Skip to content

Enhanced Edge Directed Interpolation implemented in CUDA

License

Notifications You must be signed in to change notification settings

yunginnanet/VapourSynth-EEDI2CUDA

 
 

Repository files navigation

VapourSynth-EEDI2CUDA

EEDI2 filter using CUDA

Ported from HolyWu's plugin.

Usage

Doubling Height

eedi2cuda.EEDI2(clip clip, int field[, int mthresh=10, int lthresh=20, int vthresh=20, int estr=2, int dstr=4, int maxd=24, int map=0, int nt=50, int pp=1, int[] planes=[0, 1, 2], int num_streams=1, int device_id=-1])

Arguments have the exactly same meanings with the ones of Holy's EEDI2. See descriptions there. Only pp=1 is implemented.

Additional arguments:

  • planes: specify which planes to process. For EEDI2 and Enlarge2, unprocessed planes in the result clip contain garbage. For AA2, unprocessed planes are copied as-is.
  • num_streams: specify the number of CUDA streams. The value must less than or equal to core.num_threads. A larger value increases the concurrency and also increases the GPU memory usage. The default value num_streams=1 is already fast enough.
  • device_id: set the GPU device ID to use. You must specify this argument for each call if you have multiple GPUs.

Enlarge 2X

eedi2cuda.Enlarge2(clip clip[, int mthresh=10, int lthresh=20, int vthresh=20, int estr=2, int dstr=4, int maxd=24, int map=0, int nt=50, int pp=1, int[] planes=[0, 1, 2], int num_streams=1, int device_id=-1])

Enlarge the clip by 2. Offsets caused by doubling are neutralized. This can be considered as a faster equivalent of the code below:

el = core.eedi2cuda.EEDI2(clip, 1)
el = core.fmtc.resample(el, sy=[-.5, -.5 * (1 << clip.format.subsampling_h)])
el = core.std.Transpose(el)
el = core.eedi2cuda.EEDI2(el, 1)
el = core.fmtc.resample(el, sy=[-.5, -.5 * (1 << clip.format.subsampling_w)])
el = core.std.Transpose(el)

Anti-aliasing

eedi2cuda.AA2(clip clip[, int mthresh=10, int lthresh=20, int vthresh=20, int estr=2, int dstr=4, int maxd=24, int map=0, int nt=50, int pp=1, int[] planes=[0, 1, 2], int num_streams=1, int device_id=-1])

Double and then scale back to do anti-aliasing. Offsets caused by doubling are neutralized. This can be considered as a faster equivalent of the code below:

w = clip.width
h = clip.height
aa = core.eedi2cuda.EEDI2(clip, 1)
aa = core.fmtc.resample(aa, w, h, sy=[-.5, -.5 * (1 << clip.format.subsampling_h)])
aa = core.std.Transpose(aa)
aa = core.eedi2cuda.EEDI2(aa, 1)
aa = core.fmtc.resample(aa, h, w, sy=[-.5, -.5 * (1 << clip.format.subsampling_w)])
aa = core.std.Transpose(aa)

It uses spline36 with extended filter size for downscaling.

Using in AviSynth+

  • EEDI2_CUDA
  • EEDI2_CUDA_Enlarge2
  • EEDI2_CUDA_AA2

Arguments are the same. All bindings share a unified DLL.

Using in mpv

You can use eedi2cuda in mpv as the upscaler for realtime playback. First ensure the VapourSynth video filter is available in mpv. Then copy the script below and save it as eedi2enlarge.vpy:

src16 = video_in.resize.Point(format=video_in.format.replace(bits_per_sample=16))
src16.eedi2cuda.Enlarge2().set_output()

In commandline option, load the script:

mpv --vf=vapoursynth=eedi2enlarge.vpy

Or you can specify it in mpv.conf.

In most scenarios, it is not necessary to use an advanced upscaler for UV planes. The script below does per-plane upscaling:

import vapoursynth as vs
from vapoursynth import core

src16 = video_in.resize.Point(format=video_in.format.replace(bits_per_sample=16))
w = src16.width
h = src16.height
y = core.eedi2cuda.Enlarge2(src16, planes=0)
uv = core.resize.Point(src16, w * 2, h * 2,
                       format=vs.YUV444P16,
                       resample_filter_uv="spline36")
core.std.ShufflePlanes([y, uv], [0, 1, 2], vs.YUV).set_output()

Requirements

  • A CUDA-enabled GPU of compute capability 5.0 or higher (Maxwell+).
  • GPU driver 461.33 or newer.

Compilation

Please refer to build.yml.

Performance

EEDI2CUDA Holy's EEDI2
EEDI2 175.33 fps 18.44 fps
Enlarge2 71.78 fps 6.42 fps
AA2 104.56 fps 9.62 fps

1080P YUV420P16 input. num_streams=16. Tested on i7-10875H and RTX 2070.

About

Enhanced Edge Directed Interpolation implemented in CUDA

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Cuda 78.3%
  • C++ 19.2%
  • CMake 2.5%