarray(f[0])
    dft_shift = np.fft.fftshift(dft(src, flags=DFT_COMPLEX_OUTPUT))
    mag = np.sqrt(np.power(dft_shift[:, :, 1], 2) + np.power(dft_shift[:, :, 0], 2))
    dst = f.copy()
    np.copyto(np.asarray(dst[0]), np.log(mag) / 10)
    return dst

y = get_y(src_32)
set_output(y.std.ModifyFrame(y, _to_polar), "numpy ModifyFrame")

# fftspectrum_rs
set_output(src_32.fftspectrum_rs.FFTSpectrum(), "fftspectrum_rs")

$ hyperfine --warmup 1 'vspipe test.py --end 1999 -o {output_node} .' -P output_node 0 2
Benchmark 1: vspipe test.py --end 1999 -o 0 .
  Time (mean ± σ):     31.878 s ±  0.853 s    [User: 44.497 s, System: 1.592 s]
  Range (min … max):   30.672 s … 33.681 s    10 runs

Benchmark 2: vspipe test.py --end 1999 -o 1 .
  Time (mean ± σ):     62.379 s ±  1.839 s    [User: 59.013 s, System: 22.801 s]
  Range (min … max):   59.342 s … 64.992 s    10 runs

Benchmark 3: vspipe test.py --end 1999 -o 2 .
  Time (mean ± σ):     13.909 s ±  0.146 s    [User: 178.334 s, System: 18.904 s]
  Range (min … max):   13.669 s … 14.166 s    10 runs

Summary
  vspipe test.py --end 1999 -o 2 . ran
    2.29 ± 0.07 times faster than vspipe test.py --end 1999 -o 0 .
    4.48 ± 0.14 times faster than vspipe test.py --end 1999 -o 1 .

Benchmark	Mean (fps)	Range (fps)
fftspectrum	62.74 ± 1.68	59.38 … 65.21
numpy ModifyFrame	32.06 ± 0.95	30.77 … 33.70
fftspectrum_rs	143.79 ± 1.51	141.18 … 146.32

Build

Rust v1.91.0-nightly and cargo may be used to build the project. Older versions will likely work fine but they aren't explicitly supported.

$ git clone https://github.com/sgt0/vapoursynth-fftspectrum-rs.git
$ cd vapoursynth-fftspectrum-rs

# Debug build.
$ cargo build

# Release (optimized) build.
$ cargo build --release

# Release build optimized for the host CPU.
$ RUSTFLAGS="-C target-cpu=native" cargo build --release