VapourSynth Portable FATPACK - 64Bit, Editor, Plugins, Scripts, Apps & endless FUN!

[asarian]

@Wolfberry & @WolframRhodium, thank you both kindly for your great assistance! I finally got it working! My, that was a lot of sweat for just bilateralGPU.dll. But the good news is, I went from 2.9 fps to 4.78 fps (including the vspipe to x264 process, of course); and I can visually see (GPU-Z) my video card getting taxed more. So, gentlemen, this was well worth the effort! So, thank you both again.

I'm not sure the forum rules would allow such a thing, but, if so, would be nice if the starter post could contain a link to a small 'dependency' repository for bilateralGPU. I know it's just 1 filter, but debanding is something I'm sure a lot of ppl are doing (because of compression), and with the big GPU's out there, I can attest that it's definitely a worthwhile filter. And for ppl who didn't have those dependencies, some of those are impossible to find on the net.

P.S. Actually, the whole 'Dllref = ctypes.windll.LoadLibrary' trick didn't work for me. I simply put core.std.LoadPlugin ("C:/VS/cuda/bilateralGPU.dll"), and had stuck all dependencies inside the C:/VS/cuda/ folder (figuring, as expected, that libraries always look inside their own directory for dependencies first, which worked).

Again, thank you guys, and especially you, @Wolfberry. I know I tend to ask a lot of questions, but you made me very happy today, and I'm pretty sure others will be happy with the result too.

[ChaosKing]

So these two links contains all dependencies for bilateralGPU? Can't test myself.
https://mega.nz/#!3ygkwKIA!Oa1vAO9FK..._IMddHZeSEcAeA
https://mega.nz/#!ejAUSCjD!x-s9Aqr95...9JEroASOJY0bt8

[Wolfberry]

This one is needed, too.

[asarian]

Speaking of GradFun3, I'm really loving the GPU-assisted speed. GradFun3 doesn't come with presets, though. Usually I did:

vid = core.f3kdb.Deband (vid, preset="medium", dither_algo=2)

What settings could I use on GradFun3 to achieve a similar strength as on f3kdb.Deband? In its default state, the debanding of GradFun3 is too little, for my taste.

[WolframRhodium]

Originally Posted by asarian

Speaking of GradFun3, I'm really loving the GPU-assisted speed. GradFun3 doesn't come with presets, though. Usually I did:

vid = core.f3kdb.Deband (vid, preset="medium", dither_algo=2)

What settings could I use on GradFun3 to achieve a similar strength as on f3kdb.Deband? In its default state, the debanding of GradFun3 is too little, for my taste.

You can increase 'thr', 'radius' and 'elast', especially the first one, which should be the most important parameter that controls the strength.

[asarian]

Originally Posted by WolframRhodium

You can increase 'thr', 'radius' and 'elast', especially the first one, which should be the most import parameter that controls the strength.

Thx. I'm not that mathematical, so most of these options usually dazzle me.

[asarian]

After several days of testing, I hate to admit that vid = core.f3kdb.Deband (vid, preset="medium", dither_algo=2) does a much better job of preserving the gradients than, for example, vid = fvs.GradFun3 (src=vid, thr=0.6, radius=16, elast=4.0, smode=5), even though the loss of non GPU-assisted processing is felt dearly. Dunno, but for some reason x264 just eats thru GradFun3 like crazy (and I use x264 with very high settings, for quality). Feels like the pattern of f3kdb.Deband is just more compression-resilient somehow. Unless I'm doing GradFun3 wrong, of course (always a possibility).

So, here's to hoping f3kdb.Deband gets ported to GPU one day!

[asarian]

Originally Posted by HolyWu

Are you kidding? f3kdb already runs at 100+ fps with a 1920x1080 video. What makes you think that f3kdb is the bottleneck of your filtering chain?

Because it is. Keep in mind, that if you use a GPU-assisted debander, it doesn't take any CPU (or nothing you will notice, at least).

The x264 process usurps about 85-92%; vspipe around 1.7% (max). Which means the x246 process doesn't get enough data fast enough to fully saturate the CPU. The whole vspipe process (with x264) takes around 93% CPU, on average. In practice, this means I process around 2.95 fps with f3kdb (as last in the chain), and ca. 4.25 fps with bilateralGPU-assisted GradFun3.

The script itself is nothing special; like:

import vapoursynth as vs

core = vs.get_core ()
core.max_cache_size = 16384

core.avs.LoadPlugin ("C:/Program Files (x86)/dgdecnv/x64 Binaries/DGDecodeNV.dll")

vid = core.avs.DGSource (r'f:\jobs\jup.dgi', resize_w=1920, resize_h=1080, fulldepth=True)
vid = core.dghdrtosdr.DGHDRtoSDR (clip=vid, fulldepth=False)
vid = core.avs.DGDenoise (vid, strength=0.1, searchw=9)
vid = core.f3kdb.Deband (vid, preset="medium", dither_algo=2)

vid.set_output ()

Versus GradFun3:

import vapoursynth as vs
import fvsfunc as fvs

core = vs.get_core ()
core.max_cache_size = 16384

core.avs.LoadPlugin ("C:/Program Files (x86)/dgdecnv/x64 Binaries/DGDecodeNV.dll")
core.std.LoadPlugin ("C:/VS/cuda/bilateralGPU.dll")

vid = core.avs.DGSource (r'i:\jobs\jup.dgi', resize_w=1920, resize_h=1080, fulldepth=True)
vid = core.dghdrtosdr.DGHDRtoSDR (clip=vid, fulldepth=False)
vid = core.avs.DGDenoise (vid, strength=0.1, searchw=9)
vid = fvs.GradFun3 (src=vid, thr=0.6, radius=16, elast=4.0, smode=5)

vid.set_output ()

2.95 fps vs. 4.25 fps may not seem like much, but it's the difference between roughly waiting 16 hours before job-completion, or a full 24.

[poisondeathray]

Originally Posted by asarian

After several days of testing, I hate to admit that vid = core.f3kdb.Deband (vid, preset="medium", dither_algo=2) does a much better job of preserving the gradients than, for example, vid = fvs.GradFun3 (src=vid, thr=0.6, radius=16, elast=4.0, smode=5), even though the loss of non GPU-assisted processing is felt dearly. Dunno, but for some reason x264 just eats thru GradFun3 like crazy (and I use x264 with very high settings, for quality). Feels like the pattern of f3kdb.Deband is just more compression-resilient somehow. Unless I'm doing GradFun3 wrong, of course (always a possibility).

So, here's to hoping f3kdb.Deband gets ported to GPU one day!

2.95 fps vs. 4.25 fps

What about 10bit encoding ? In general, 10bit encoding is slower than 8bit , but you likely do not even need to apply a debanding filter

[asarian]

Originally Posted by poisondeathray

What about 10bit encoding ? In general, 10bit encoding is slower than 8bit , but you likely do not even need to apply a debanding filter

That's a good idea. I haven't done 10-bit encodings yet; primarily, really, because I've been off this world for a spell, x264-wise, and only recently got interested in re-encoding again, now UHD Blu-rays became available. And I haven't even checked whether Kodi can even play it properly (although it should, as it can handle my UHD disc too).

What you're saying sounds very logical, though: 10-bit encoding should work wonders for gradients. And without a debanding filter, things go much faster still (ca. 5.15 fps). My next encode can start in about 6 hours. I shall try out a 10-bit conversion, and will report back. Thanks.

[ChaosKing]

New update, see first post.
I used my build script this time. I hope everything works: https://github.com/theChaosCoder/vap...rtable-FATPACK

[ThePi7on]

Originally Posted by ChaosKing

New update, see first post.
I used my build script this time. I hope everything works: https://github.com/theChaosCoder/vap...rtable-FATPACK

Thanks for the new update!
Would it be possible to add the new cunet waifu2x models in the next update?

[ChaosKing]

If the models are compatible with the VS version of waifu2x, sure.

[brucethemoose]

Love this pack, thanks!

Originally Posted by ChaosKing

If the models are compatible with the VS version of waifu2x, sure.

Speaking of Waifu, could you throw in some of muvsfunc's super_resolution dependancies? I guess you can't distribute cuDNN/CUDA, but I got most of what I needed in the Python directory with "pip install mxnet-cu92 --pre"

You could even add a few of the upscaling models with presets, but that's kinda a stretch since they're big files. Also, I'm not sure which ones are best to add yet... Need more testing.


Another small request: ffmpeg, just for a GPU encoding preset. GPU 10-bit hevc isn't so bad these days, especially if speed is a priority.

[ChaosKing]

Hmm I don't really want to ship a big chunk of dependencies that only nvidia users can use and since I only have a amd card here I can't test it myself. But I could link to a zip package similar to the bilineargpu plugin...

I think I can at least add some bat or vseditor-templates for ffmpeg (gpu) encoding. You use NVenc I guess?

[brucethemoose]

Originally Posted by ChaosKing

Hmm I don't really want to ship a big chunk of dependencies that only nvidia users can use and since I only have a amd card here I can't test it myself. But I could link to a zip package similar to the bilineargpu plugin...

I think I can at least add some bat or vseditor-templates for ffmpeg (gpu) encoding. You use NVenc I guess?

Yeah. Reasonably new AMD GPUs and Intel IGPs support hardware encoding too, I used to do it on my 7950 pretty regularly.

Also, there's a mxnet package that supports either Intel CPUs or Nvidia GPUs, depending on which is available. It's not too big, so I don't think it would hurt to include it: https://pypi.org/project/mxnet-cu92mkl/

AMD support for MXNet is coming soon, AFAIK... I'm on an Nvidia GPU atm, so in the meantime I can help test if you want.

[ChaosKing]

Originally Posted by WorBry

Just checked out VapourSynth Portable FATPAK. Don't seem to be able to open the scripts in VirtualDub2 (64bit) for encoding - 'AVI Import Error' - so I assume that, like Myrsloik's portable version, output is VSPipe only ?

Finally tested the vfw stuff. It seems these reg entries are enough to get vdub to read vpy files:^
(don't use it if you have vapoursynth already installed)

Windows Registry Editor Version 5.00

[HKEY_LOCAL_MACHINE\SOFTWARE\Classes\CLSID\{58F74CA0-BD0E-4664-A49B-8D10E6F0C131}]
@="VapourSynth"

[HKEY_LOCAL_MACHINE\SOFTWARE\Classes\CLSID\{58F74CA0-BD0E-4664-A49B-8D10E6F0C131}\InProcServer32]
@="D:\\VapourSynth64Portable\\VapourSynth64\\VSVFW.dll"
"ThreadingModel"="Apartment"


[HKEY_LOCAL_MACHINE\SOFTWARE\Classes\AVIFile]

[HKEY_LOCAL_MACHINE\SOFTWARE\Classes\AVIFile\Extensions\VPY]
@="{58F74CA0-BD0E-4664-A49B-8D10E6F0C131}"

[unix]

Why I'm getting this error?!

File "...VapourSynth64Portable\Scripts\CSMOD.py", line 1117, in CSMOD
sharpdiff2 = core.focus.TemporalSoften(sharpdiff, 1, 255, 255 if chroma else 0, 32, 2)
File "src\cython\vapoursynth.pyx", line 1803, in vapoursynth.Function.__call__
vapoursynth.Error: TemporalSoften: Too many unnamed arguments specified

script:

import fvsfunc as fvf#
import havsfunc as hvf #
import hnwvsfunc as hnwvsfunc
import CSMOD as cs

src = core.lsmas.LWLibavSource(source='69.mkv')
src = hvf.daa(src).std.AssumeFPS(fpsnum=24000,fpsden=1001)
c = cs.CSMOD(src, ss_w=2, ss_h=2, ssout=True)
c.set_output()

[ChaosKing]

Remove libfocus-em64t.dll and try again. There are currently 2 TemporalSoften dlls with the same namespace inside the plugins folder. (VapourSynth64Portable\VapourSynth64\vapoursynth64\plugins)

[unix]

Originally Posted by ChaosKing

Remove libfocus-em64t.dll and try again. There are currently 2 TemporalSoften dlls with the same namespace inside the plugins folder. (VapourSynth64Portable\VapourSynth64\vapoursynth64\plugins)

It works
Thanx.