TemporalDegrainV2

[ChaosKing]

Bonus: degrainTR 5-6

  148.04684 47280ms degrainTR=6 postFFT=3 postSigma=4
  148.01166 42710ms degrainTR=5 postFFT=3 postSigma=4
  147.94945 41310ms degrainTR=5 postFFT=3 postSigma=3
  147.78264 37160ms degrainTR=6 postFFT=1 postSigma=4
  147.7377 36780ms degrainTR=6 postFFT=1 postSigma=3
  147.73164 30890ms degrainTR=5 postFFT=1 postSigma=4
  147.67194 30840ms degrainTR=5 postFFT=1 postSigma=3
  147.44595 30810ms degrainTR=5 postFFT=1 postSigma=2
  144.64229 27510ms degrainTR=5 postFFT=0 postSigma=1

@zorr I had set it to -iters 30 by mistake but it ran 32 times nonetheless
LOG (x64, 0.9.8 beta)

Output 1:
  name  ssim
  type  FLOAT
  goal  MAX
  unit
  round 0
Output 2:
  name  time
  type  TIME
  goal  MIN
  unit  ms
  round 10

Found following optimizable parameters:
  # optimize degrainTR = _n_ | 5..6 | degrainTR
  # optimize postFFT = _n_ | 0,1,3 | postFFT
  # optimize postSigma = _n_ | 1..4 | postSigma
found 3 parameters to optimize

Running SPEA2
* 1 / 30 : 145.00519 56410ms degrainTR=6 postFFT=0 postSigma=1
+ 2 / 30 : 145.00519 55430ms degrainTR=6 postFFT=0 postSigma=3
+ 3 / 30 : 144.64229 28650ms degrainTR=5 postFFT=0 postSigma=2
  4 / 30 : 144.64229 28710ms degrainTR=5 postFFT=0 postSigma=1
* 5 / 30 : 147.81573 42460ms degrainTR=5 postFFT=3 postSigma=2
+ 6 / 30 : 146.67145 31380ms degrainTR=5 postFFT=1 postSigma=1
  7 / 30 : 145.00519 56330ms degrainTR=6 postFFT=0 postSigma=4
+ 8 / 30 : 144.64229 28330ms degrainTR=5 postFFT=0 postSigma=2
Parameter sensitivity estimation with 64 result combinations
  -> degrainTR 0,542 postFFT 1,546 postSigma 0,990
Mutating 2 params by 22,0 %
+ 9 / 30 : 147.78264 37160ms degrainTR=6 postFFT=1 postSigma=4
+ 10 / 30 : 147.44595 30810ms degrainTR=5 postFFT=1 postSigma=2
  11 / 30 : 144.64229 28410ms degrainTR=5 postFFT=0 postSigma=4
  12 / 30 : 145.00519 53030ms degrainTR=6 postFFT=0 postSigma=1
  13 / 30 : 147.45505 42150ms degrainTR=5 postFFT=3 postSigma=1
  14 / 30 : 146.67145 30820ms degrainTR=5 postFFT=1 postSigma=1
  15 / 30 : 146.8749 36550ms degrainTR=6 postFFT=1 postSigma=1
+ 16 / 30 : 144.64229 27580ms degrainTR=5 postFFT=0 postSigma=2
Parameter sensitivity estimation with 256 result combinations
  -> degrainTR 0,556 postFFT 1,403 postSigma 0,968
Mutating 1 params by 14,0 %
mutation failed 1000 times, increasing mutation count to 2
+ 17 / 30 : 144.64229 27510ms degrainTR=5 postFFT=0 postSigma=1
  18 / 30 : 145.00519 52870ms degrainTR=6 postFFT=0 postSigma=2
  19 / 30 : 147.81573 42490ms degrainTR=5 postFFT=3 postSigma=2
  20 / 30 : 144.64229 27880ms degrainTR=5 postFFT=0 postSigma=3
+ 21 / 30 : 147.7377 36780ms degrainTR=6 postFFT=1 postSigma=3
  22 / 30 : 147.54588 36850ms degrainTR=6 postFFT=1 postSigma=2
+ 23 / 30 : 147.67194 30840ms degrainTR=5 postFFT=1 postSigma=3
  24 / 30 : 147.57674 48110ms degrainTR=6 postFFT=3 postSigma=1
Parameter sensitivity estimation with 256 result combinations
  -> degrainTR 0,559 postFFT 1,282 postSigma 1,054
Mutating 1 params by 6,0 %
mutation failed 1000 times, increasing mutation count to 2
mutation failed 1000 times, increasing mutation count to 2
mutation failed 1000 times, increasing mutation count to 2
mutation failed 2000 times, increasing mutation count to 3
  25 / 30 : 145.00519 53010ms degrainTR=6 postFFT=0 postSigma=3
+ 26 / 30 : 147.73164 30890ms degrainTR=5 postFFT=1 postSigma=4
* 27 / 30 : 147.94945 41310ms degrainTR=5 postFFT=3 postSigma=3
  28 / 30 : 147.88573 46830ms degrainTR=6 postFFT=3 postSigma=2
* 29 / 30 : 148.01166 42710ms degrainTR=5 postFFT=3 postSigma=4
  30 / 30 : 147.99655 47120ms degrainTR=6 postFFT=3 postSigma=3
* 31 / 30 : 148.04684 47280ms degrainTR=6 postFFT=3 postSigma=4
  32 / 30 : 145.00519 53230ms degrainTR=6 postFFT=0 postSigma=4
Parameter sensitivity estimation with 256 result combinations
  -> degrainTR 0,566 postFFT 1,161 postSigma 1,152
Total execution time: 1339735ms
Pareto front:
  148.04684 47280ms degrainTR=6 postFFT=3 postSigma=4
  148.01166 42710ms degrainTR=5 postFFT=3 postSigma=4
  147.94945 41310ms degrainTR=5 postFFT=3 postSigma=3
  147.78264 37160ms degrainTR=6 postFFT=1 postSigma=4
  147.7377 36780ms degrainTR=6 postFFT=1 postSigma=3
  147.73164 30890ms degrainTR=5 postFFT=1 postSigma=4
  147.67194 30840ms degrainTR=5 postFFT=1 postSigma=3
  147.44595 30810ms degrainTR=5 postFFT=1 postSigma=2
  144.64229 27510ms degrainTR=5 postFFT=0 postSigma=1
Running SPEA2
...

[ErazorTT]

What is actually the script you are using?

[ChaosKing]

https://forum.doom9.org/showpost.php...3&postcount=98

[ChaosKing]

Now I tested 100 frames from tears of steel, this scene https://www.youtube.com/watch?v=R6Ml...youtu.be&t=271

And here I noticed also some halos. You can see it very good on his hair in #2 http://screenshotcomparison.com/comparison/123735/

My script

ffms2("TearsOfSteelFull12min_1080p_24fps_27qp_1474kbps_GPSNR_42.29_HM11.mkv").ConvertBits(8).bicubicresize(880,500).trim(6460,6560)
source = last
last=source.AddGrain(40, 0, 0, seed=2)
degrainTR = 5					# optimize degrainTR = _n_ | 1..5 | degrainTR
postFFT = 3					# optimize postFFT = _n_ | 0,1,3 | postFFT
postSigma = 1					# optimize postSigma = _n_ | 1..4 | postSigma
denoised = TemporalDegrain2(degrainTR=degrainTR, postFFT=postFFT, postSigma=postSigma)
# no Prefetch, i forgot to add it :)

@zorr
- Evaluate mode still shows version 0.9.8 instead of 0.9.9
- Worst result is still wrong in 0.9.9 and the visualisation also I guess (96.75 value is cut off)

AvisynthOptimizer v0.9.8-beta

Arguments
  mode = evaluate
  autorefresh = true

Running evaluation for log(s) .\avs_version (2018-11-04 10-10-30) optimize spea2 pop-8 mutcount-60%-1 mutamount-0.3-0.01 cross-0.1-20 iters-70 run-01.log
Using these settings:
ARGUMENT       DESCRIPTION                     VALUE
-vismode       visualization mode              single
-scripts       write scripts                   none
-globalpareto  show series pareto              true
-autorefresh   autorefresh                     true
-highlight     highlight latest                10
-multipareto   show multiple pareto fronts     0 (OFF)
-top           zoom into top x% results        20
-groupby       group by parameter              (OFF)
-maxgroups     group by max groups             N/A
-groupmin      group by min value              N/A
-groupmax      group by max value              N/A
-animation     animation iterations per frame  0 (OFF)

Reading log file .\avs_version (2018-11-04 10-10-30) optimize spea2 pop-8 mutcount-60%-1 mutamount-0.3-0.01 cross-0.1-20 iters-70 run-01.log
Single run results: 64
Series results: 0

Run 1 best: 99.42768 51190 degrainTR=5 postFFT=3 postSigma=1

Best result: 99.42768
Worst result: 99.42768 <----- NOPE!

Pareto front:
  99.42768 51190 degrainTR=5 postFFT=3 postSigma=1
  99.42644 44710 degrainTR=4 postFFT=3 postSigma=1
  99.3991 38630 degrainTR=3 postFFT=3 postSigma=1
  99.39909 38140 degrainTR=3 postFFT=3 postSigma=1
  99.37821 37100 degrainTR=5 postFFT=1 postSigma=1
  99.35609 30730 degrainTR=4 postFFT=1 postSigma=2
  99.3523 30310 degrainTR=4 postFFT=1 postSigma=1
  99.34066 24040 degrainTR=3 postFFT=1 postSigma=2
  99.26441 17570 degrainTR=2 postFFT=1 postSigma=2
  99.15682 17560 degrainTR=2 postFFT=1 postSigma=4
  99.067345 17490 degrainTR=2 postFFT=1 postSigma=1
  99.05565 11640 degrainTR=1 postFFT=1 postSigma=4
  96.75158 8700 degrainTR=1 postFFT=0 postSigma=2

98.24987 14380 degrainTR=2 postFFT=0 postSigma=3 
99.31801 31720 degrainTR=2 postFFT=3 postSigma=4 
99.34313 32060 degrainTR=2 postFFT=3 postSigma=2 
99.3523 30310 degrainTR=4 postFFT=1 postSigma=1 
96.75158 8700 degrainTR=1 postFFT=0 postSigma=2 
99.39007 44140 degrainTR=4 postFFT=3 postSigma=2 
96.75158 8720 degrainTR=1 postFFT=0 postSigma=3 
98.24987 14580 degrainTR=2 postFFT=0 postSigma=4 
98.72428 21350 degrainTR=3 postFFT=0 postSigma=3 
99.052536 11670 degrainTR=1 postFFT=1 postSigma=3 
99.02302 25530 degrainTR=1 postFFT=3 postSigma=2 
99.05565 11640 degrainTR=1 postFFT=1 postSigma=4 
98.72428 21600 degrainTR=3 postFFT=0 postSigma=4 
98.12291 11930 degrainTR=1 postFFT=1 postSigma=1 
99.05177 34090 degrainTR=5 postFFT=0 postSigma=1 
99.15682 17560 degrainTR=2 postFFT=1 postSigma=4 
96.75158 8820 degrainTR=1 postFFT=0 postSigma=1 
98.878456 12010 degrainTR=1 postFFT=1 postSigma=2 
99.28696 32500 degrainTR=2 postFFT=3 postSigma=1 
99.24217 17770 degrainTR=2 postFFT=1 postSigma=3 
99.37821 37100 degrainTR=5 postFFT=1 postSigma=1 
96.75158 8750 degrainTR=1 postFFT=0 postSigma=4 
98.94025 27930 degrainTR=4 postFFT=0 postSigma=1 
98.24987 14310 degrainTR=2 postFFT=0 postSigma=2 
99.27137 24320 degrainTR=3 postFFT=1 postSigma=3 
99.352585 36270 degrainTR=5 postFFT=1 postSigma=2 
98.72428 21830 degrainTR=3 postFFT=0 postSigma=2 
99.42768 51190 degrainTR=5 postFFT=3 postSigma=1 
99.188416 26020 degrainTR=1 postFFT=3 postSigma=4 
99.35609 30730 degrainTR=4 postFFT=1 postSigma=2 
99.37805 51230 degrainTR=5 postFFT=3 postSigma=2 
99.14046 25910 degrainTR=1 postFFT=3 postSigma=3 
99.2798 24180 degrainTR=3 postFFT=1 postSigma=1 
99.15887 24290 degrainTR=3 postFFT=1 postSigma=4 
99.34593 44000 degrainTR=4 postFFT=3 postSigma=3 
99.42644 44710 degrainTR=4 postFFT=3 postSigma=1 
99.05177 33440 degrainTR=5 postFFT=0 postSigma=2 
98.72428 21520 degrainTR=3 postFFT=0 postSigma=1 
99.38929 38540 degrainTR=3 postFFT=3 postSigma=2 
99.34066 24040 degrainTR=3 postFFT=1 postSigma=2 
99.39909 38140 degrainTR=3 postFFT=3 postSigma=1 
98.94025 28150 degrainTR=4 postFFT=0 postSigma=2 
99.35622 38870 degrainTR=3 postFFT=3 postSigma=3 
98.24987 14870 degrainTR=2 postFFT=0 postSigma=1 
99.32733 51160 degrainTR=5 postFFT=3 postSigma=3 
98.688225 25490 degrainTR=1 postFFT=3 postSigma=1 
99.30322 44120 degrainTR=4 postFFT=3 postSigma=4 
99.26441 17570 degrainTR=2 postFFT=1 postSigma=2 
98.94025 27280 degrainTR=4 postFFT=0 postSigma=4 
99.32066 38310 degrainTR=3 postFFT=3 postSigma=4 
99.067345 17490 degrainTR=2 postFFT=1 postSigma=1 
99.26495 30370 degrainTR=4 postFFT=1 postSigma=3 
99.280716 51640 degrainTR=5 postFFT=3 postSigma=4 
99.24911 36540 degrainTR=5 postFFT=1 postSigma=3 
99.1392 30790 degrainTR=4 postFFT=1 postSigma=4 
99.33777 31310 degrainTR=2 postFFT=3 postSigma=3 
99.05177 33620 degrainTR=5 postFFT=0 postSigma=3 
99.115715 36940 degrainTR=5 postFFT=1 postSigma=4 
99.05177 33770 degrainTR=5 postFFT=0 postSigma=4 
98.94025 27360 degrainTR=4 postFFT=0 postSigma=3 
98.12291 11790 degrainTR=1 postFFT=1 postSigma=1 
99.15887 24470 degrainTR=3 postFFT=1 postSigma=4 
98.94025 28210 degrainTR=4 postFFT=0 postSigma=2 
99.3991 38630 degrainTR=3 postFFT=3 postSigma=1

[tormento]

Originally Posted by ErazorTT

Well considering that the result from ChaosKing are more than an order of magnitude faster than yours, and yours are exactly as fast as both TempralDegrain v1 and v2 on my crappy laptop I would guess that there is something wrong for you. I don’t know what simple launcher is supposed to do but apparently there goes something wrong.

Or probably you missed the fact I am running the script on a 1080p source and you both on 720p or DVD resolution? Which encoder you feed at the end? I am using x264 --preset slow.

To clear things up, I am now using AVSMeter64 to get script results with no encoder around.

@ChaosKing to answer you about KNLMeansCL performance, I run a simple KNLMeansCL() on 1080p source:

AVSMeter64.exe sacrificio_KNLMeansCL.avs

AVSMeter 2.8.6 (x64) - Copyright (c) 2012-2018, Groucho2004
AviSynth+ 0.1 (r2728, MT, x86_64) (0.1.0.0)

Number of frames:                 1750
Length (hh:mm:ss.ms):     00:01:12.990
Frame width:                      1920
Frame height:                     1032
Framerate:                      23.976 (24000/1001)
Colorspace:                       YV12

Frames processed:               1750 (0 - 1749)
FPS (min | max | average):      2.001 | 17.93 | 12.41
Process memory usage (max):     198 MiB
Thread count:                   24
CPU usage (average):            9.6%

Time (elapsed):                 00:02:21.065

Anyway... as I told twice... SMDegrain with KNLMeansCL prefilter works with 10x performance.

[ChaosKing]

tears of steel 1080p x64 test:

#0.82fps
TemporalDegrain2(degrainTR=3, postFFT=3, postSigma=1)

#4fps
TemporalDegrain2(degrainTR=3, postFFT=3, postSigma=1, fftthreads=8)
Prefetch(8)

KNLMeansCL alone runs at 27fps on a 1080p source


I'm testing the script performance, so I don't feed to any encoder.

[tormento]

Originally Posted by ChaosKing

tears of steel 1080p x64 test:

Please try these settings to mantain some uniformity.

EDIT:

#4fps
TemporalDegrain2(degrainTR=3, postFFT=3, postSigma=1, fftthreads=8)
Prefetch(8)

@ErazorTT

1.86 fps here on i7-2600k + GTX 760. I can't really see one order of magnitude.

[ChaosKing]

All with Prefetch(8). Now tested with avsmeter64, 300 frames, 1080p source - ffms2("TearsOfSteelFull12min_1080p_24fps_27qp_1474kbps_GPSNR_42.29_HM11.mkv").trim(500,500+300)
AviSynth+ r2728, all plugins updated
Ryzen 1700 @ 3.7ghz, Rx480, 16gb ram 3000mhz

TemporalDegrain2(degrainTR=4,postFFT=0,postSigma=3,postDither=0) #5.15
TemporalDegrain2(degrainTR=4,postFFT=0,postSigma=3,postDither=0, fftthreads=8) #5.5
TemporalDegrain2(degrainTR=4,postFFT=3,postSigma=3,postDither=0) # 2.42
TemporalDegrain2(degrainTR=4,postFFT=3,postSigma=3,postDither=0, fftthreads=8) # 3.52
TemporalDegrain2(degrainTR=4,postFFT=4,postSigma=3,postDither=0) #4.23
TemporalDegrain2(degrainTR=4,postFFT=4,postSigma=3,postDither=0, fftthreads=8) #4.62

Have you the newest fftw lib installed?

[ErazorTT]

Originally Posted by tormento

Or probably you missed the fact I am running the script on a 1080p source and you both on 720p or DVD resolution?

My performance measurement (here) was using a 1080p source. Yielding around 1.7fps without postFFT on a dual core haswell laptop, i5-4288U.

Originally Posted by tormento

Which encoder you feed at the end? I am using x264 --preset slow.

For measuring the speed of the script I use x264 with preset super fast, since I want to measure the differences between scripts not the impact of the encoder. However I very much see the point of measuring both, since in the end one has to do both at a time.

Originally Posted by tormento

Anyway... as I told twice... SMDegrain with KNLMeansCL prefilter works with 10x performance.

Yes.. Apparently there is somewhere something wrong but I just dont know where... Without being able to test it myself, there nothing more I can do than guessing..

[tormento]

Originally Posted by ErazorTT

For measuring the speed of the script I use x264 with preset super fast

Much better to feed directly to avsmeter

Originally Posted by ErazorTT

Yes.. Apparently there is somewhere something wrong but I just dont know where... Without being able to test it myself, there nothing more I can do than guessing..

We are here to help

[tormento]

Originally Posted by ChaosKing

Have you the newest fftw lib installed?

I am using 3.3.8 AVX

[ErazorTT]

Originally Posted by tormento

Much better to feed directly to avsmeter

Thats a good suggestion!

Originally Posted by tormento

I am using 3.3.8 AVX

Oh wow I test all the version from your link and the fastest version was "simd128+256" which made dfttest alsmot 10% faster!

EDIT:
Hm however fft3dfilter does not work with this version of fftw lib, it produces access violations...

[zorr]

Originally Posted by ChaosKing

@zorr
- Evaluate mode still shows version 0.9.8 instead of 0.9.9

Yes, I forgot to update the version number. It has happened a couple of times before.

Originally Posted by ChaosKing

- Worst result is still wrong in 0.9.9 and the visualisation also I guess (96.75 value is cut off)

The visualization by default "zooms in" into the best 20% of results. You can change that with argument -top, for example -top 50 would show the top 50% of results. If you want the zooming completely disabled use -top 100 but that doesn't usually look very good. Perhaps -top 99 will do what you want.

I wasn't able to reproduce the problem with the worst result, I also noticed that in your log the Series has zero results, probably has something to do with that.

Single run results: 64
Series results: 0

Can you give me the exact command line you use to call the optimizer? Are you calling it from the <optimizer>/bin directory or somewhere else? Where are your log files located at?

You also might try the latest version which has one path related problem fixed.

@zorr I had set it to -iters 30 by mistake but it ran 32 times nonetheless

That's a limitation of the algorithms, they evaluate a whole population until they can stop. So the iteration count is always divisible by the population count.

[tormento]

Originally Posted by ErazorTT

Hm however fft3dfilter does not work with this version of fftw lib, it produces access violations...

Try to report in that thread.

[zorr]

Originally Posted by ChaosKing

Run 1 best: 99.42768 51190 degrainTR=5 postFFT=3 postSigma=1

Best result: 99.42768
Worst result: 99.42768 <----- NOPE!

Ah, I forgot what the worst result means. It's actually the best result of the worst run. When there is only one run then best == worst. The idea is that you can see how different results the runs give. I will change the labels to "Best run" and "Worst run" to make it easier to understand.

I have tracked down the Series results being 0, will be fixed in the next version.

And more on topic: it might be a good idea to test if there is any difference in quality when running in multithreaded mode versus single threaded. There shouldn't be, but it doesn't hurt to make sure...

[Zetti]

Originally Posted by ErazorTT

Thats a good suggestion!



Oh wow I test all the version from your link and the fastest version was "simd128+256" which made dfttest alsmot 10% faster!

EDIT:
Hm however fft3dfilter does not work with this version of fftw lib, it produces access violations...

The problem is fftw lib and not fft3dfilter.

See here: https://forum.doom9.org/showthread.p...82#post1857182

[tormento]

Originally Posted by ErazorTT

I want to share an updated version of the script TemporalDegrain.

What about a switch to use FFT3DGPU instead of FFT3DFilter? I'd like to see speed/quality differences. Problem is AFAIK we don't have a x64 version of it.

It could be nice to have a new compiled version from Pinterf, who already dealt with FFT3DFilter or (my dream) a FFT3DCL from Khanattila

[ErazorTT]

Originally Posted by tormento

What about a switch to use FFT3DGPU instead of FFT3DFilter?

FFT is used in two instances in the script. You can already change both to using FFT3DGPU via function arguments. One by changing limitFFT to 2 and the other by changing postFFT to 2. PostFFT also has the options 3 (dfttest) and 4 (meansCL) which do not make sense for limitFFT.

Originally Posted by tormento

I'd like to see speed/quality differences.

Yeah me too but I just have an integrated Intel GPU. Quality wise I would expect FFT3DGPU to behave very close to FFT3Dfilter. And especially for limitFFT there is will be no difference in quality, though perhaps very significantly in speed.

That said, FFT3DGPU has not seen any new version in ages while there have been updates to FFT3Dfilter.

[tormento]

Originally Posted by ErazorTT

One by changing limitFFT to 2

Ok... I did not notice that. Now that you point it, I can see it in the script code. Now I only need to find a FFT3DGPU x64...

EDIT: Here.

Originally Posted by ErazorTT

That said, FFT3DGPU has not seen any new version in ages while there have been updates to FFT3Dfilter.

A real pity. Just imagine to have a FFT3DCL...

[tormento]

TemporalDegrain2(degrainTR=4,postFFT=0,postSigma=3,postDither=-1,limitFFT=1) 1.70 fps
TemporalDegrain2(degrainTR=4,postFFT=0,postSigma=3,postDither=-1,limitFFT=2) 1.66 fps
TemporalDegrain2(degrainTR=4,postFFT=4,postSigma=3,postDither=-1,limitFFT=1) 1.12 fps
TemporalDegrain2(degrainTR=4,postFFT=4,postSigma=3,postDither=-1,limitFFT=2) 1.03 fps

So there should be a bottleneck somewhere...