porting LSFmod for HBD in avs+

[LeXXuz]

can you try this update? https://github.com/realfinder/AVS-Stuff/raw/Community/avs%202.5%20and%20up/LSFmod.avsi

Today I'm not at home. I'll test it tomorrow and report back to you. :)

[LeXXuz]

Okay, here are the results:

Source:
https://abload.de/thumb/2020-07-07_source75jcc.png (https://abload.de/image.php?img=2020-07-07_source75jcc.png)

LSFmod (new) in 8-bit:
https://abload.de/thumb/2020-07-071_8bit5xjij.png (https://abload.de/image.php?img=2020-07-071_8bit5xjij.png)

LSFmod (new) in 10-bit:
https://abload.de/thumb/2020-07-072_10bit7xjhk.png (https://abload.de/image.php?img=2020-07-072_10bit7xjhk.png)

LSFmod (new) in 16-bit:
https://abload.de/thumb/2020-07-073_16bitjtkwu.png (https://abload.de/image.php?img=2020-07-073_16bitjtkwu.png)

LSFmod (old) in 16-bit:
https://abload.de/thumb/2020-07-074_16bit-oldb2ko1.png (https://abload.de/image.php?img=2020-07-074_16bit-oldb2ko1.png)

Looks alright to my eyes. Out of curiosiy, what did you change exactly?

Anyhow thank you very much for maintaining this script and looking into this! :) :thanks:

[real.finder]

Out of curiosiy, what did you change exactly?


https://github.com/realfinder/AVS-Stuff/commit/101bb78673aec19e71dd1613a93f93a0ead9e2ff

I did this, in 8bit will still same since the lut will round it (I used 0.51, 0.5 should also work but it's not for some reason)

[LeXXuz]

I have the impression the sharpening effekt is a little weaker in 10/16 but maybe my old eyes are just messing with me. :D

Will sharpening strength be about the same in HBD than in LBD?
Or should I use slightly higher values when filtering my videos in 10bit, compared to 8bit?

[real.finder]

I have the impression the sharpening effekt is a little weaker in 10/16 but maybe my old eyes are just messing with me. :D

Will sharpening strength be about the same in HBD than in LBD?
Or should I use slightly higher values when filtering my videos in 10bit, compared to 8bit?

it should be in same strength, and it also depends on how you display the HBD (like convert to 8bit at the end with dither or not, and in megui it just convert it to rgb24 for any bitdepth), but anyway the HBD outputs for most filters are not similar to 8bit in first place and that why this problem happened

[LeXXuz]

Okay, thanks again. :)

[WorBry]

edit: all old talk that was here moved to the 2nd post now

the porting is done https://github.com/realfinder/AVS-Stuff/raw/master/avs%202.5%20and%20up/LSFmod.avsi

I've just tried to run the script and I'm getting "I don't know what AvsPlusVersionNumber means, LSFMod.avsi, line 433"

What to do? I'm running AviSynthPlus 3.6.1

Thanks.

[real.finder]

I've just tried to run the script and I'm getting "I don't know what AvsPlusVersionNumber means, LSFMod.avsi, line 433"

What to do? I'm running AviSynthPlus 3.6.1

Thanks.

you need this https://github.com/realfinder/AVS-Stuff/raw/master/avs%202.5%20and%20up/Zs_RF_Shared.avsi

[WorBry]

OK thanks.

[WorBry]

So I've been (re)evaluating denoise and sharpening filter combinations for post processing of graded DaVinci Resolve exports - these are 1080/60p Cineform (Filmscan 2) avi files exported at 'Video Levels' but retaining super-white data. The original source material is 1080/60p HD-AVC.mp4 shot on a Canon HF-G40 camcorder in 'Wide DR' gamma mode. Anyone familiar with Wide DR on Canon cameras will know, the 'extended dynamic range', as perceived, comes at the expense of a softer image and increased levels of gain noise in the shadows. Adding back local contrast ('clarity', or 'Midtone Detail' in Resolve), coupled with light sharpening, offers the best prospect for 'enhancing' more detail. The sharpen/blur tool-set in Resolve is very good (IMO), but I'm using the free version and NR is only available in the Studio version. I'd consider the 'NeatVideo' OFX plugin but, AFAIK, that requires the Studio version also.

Hence, my reason for turning to AVISynth.

I've been getting very nice results with KNLMeans and LSFMod processed in 10bit 422 (p210) pipeline. Typical script, for Wide DR:

LWLibavVideoSource("{Path}:\Cineform_export.avi", fpsnum=60000, fpsden=1001,format="YUV422P10")
KNLMeansCL(h=1.2) #up to 1.5 for noisier scenes.
LSFMod(LSFMod(strength=110, defaults="slow")#ranging between 90 - 120.

I'm quite happy with the results as they stand but, if possible, I'd like to look at adding back noise after sharpening. Searching back through the forum I found this snippet from Didee:

https://forum.doom9.org/showthread.php?p=1651062#post1651062

source = whatever
filter = source.AnyDenoiserThatFitsTheSource()
sharpy = filter.YourSharpenerOfChoice()
result = source.mt_makediff(mt_makediff(filter,sharpy),U=2,V=2)

But I think the MaskTools expression would need to be updated for HBD processing - wouldn't it ? - and I'm not sure how to do that. Also how could it be modified to add back noise in varying degrees ?

Thanks.

[poisondeathray]

But I think the MaskTools expression would need to be updated for HBD processing - wouldn't it ? - and I'm not sure how to do that. Also how could it be modified to add back noise in measured degrees ?


Modern MaskTools2 for avs+ supports HBD and float for most functions. There is a chart in the readme
https://github.com/pinterf/masktools/releases

Merge can "merge" clips with weighting and supports HBD in avs+ .
http://avisynth.nl/index.php/Merge

[WorBry]

OK thanks. I'm not at my PC just now to try this, but just to be clear - I should be able to use that MakeDiff script for adding back noise as is, or do I need to add some operator for 10bit scaling ?

[poisondeathray]

OK thanks. I'm not at my PC just now to try this, but just to be clear - I should be able to use that MakeDiff script for adding back noise as is, or do I need to add some operator for 10bit scaling ?

You don't need anything for 10bit scaling, masktools works up to 16bits and some functions at float

That script is not "adding back noise" , it's applying sharpening to a denoised version

It might look something like this




source=last

#d is denoiser apply only

d = source.whateverdenoisers()

#ds is denoiser+sharpener

ds = d.whateversharpeners()

#result1 is add back to ds, the diff between denoised and source

result1=mt_adddiff(ds, mt_makediff(d,source),U=2,V=2)

#modulate strength of addback to ds from 0.0 to 1.0 .

merge(ds, result1, 0.5)

[WorBry]

That script is not "adding back noise" , it's applying sharpening to a denoised version

True, strictly it's by-passing noise - sharpening a denoised version and applying the difference between that and the denoised version (i.e. the denoised sharpening) to the original noisy version, so leaving the noise in (not adding it back).



source=last

#d is denoiser apply only

d = source.whateverdenoisers()

#ds is denoiser+sharpener

ds = d.whateversharpeners()

#result1 is add back to ds, the diff between denoised and source

result1=mt_adddiff(ds, mt_makediff(d,source),U=2,V=2)

#modulate strength of addback to ds from 0.0 to 1.0 .

merge(ds, result1, 0.5)


OK thanks, I'll give it a go.

Edit: Yes, that works. Thanks.

[Dogway]

Updated LSFmod to use internal Expr() and dropped avs 2.5 support. Runs faster on my tests. Still needs masktools2 for the convolutions.
https://github.com/Dogway/Avisynth-Scripts/blob/master/Mods/LSFmod.v2.192m.avsi

[real.finder]

Updated LSFmod to use internal Expr() and dropped avs 2.5 support. Runs faster on my tests. Still needs masktools2 for the convolutions.
https://github.com/Dogway/Avisynth-Scripts/blob/master/Mods/LSFmod.v2.192m.avsi

how much faster? masktools2 lut things can use internal Expr (use_expr) but yes pinterf said back then that using Expr directly will be faster but even so with most 8bit RPN lut still faster from my tests

[Reel.Deel]

Updated LSFmod to use internal Expr() and dropped avs 2.5 support. Runs faster on my tests. Still needs masktools2 for the convolutions.
https://github.com/Dogway/Avisynth-Scripts/blob/master/Mods/LSFmod.v2.192m.avsi

Thanks for the update Dogway. Death to AviSynth 2.5 :devil:

[Dogway]

Near 4% faster. The limiting factor here is the convolutions (mt_edge, mt_xxpand, mt_xxflate, etc). If VS convolution is ever ported probably the difference would be bigger. I kinda want to deprecate masktools2 but some functions are still needed.

CPU: i7-4790K (4C/8T)
2.192: 41.5fps
2.192m: 43fps

setmemorymax(2048)
DGSource("source.dgi") # 1080p

ConvertBits(16)
LSFmod(defaults="slow",strength=200,edgemode=0,soothe=true,ss_x=1.0,ss_y=1.0)

trim(7529,12601)
Prefetch(4)

@Reel.Deel: lol, yeah even me haven't found a reason to go back.

[real.finder]

yeah there are some functions need to be ported from vs like https://forum.doom9.org/showpost.php?p=1941761&postcount=318 and as you said VS convolution

anyway, you test with 16bit, did you tried with 8bit?

@Reel.Deel, Dogway also killed avs 2.6 :) and any old avs+ that don't has expr (that not as important as kill avs 2.6 :) anyway)

[Dogway]

Yes, box blur is a type a convolution filter, in that case one of variable size.

In 8-bit masktools2 seems faster. This was unexpected. I will test with simple calls if there's something else that can be done.

2.192: 97fps
2.192m: 67fps

[Reel.Deel]

@Reel.Deel, Dogway also killed avs 2.6 :) and any old avs+ that don't has expr (that not as important as kill avs 2.6 :) anyway)

Good, there is no reason to use classic AviSynth over AviSynth+ ... And if people don't want to update to the latest version, well, shame on them. Your script modifications would greatly benefit from you doing the same thing, they would be much cleaner without the hundreds of calls to "sisavs26", "isyuy2", etc. But hey, to each his own :)

[StainlessS]

And if people don't want to update to the latest version, well, shame on them.
No shame at all, personal choice, however so long as it runs on XP, I'm stickin' with + [although I feel myself quivering a bit on that XP thing].
Not sure, I think I did last Pole on Avisynth version usage [or maybe the penultimate one], maybe Reel.Deel
or one of the other NonPlus haters would hold a new pole, see how the land lies.
[surely nobody is still usin' 2.58].

[Dogway]

Good news I found how to use convolutions with Expr()

BoxBlur example, same as Blur(1.0) but faster:
removegrain(19)

BoxBlur with Expr(), slower by 10-15% than removegrain in either bitdepth, but still need to compare with masktools2.
Expr(last,"x[-1,-1] x[1,1] x[-1,0] x[1,0] x[0,1] x[0,-1] x[-1,1] x[1,-1] + + + + + + + 8 /")



EDIT: yep, Expr() is slower (no AVX2)
487fps
mt_edge(mode="sobel",thY1=0.0,thY2=0.0,thC1=0.0,thC2=0.0,chroma="-128")

353fps
Expr(last,"x[-2,0] x[-1,1] x[1,1] x[2,0] x[1,-1] x[-1,-1] + - - - + 2 ^ " \
+"x[-1,0] x[-1,1] x[0,2] x[1,1] x[1,-1] x[0,-2] x[-1,-1] + + + - - - 2 ^ * sqrt", "range_half")

[StainlessS]

DogWay

Avisource("D:\Parade.avi")
return Expr(last,"x[-1,-1] x[1,1] x[-1,0] x[1,0] x[0,1] x[0,-1] x[-1,1] x[1,-1] + + + + + + + 8 /") # Dogway

EDIT: Oops, I posted the mod above, instead of DogWay original, fixed.


AVSMeter 3.0.8.0 (x86), (c) Groucho2004, 2012-2021
AviSynth+ 3.7.0 (r3382, 3.7, i386) (3.7.0.0)

Number of frames: 7373
Length (hh:mm:ss.ms): 00:04:06.012
Frame width: 488
Frame height: 360
Framerate: 29.970 (30000/1001)
Colorspace: YV12
Audio channels: 2
Audio bits/sample: 16
Audio sample rate: 44100
Audio samples: 10849148

Frames processed: 7373 (0 - 7372)
FPS (min | max | average): 204.1 | 343.6 | 307.6 # Dogways
Process memory usage (max): 39 MiB
Thread count: 9
CPU usage (average): 25.3%

Time (elapsed): 00:00:23.970


Mod (assuming that is in form [x,y]), then all Y-1, then all Y, then all Y+1. [& all left to right,instead of jumping all over the place]
EDIT: I presume that Expr() may be able to take shortcuts, if ordering is as expected in raster order.
EDIT: + possibly more CPU cache friendly.

Avisource("D:\Parade.avi")
return Expr(last,"x[-1,-1] x[0,-1] x[1,-1] x[-1,0] x[1,0] x[-1,1] x[0,1] x[1,1] + + + + + + + 8 /")



AVSMeter 3.0.8.0 (x86), (c) Groucho2004, 2012-2021
AviSynth+ 3.7.0 (r3382, 3.7, i386) (3.7.0.0)

Number of frames: 7373
Length (hh:mm:ss.ms): 00:04:06.012
Frame width: 488
Frame height: 360
Framerate: 29.970 (30000/1001)
Colorspace: YV12
Audio channels: 2
Audio bits/sample: 16
Audio sample rate: 44100
Audio samples: 10849148

Frames processed: 7373 (0 - 7372)
FPS (min | max | average): 142.9 | 357.7 | 319.1 # Mod
Process memory usage (max): 37 MiB
Thread count: 9
CPU usage (average): 26.5%

Time (elapsed): 00:00:23.104


However, should this boxblur thingy not include the original pixel itself, and not just the ones surrounding it.
EDIT: I've no idea how boxblur works. [EDIT: Yes, boxblur should include the original pixel value too, ie 9 coords]

This below would seem to suggest the Blur(1), RemoveGrain(19) and the Expr thingy, aint the same at all.

Avisource("D:\Parade.avi")
#Greyscale
AMP=True

A= Blur(1)
B= removegrain(19)
C=Expr(last,"x[-1,-1] x[1,1] x[-1,0] x[1,0] x[0,1] x[0,-1] x[-1,1] x[1,-1] + + + + + + + 8 /") # Dogway
D=Expr(last,"x[-1,-1] x[0,-1] x[1,-1] x[-1,0] x[1,0] x[-1,1] x[0,1] x[1,1] + + + + + + + 8 /") # Mod
E=Expr(last,"x[-1,-1] x[0,-1] x[1,-1] x[-1,0] x[0,0] x[1,0] x[-1,1] x[0,1] x[1,1] + + + + + + + + 9 /") # Mod, 9 coords

P=ClipDelta(A,B,AMP) # blur(1), RemoveGrain(19)
Q=ClipDelta(A,C,AMP) # blur(1), Dogway
R=ClipDelta(A,E,AMP) # blur(1), 9 coords
S=ClipDelta(B,C,AMP) # RemoveGrain(19), Dogway
T=ClipDelta(B,E,AMP) # RemoveGrain(19), 9 coords
U=ClipDelta(C,D,AMP) # Dogway, Mod : The only one that is identical
V=ClipDelta(C,E,AMP) # Dogway, 9 coords
P
return Last

Function ClipDelta(clip clip1,clip clip2,bool "amp",bool "show") {
amp=Default(amp,false)
show=Default(show,false)
c2=clip1.levels(128-32,1.0,128+32,128-32,128+32).greyscale()
c1=clip1.subtract(clip2)
c1=(amp)?c1.levels(127,1.0,129,0,255):c1
return (show)?c1.Merge(c2):c1
}

EDIT: Above changed around a bit.

ClipDelta(amp=true) shows any difference at all as pretty much max.

[real.finder]

well I do keep avs25 for some reasons, like use it as a reference either for experimentation or for code, aside from satisfaction the simple or poor peoples that still use old things, also in my personal opinion, I prefer to maintain compatibility with older versions whenever possible (indeed they will not have all privileges), finally it's kinda fun to do this :P

[StainlessS]

Masochism is a hard habit to break :) [I know, I've been there, partly still am there]

[real.finder]

seems I will do another update since I got https://github.com/realfinder/AVS-Stuff/blob/Community/avs%202.5%20and%20up/UnsharpMask_avsi.avsi

the only missed HBD now is http://avisynth.nl/index.php/VariableBlur/Unsharp

here https://github.com/realfinder/AVS-Stuff/blob/Community/avs%202.5%20and%20up/LSFmod.avsi

[StainlessS]

seems I will do another update since I got https://github.com/realfinder/AVS-Stuff/blob/Community/avs%202.5%20and%20up/UnsharpMask_avsi.avsi


RF, thanks, above, too many avsi's.

[Dogway]

mt_convolution is very very slow, in this case I recommend using Expr() for over double the speed. Stepping is not actually the same as Expr() conv doesn't accept floats.

function UnsharpMask_HBD(clip clp, float "strength", int "radius", float "threshold")
{
str = Default(strength,64)
radius = Default(radius,3)
thres = Default(threshold,8)

rd = radius - 1
str = str/128.

blurclip = Expr(clp,Format("x[-{rd},-{rd}] x[0,-{rd}] x[{rd},-{rd}] x[-{rd},0] x[0,0] x[{rd},0] x[-{rd},{rd}] x[0,{rd}] x[{rd},{rd}] + + + + + + + + 9 /"))

sIsRGB=clp.IsRGB()
e = Format("x y - abs {thres} scalef > x y - {str} * x + x ?")
Expr(clp, blurclip, e, sIsRGB?e:"x") }

[StainlessS]

Concerning the timing of the 8 pixel coord convolution blur whotsit [which looked more like a de-dot thingy to me, missed out the original pixel],
Doggy changed to 9 coord Expr() convolution so including the missing pixel, looks like 9 coords was correct.
This also is correct, according to wikiPedia Boxblur:- https://en.wikipedia.org/wiki/Box_blur

and,
A number of optimizations can be applied when implementing the box blur of a radius r and N pixels:[6]

The box blur is a separable filter, so that only two 1D passes of averaging 2 r + 1 pixels will be needed, one horizontal and one vertical, for each pixel. This lowers the complexity from O(Nr2) to O(Nr). In digital signal processing terminology, each pass is a moving-average filter.

"Separable filter", so same kind of thing avisynth does with 2 * 1D resizer instead of 2D resizer.

So, I tried this

# boxblur,, 9 coords
Avisource("D:\Parade.avi")
Expr("x[-1,-1] x[0,-1] x[1,-1] x[-1,0] x[0,0] x[1,0] x[-1,1] x[0,1] x[1,1] + + + + + + + + 9 /")
Return Last


Now has 9 coords, so slightly slower than the previously timed 8 coord whotsit.

AVSMeter 3.0.8.0 (x86), (c) Groucho2004, 2012-2021
AviSynth+ 3.7.0 (r3382, 3.7, i386) (3.7.0.0)

Number of frames: 7373
Length (hh:mm:ss.ms): 00:04:06.012
Frame width: 488
Frame height: 360
Framerate: 29.970 (30000/1001)
Colorspace: YV12
Audio channels: 2
Audio bits/sample: 16
Audio sample rate: 44100
Audio samples: 10849148

Frames processed: 7373 (0 - 7372)
FPS (min | max | average): 168.1 | 321.4 | 290.3
Process memory usage (max): 38 MiB
Thread count: 9
CPU usage (average): 25.7%

Time (elapsed): 00:00:25.395


2 * 1D [EDIT: or should it be horizontal(long side), then vertical(short side), I think there is mention of it somewhere about resizers]

# Vertical 3, then horizontal 3
Avisource("D:\Parade.avi")
Expr("x[0,-1] x[0,0] x[0,1] + + 3 /")
Expr("x[-1,0] x[0,0] x[1,0] + + 3 /")
return Last



AVSMeter 3.0.8.0 (x86), (c) Groucho2004, 2012-2021
AviSynth+ 3.7.0 (r3382, 3.7, i386) (3.7.0.0)

Number of frames: 7373
Length (hh:mm:ss.ms): 00:04:06.012
Frame width: 488
Frame height: 360
Framerate: 29.970 (30000/1001)
Colorspace: YV12
Audio channels: 2
Audio bits/sample: 16
Audio sample rate: 44100
Audio samples: 10849148

Frames processed: 7373 (0 - 7372)
FPS (min | max | average): 237.4 | 466.8 | 406.5
Process memory usage (max): 38 MiB
Thread count: 9
CPU usage (average): 26.8%

Time (elapsed): 00:00:18.136

So a pretty fair speed improvement.
However, is not exaclty the same as 9 coord boxblur, probably due to being rounded to int in both vertical and horizontal filtering.
If you do a subtract of 9 coord Expr() thingy, and the 2 * 1D whotsit, it does not show any obvious difference,
and if put through ClipDelta(AMP=True) then the rounding differences look pretty much like random noise,
except in occasional areas of the original clip which were flat in color.

EDIT: Assuming that original 8 coord try at BoxBlur was actaully a DeDot thingy (is it)

# De-Dot
Avisource("D:\Parade.avi")
Expr("x[0,-1] x[0,1] + 2 /")
Expr("x[-1,0] x[1,0] + 2 /")
return Last

Do we have a HBD UnDot ? [if it is an undot/dedot, then maybe one for real.finder collection]

[Dogway]

Wow, pretty cool! Yes, I thought on shaving some speed with the 8 coords but was wrong.

I think here we might have a masktools2 killer. For me the speed improvement defeats the subtle change.
In any case I have yet to test in HBD.


Here is mt_expand(mode="both"), slower in any case:
Expr("x[0,0] x[0,-1] x[0,1] x[-1,0] x[1,0] + + + +","")

mt_inpand(mode="both")
Expr("x[0,0] x[0,-1] x[0,1] x[-1,0] x[1,0] min min min min","")


I'm going to create a library. EDIT: done, check here -> ExTools (https://github.com/Dogway/Avisynth-Scripts/blob/master/ExTools.avsi)

[kedautinh12]

new mod ver of dogway v2.193mod support > avs+ 2724
https://github.com/Dogway/Avisynth-Scripts/blob/master/Mods/LSFmod.v2.193m.avsi

[Boulder]

Do we have a HBD UnDot ? [if it is an undot/dedot, then maybe one for real.finder collection]

I believe RemoveGrain(1) is the same as Undot().

[Reel.Deel]

I believe RemoveGrain(1) is the same as Undot().

Indeed it is :)