fft3dGPU 0.8.2 - Page 47

poisondeathray · 24th March 2021, 17:47

Quote:

Originally Posted by Arx1meD

When I used FFT3dGPU v0.8.6 I noticed that when plane = 0 or 4 and sharpen > 0 a small mesh appears. Is this a bug in the filter or not?

probably a bug; the same call with fft3dfilter does not produce the grid

bt=1 (spatial only) instead of bt=3 does not produce grid either , at least on that single still image that was looped with imagesource

Fjord · 3rd May 2021, 09:56

Quote:

Originally Posted by Arx1meD

When I used FFT3dGPU v0.8.6 I noticed that when plane = 0 or 4 and sharpen > 0 a small mesh appears. Is this a bug in the filter or not?

Quote:

Originally Posted by poisondeathray

probably a bug; the same call with fft3dfilter does not produce the grid

bt=1 (spatial only) instead of bt=3 does not produce grid either , at least on that single still image that was looped with imagesource

@Arx1meD, I suggest that you open an issue on Pinterf's FFT3dGPU github site: https://github.com/pinterf/FFT3dGPU/issues, to make sure he is aware of this.

Dogway · 18th September 2021, 12:17

Is there a way to match FFT3DGPU to fft3dfilter? The following calls give me very different results FFT3DGPU being worse with blurrier lines and also some halo. precision=0 is ok if the pipeline is in HBD.

Code:

fft3dfilter(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=12,bh=12,ow=6,oh=6,bt=4,plane=4,ncpu=4)
FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=12,bh=12,ow=6,oh=6,bt=4,plane=4,precision=0)

StainlessS · 18th September 2021, 14:39

Doggy, methinks that FFT3dFilter and the gpu version will often if not always give differing results, I seem to remember that
one good reason not to use gpu version was thats it uses less precision, and maybe truncation rather than rounding.
Although visible blurring and halo does sound disturbing.

pinterf · 18th September 2021, 14:58

Less precision means that in 8 bits the default precision is using only 16 bit floating point type (half). At least for part of the algorithm. I suppose this was chosen because of speed reasons. For 10+ bits I made it using 32 bit float always as I remember. See docs.
I cannot tell you about other comparisons, I never used the plugin. I guess fft3dgpu was based on an earlier version of fft3dfilter but I don't know when this happened.

Reel.Deel · 18th September 2021, 18:41

Quote:

Originally Posted by Dogway

Is there a way to match FFT3DGPU to fft3dfilter?...

Does changing the values of bw,bh,ow,oh to a power of 2 change anything?

Quote:

Originally Posted by tsp

fft3dfilter 1.8.5 that the current version of fft3dgpu is based on (expect bw,bh,ow,oh value as they need to be a power of two in fft3dgpu)

Also there have been reports of decreased quality in fft3dgpu when bt=4 and/or precision=2 (in pinterf's mod precision is always set to 2 when bitdepth is greater than 8).

--------

Quote:

Originally Posted by pinterf

I guess fft3dgpu was based on an earlier version of fft3dfilter but I don't know when this happened.

According to tsp, fft3dgpu is based on fft3dfilter v1.8.5 (post #683). But aside from delaho, it seems that most changes to fft3dfilter after v1.8.5 were optimizations and bug fixes:

Code:

v1.8.5 - 2005/12/04 - fixed bug with memory leakage (thanks to tsp).
v1.9.0 - 2006/04/25 - added dehalo options; corrected sharpen mode; 
                    - re-enabled SSE optimization for degrid=0; 
                    - added SSE optimization for bt=3,-1 with degrid>0 (faster by 15%)
v1.9.1 - 2006/05/10 - added SSE optimization for bt=4 with degrid>0 (faster by 30%).
v1.9.2 - 2006/09/11 - changed default bw=bh=32; added mode bt=5
v2.0.0 - 2006/11/29 - added experimental internal motion compensation mode; 
                    - window reorganized; internal multithread support
v2.1.0 - 2007/01/17 - removed (temporary ?) experimental internal motion compensation mode.
v2.1.1 - 2007/02/20 - fixed bug with bw not mod 4 (restored v1.9.2 method).

Dogway · 18th September 2021, 21:48

Thanks Reel.Deel, has some time to check what's going on.

So power of 2 only, 4, 8, 16... that leaves out 12.

Code:

FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=16,bh=16,ow=8,oh=8,bt=4,plane=0,precision=0)
fft3dfilter(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=16,bh=16,ow=8,oh=8,bt=4,plane=0,ncpu=2)

Yes, this is a match.

This one though isn't:

Code:

FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=8,bh=8,ow=4,oh=4,bt=2,plane=0,precision=0)
fft3dfilter(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=8,bh=8,ow=4,oh=4,bt=2,plane=0,ncpu=2)

This call just outputs garbage (bug?)

Code:

FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=16,bh=16,ow=6,oh=6,bt=2,plane=0,precision=0)

And here a comparison between fft3dfilter and FFT3DGPU, very different, background is blurred more on fft3dfilter, and lineart is cleaner. On FFT3DGPU you can see some halo and ringing, easily seen on the inner side of the glasses.

Code:

FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=12,bh=12,ow=4,oh=4,bt=2,plane=0,precision=0)

fft3dfilter

FFT3DGPU

minusthebear · 3rd May 2022, 02:42

I'm a bit at a loss for how to install and use FFT3dGPU. Does it replace FFT3d entirely? Does it only work with some plugins?

I'm looking to speed up my workflow in any way possible, can someone please direct me to a good guide/resource for using and installing FFT3dGPU?

Thanks!!

Reel.Deel · 3rd May 2022, 03:24

Quote:

Originally Posted by minusthebear

I'm a bit at a loss for how to install and use FFT3dGPU. Does it replace FFT3d entirely? Does it only work with some plugins?

I'm looking to speed up my workflow in any way possible, can someone please direct me to a good guide/resource for using and installing FFT3dGPU?

Thanks!!

To autoload FFT3dGPU, copy both fft3dgpu.hlsl and FFT3dGPU.dll into the appropriate folder. Or load the plugin manually with LoadPlugin, just make sure fft3dgpu.hlsl is in the same folder as the dll. It does not replace FFT3DFilter since the function names are different. You can have both plugins installed/loaded and even used in the same script. Not sure what you mean by "Does it only work with some plugins?" but it works like a normal plugin and it is not dependent on what other plugins you use or not.

Lan4 · 6th May 2023, 21:37

The bug happened several times. At first, encoding with the FFT3DGPU filter proceeds normally. Then the encoding time starts to increase, from several seconds to several minutes. This problem has been around for about 2 months. This may be due to a video driver update. Has anyone had the same problem? And what could be its reason?

18th September 2021, 12:17	#923 \| Link
Dogway Registered User Join Date: Nov 2009 Posts: 2,361	Is there a way to match FFT3DGPU to fft3dfilter? The following calls give me very different results FFT3DGPU being worse with blurrier lines and also some halo. precision=0 is ok if the pipeline is in HBD. Code: fft3dfilter(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=12,bh=12,ow=6,oh=6,bt=4,plane=4,ncpu=4) FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=12,bh=12,ow=6,oh=6,bt=4,plane=4,precision=0) __________________ i7-4790K@Stock::GTX 1070] AviSynth+ filters and mods on GitHub + Discussion thread

18th September 2021, 14:39	#924 \| Link
StainlessS HeartlessS Usurer Join Date: Dec 2009 Location: Over the rainbow Posts: 10,980	Doggy, methinks that FFT3dFilter and the gpu version will often if not always give differing results, I seem to remember that one good reason not to use gpu version was thats it uses less precision, and maybe truncation rather than rounding. Although visible blurring and halo does sound disturbing. __________________ I sometimes post sober. StainlessS@MediaFire ::: AND/OR ::: StainlessS@SendSpace "Some infinities are bigger than other infinities", but how many of them are infinitely bigger ???

18th September 2021, 14:58	#925 \| Link
pinterf Registered User Join Date: Jan 2014 Posts: 2,314	Less precision means that in 8 bits the default precision is using only 16 bit floating point type (half). At least for part of the algorithm. I suppose this was chosen because of speed reasons. For 10+ bits I made it using 32 bit float always as I remember. See docs. I cannot tell you about other comparisons, I never used the plugin. I guess fft3dgpu was based on an earlier version of fft3dfilter but I don't know when this happened. __________________ AviSynth+ on github, Other repos: RgTools, Masktools2, MvTools2, TIVTC, Average

18th September 2021, 21:48	#927 \| Link
Dogway Registered User Join Date: Nov 2009 Posts: 2,361	Thanks Reel.Deel, has some time to check what's going on. So power of 2 only, 4, 8, 16... that leaves out 12. Code: FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=16,bh=16,ow=8,oh=8,bt=4,plane=0,precision=0) fft3dfilter(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=16,bh=16,ow=8,oh=8,bt=4,plane=0,ncpu=2) Yes, this is a match. This one though isn't: Code: FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=8,bh=8,ow=4,oh=4,bt=2,plane=0,precision=0) fft3dfilter(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=8,bh=8,ow=4,oh=4,bt=2,plane=0,ncpu=2) This call just outputs garbage (bug?) Code: FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=16,bh=16,ow=6,oh=6,bt=2,plane=0,precision=0) And here a comparison between fft3dfilter and FFT3DGPU, very different, background is blurred more on fft3dfilter, and lineart is cleaner. On FFT3DGPU you can see some halo and ringing, easily seen on the inner side of the glasses. Code: FFT3DGPU(sigma=16,sigma2=6,sigma3=1,sigma4=0,bw=12,bh=12,ow=4,oh=4,bt=2,plane=0,precision=0) fft3dfilter FFT3DGPU __________________ i7-4790K@Stock::GTX 1070] AviSynth+ filters and mods on GitHub + Discussion thread

3rd May 2022, 02:42	#928 \| Link
minusthebear Registered User Join Date: Oct 2018 Posts: 10	I'm a bit at a loss for how to install and use FFT3dGPU. Does it replace FFT3d entirely? Does it only work with some plugins? I'm looking to speed up my workflow in any way possible, can someone please direct me to a good guide/resource for using and installing FFT3dGPU? Thanks!!

6th May 2023, 21:37	#930 \| Link
Lan4 Registered User Join Date: Dec 2022 Posts: 64	The bug happened several times. At first, encoding with the FFT3DGPU filter proceeds normally. Then the encoding time starts to increase, from several seconds to several minutes. This problem has been around for about 2 months. This may be due to a video driver update. Has anyone had the same problem? And what could be its reason?