fft3dGPU 0.8.2 [Archive] - Page 15

grennis

31st January 2008, 06:08

That looks liek you are using it in QMF, may very well be this filter not liking to be used within QMF.

I have no idea what QMF is, so I don't think I'm doing that :)

I have also noticed that I get D3D errors when using FFT3DGPU over remote desktop and disconnect/reconnect. There probably isn't much you can do about it?

Is the sigma value exactly the same as in the fft3dfilter? It seems like I need to use a higher sigma in fft3dgpu to get the same result.

rkalwaitis

7th February 2008, 16:58

TSP or someone else in the know.

I really like what I can do with this filter, Im very impressed by the quality I can achieve and results when compressing.

Just a few questions? I like to use x264. Are the cqm/cqf (Sharktooth's) matrixes necessary since the filter is so efficient?

I also seem to notice that the colors, (at least for me) tend to be a bit different? Is this a setting I'm dorking up? Perhaps a correction I can make? The colors seem to lend themselves to a slight blueish-green and the picture seems to brighten up a bit resulting in a slightly bleached look. I can tweak filter it and get back some of it. Is this another setting Im messing up!

Well anyways I like it. I wonder if using this filter with say 'snow' or 'msu' it would work faster?:stupid:

Adub

13th February 2008, 07:18

Matrices are not specifically designed for the removal of noise, that's what denoisers are for. Sure a matrix can be tuned to help in the removal of noise, but not like a denoiser can. So, yes, as long as the quality is better than not using one, keep using a matrix.

If you get better results without a matrix then don't use it. These instances have nothing to do with fft3dgpu however, as it is a denoiser. Denoise your source until it looks as good as you want it to, and then encode it (with or without a matrix, your choice).

As to the colors, it maybe that you need to add ColorMatrix() to your script. It depends on what your source is and what you are doing with it.

For instance, if you are converting a DVD to mp4 or mkv, then chances are you will need to edit your script to look something like this:

DGDECODE_Mpeg2Source("video.d2v", info=3)
Colormatrix(hints=true)
...other filters...

It's correct use depends on the source, as I said above.

rkalwaitis

14th February 2008, 07:17

Thanks for the info. Adding the color matrix as you suggested worked. It was a dv2 source. Lots to learn. I can't tell a difference with the matrix I was trying to use, so Ill just not use it. They are confusing to me, ulr, vlr, .....and so on. I thought they could help me crunch a bit more, but in the end it is not that important. I like the best possible picture, then worry about size.

Thanks again.

XBoy

19th February 2008, 23:03

I'm trying to compile the latest source for x64 but it seems to be missing a few files, mainly

filtersps.h, filtersps.cpp, dxinput.h, dxinput.cpp

Where do these come from??

Leak

20th February 2008, 13:28

filtersps.h, filtersps.cpp, dxinput.h, dxinput.cpp

Where do these come from??
The latter 2 definitely come from the DirectX SDK, and I wouldn't be surprised if the same goes for the former...

tsp

20th February 2008, 18:32

sorry I have updated the file to include all 4 files although dxinput might cause problem the the latest version of directx SDK as support for dxinput has been removed.

http://www.avisynth.org/tsp/fft3dgpu0.8.2a.exe

Mosfetov

23rd April 2008, 04:55

Oh, nice filter. It's very fast and gives high quality filtered image with minimum artifacts (even at high sigma). At 720p this filter works with my X800XT in realtime (framerate more than 24fps). There is only one restriction - size of GPURAM. It must be at least 256MB to work at 720p. Otherwise, speed will downgrade, because speed between GPU and aperture memory (System RAM) more-more slower than between GPU - GPURAM. But even in this case, FFT3DGPU works faster than FFT3DFilter. By default, i use bw=32, bh=32, ow=16, oh=16, or in some cases bw=64, bh=64, ow=32, oh=32. This takes approx up to 200+MB of GPURAM at 720p. If GPURAM only 128 MB, it's possible to use bw=64, bh=64, ow=16, oh=16. It reduces GPURAM utilization to 100-120MB at 720p and speed will boost. But quality will a little bit less.

Fizick, tsp, thank you for this useful filter! :thanks:

PS: My box - X800XT 256MB, Athlon XP 2300MHz (FSB 192MHz, sync to RAM), nF2U400, RAM 2048MB

murrsturr

28th April 2008, 03:01

Thanks to both TSP and Fizick!

This stuff is very much appreciated by myself and thousands of others.. and at this point in time... I speak for them all! ;)

Blue_MiSfit

29th April 2008, 21:33

Yeah seriously. I use this filter on practically every single stupid thing I encode - to kill chroma noise.

Amazingly effective, and basically free, since it runs on the GPU!

I like free!!!! Thanks guys!

~MiSfit

Quark.Fusion

19th June 2008, 15:20

I know you can assume that GPU-based filter isn't benefit from multithreading, but think about next filter in chain — while thread one doing it's part on GPU thread two still can do CPU part, and when it is thread two time for GPU, thread one can continue on next CPU part.

So is it possible to make thread-safe GPU filter? What about using CAL/CUDA?

Example:

MTsource(input)
MT("""
FFT3DGPU()
DeBlock_QED()
""")
in this example script can do deblock on first part while second is processing on GPU by using extra thread.

Why not just use mutex for synchrotisation?

Terka

20th June 2008, 09:07

2>tsp
on page 1 of this thread:
Have you plan to implement in GPU all my other plugins?
Only the FPU heavy filter

Could you please implement mvtools in GPU too? :)

Zep

22nd June 2008, 06:43

2>tsp
on page 1 of this thread:

Could you please implement mvtools in GPU too? :)

yeah MVtools is the main filter I wish was GPU also :)

Terka

22nd June 2008, 11:26

everybody wishes :p

Underground78

24th June 2008, 10:37

Hello,

Is there a way to take advantage of a CrossFire system with fft3dGPU ? I've made some searches in this thread but I can't find anything new about that ...

Underground78

27th June 2008, 14:11

Hello,

There is something that puzzled me a bit, when using a script without fft3dGPU, x264 first pass runs at ~100 fps with a CPU usage near 100%, the same script with fft3dGPU (default settings) runs at ~45 fps but with a CPU usage of about 35% and a GPU usage of about 50% so I assume it could be faster ... What could explain this ?

Thanks in advance

talen9

27th June 2008, 15:11

Do you have a dual-GPU video adapter? If so, then I think FFT3DGPU is using only one of the two GPU cores, and the processing speed of that one core is lower than the one you can obtain using your (dual/quad core) CPU ... your best using that, I'm afraid, if my analysis is correct.

Underground78

27th June 2008, 15:14

Quark.Fusion

28th June 2008, 04:16

I think that because GPU have many different blocks (ROPS, Shaders, etc) and fft3dGPU uses only vertex shaders. Look at shaders usage in GPU profiler.

Underground78

28th June 2008, 09:35

I think that because GPU have many different blocks (ROPS, Shaders, etc) and fft3dGPU uses only vertex shaders. Look at shaders usage in GPU profiler.

I will test this, thanks ! But where can I find this ? It seems to be a quite uncommon tool ...

Quark.Fusion

29th June 2008, 15:21

For nVidia card try NVIDIA PerfHUD and enable support in fft3dGPU, don't know about Radeons profiling. BTW, where you look at GPU usage?

Underground78

29th June 2008, 15:58

For nVidia card try NVIDIA PerfHUD and enable support in fft3dGPU, don't know about Radeons profiling. BTW, where you look at GPU usage?

I have Radeon cards, I look at GPU usage in GPU-Z ...

Quark.Fusion

30th June 2008, 12:15

Where it show that? Or it's supported only for Radeons?

Underground78

30th June 2008, 12:25

It's in the "Sensors" tab but it may be not always available ... I'm not sure my problem is only related to vertex shaders because it seems that using TimeCodec I can obtain a GPU usage near 80% ...

Quark.Fusion

1st July 2008, 11:00

If I understand GPU architecture right, then it isn't necessary to reach 100% of GPU usage to get optimum performance. I'm running folding@home at GPU and my current problem is what clocks to underclock to lower noise level and don't hurt folding performance too much :)

Quark.Fusion

1st July 2008, 11:48

Also there can be GPU idle sensor — do you have that? If it means time when GPU doesn't do anything then it's must be possible to optimize GPU-CPU code. And I spotted that fft3dGPU spend much time in GPU driver — maybe moving to CAL/CUDA can also speed up things.

lexor

4th July 2008, 16:12

Sup everyone, I'm getting a weird issue with fft3dgpu. Blocks on fairly solid background (with slight gradient), but not all such backgrounds, as the following picture illustrates:

dss("movie")
fft3dgpu() #also tried pretty much any combination of settings

produces:
http://img390.imageshack.us/img390/7860/fryedblocksba3.png

brightness/contrast pumped to show that walls are artifacts free:
http://img368.imageshack.us/img368/7691/brightfryedblockszu3.png

This is on XP SP3 running ATI 2600 XT with catalyst 8.6 and filter 0.8.2 under avisynth 2.5.7

type1encryption

5th July 2008, 09:24

atm I'm using fft3dgpu a lot, but a downside (possible bug) is: when I lock down vista, the filter seams to stop working

acrespo

8th July 2008, 03:36

Yes, I have the same problem with Vista. Lock computer and the enconding process freeze.

Mystery Keeper

8th July 2008, 08:31

Lexor, your sample seem gibbsed. Ever tried MSUCartoonRestore?

type1encryption

8th July 2008, 09:25

In "power options" I can choose to disable energy saving options for pci-e, but that won't work either.

Warpman

8th July 2008, 09:28

Sup everyone, I'm getting a weird issue with fft3dgpu. Blocks on fairly solid background (with slight gradient), but not all such backgrounds, as the following picture illustrates:

dss("movie")
fft3dgpu() #also tried pretty much any combination of settings

produces:

brightness/contrast pumped to show that walls are artifacts free:

This is on XP SP3 running ATI 2600 XT with catalyst 8.6 and filter 0.8.2 under avisynth 2.5.7

i experieced that, too. Sadly i didn't found a solution and switched to dfttest or tnlmeans(slow).
Same happens with ffd3dfilter as well if i remember right.

lexor

8th July 2008, 15:23

Lexor, your sample seem gibbsed. Ever tried MSUCartoonRestore?

Nope, first time I hear about that, I wasn't actually trying to restore the thing. I just noticed the artifacts, since I had fft3dgpu enabled in Avisynth section of FFDshow (so the artifacts are on playback, my source doesn't have that blocking). I posted this because it shows the problem the best.

fjhdavid

16th July 2008, 13:32

Dear all,

I planned to buy a new graphic card (and a new computer core2 duo at 3ghz).

Tsp told us that the filter efficiency speed relies on the number of shading processors.

But ATI give always shading processors figures bigger than nvidia (for the same price). So does I have to choose ATI?

Does shader processor V4.1 and directx 10.1 cards works with fft3dfilterGPU, or does I have to choose an older graphic cards (directx 10 and PS 4.0)?

what is your recommendation for speed graphic cards? (I have 250 old DVD to convert and to filter with dvdrebuilder and I will do a batch but I need speed!)

What fps should I expect today with a 3ghz core2 duo and a modern graphic card?

Thanks for your help!

Francois

Mystery Keeper

16th July 2008, 14:30

Leak

16th July 2008, 14:38

Sad but true, nVidia cards always have better drivers, thus giving less problems.
I can't say I agree - maybe because my last nVidia card was one of those PoS 5600 FX, but I haven't had a problem with the drivers for my Radeon 9800 Pro, my Radeon X1900 XT or my current Radeon 4850.

Also, the Radeon 4850 currently gets you the best bang for the buck.

Also, nVidia has their CUDA technology, which allows to easily write programs for execution on GPU.
And ATI has their own Stream SDK (http://ati.amd.com/technology/streamcomputing/sdkdwnld.html)... so what?

Q: Will the AMD FireStream SDK work on previous generation hardware?

A: To run the CAL/Brook+ SDK, you need a platform based on the AMD R600 GPU or later. R600 and newer GPUs are found with ATI Radeontm HD2400, HD2600, HD2900 and HD3800 graphics board.

np: Boards Of Canada - Pete Standing Alone (Music Has The Right To Children)

fjhdavid

16th July 2008, 15:14

ok, thanks for the info.

just a question.

I read that you can't use fft3dfilterGPU twice in the same script? is it true? (I need to denoise differently luma and chroma)

Blue_MiSfit

19th July 2008, 01:53

@fjhdavid
No, that's not true, I do it all the time.

Multithreading and fft3dgpu tends to be sticky - at least with SetMTMode. I tend to use MT("...") instead when using fft3dgpu in the same script.

@ Lexor:
Are you complaining about the banding on the wall, or the blocking below the lamp shade? I think the former is a result of using aggressive defaults and not dithering afterwards. Try rolling off sigma values like this:

fft3dgpu(plane=4, sigma=2, sigma2=.4, sigma3=.8, sigma4=1.1)
GradFun2DB(1.51)

If it's the latter (blocking), then that's probably in your source. Deblock if you must ;)

~MiSfit

Quark.Fusion

12th August 2008, 17:58

Problem with fft3dGPU: using filter 2 times in script with HD content crashing due to memory overflow (2 gb limit), how much memory filter really need and why using so much?

Blue_MiSfit

12th August 2008, 21:36

I get out of memory errors sometimes when I try to run two instances of fft3dgpu on 1080p content. I'm using a 9600gt at work with 512MB of memory, and RivaTuner shows almost all of it wired up for just one instance - I can't even imagine how much swapping must go on when I chain two instances!!! It's making a lot of sense why Quadros have 1.5 GB of memory :D

~MiSfit

Quark.Fusion

13th August 2008, 02:38

8 GB here :) and my best wish is 64-bit avisynch and filters :) I even tried to chain 2 processes, but problem is that my source can provide only linear seeking and it breaks with multiple processes. And i have no tools to remux into another format. (Only can recode to lossless, but don't have enought free space, lol) Memory is now cheap, but after my new PC i out of money :(

Blue_MiSfit

13th August 2008, 06:51

I meant VRAM :)

I have 4GB of RAM, but my video card runs out of RAM easily :)

~MiSfit

Blue_MiSfit

13th August 2008, 22:44

Adub

13th August 2008, 23:37

Have you updated your drivers recently?

Blue_MiSfit

14th August 2008, 00:43

Yes, actually I just did a complete system reinstall and have the latest installed.

~MiSfit

Quark.Fusion

14th August 2008, 15:10

I meant VRAM :)

I have 4GB of RAM, but my video card runs out of RAM easily :)

~MiSfit

It's no problem in Vista as Memory Manager takes care of video ram too.

Quark.Fusion

14th August 2008, 15:13

On that same note, when I'm encoding 1080p with fft3dgpu(plane=4, sigma=3) (aggressive, I know, it's a shit source), I have issues doing anything with my video card.

Trying to watch a 1080p movie (with software decoding, I'm not that masochistic) through MPC-HC with Haali renderer results in corrupted video and borken playback. I have to switch to overlay mixer to get clean results.

~MiSfit

I have 8800GT 512MB — don't remember any problems with corruption. I was tested fft3dgpu in avisynth script, playing it in WMP11 with Aero interface on and 2 displays :)

Jeremy Duncan

14th August 2008, 20:46

i see there is still no dehalo in the fft3dgpu filter.

i use dehalo in my fft3d script, and i want to know when this filter will be updated to include dehalo and all the parameters in the code i will post below.

FFT3DFILTER(sigma=0.1, bw=10, bh=10, ow=5, oh=5, bt=0, dehalo=1, hr=1, sharpen=0.5)

no rush, take your time, but i am waiting. :)

Blue_MiSfit

14th August 2008, 22:01

I guess that's a good reason to move to Vista on my encoding workstation..

I have an 8800gt and Vista at home - and that works just fine... it's just the 9600gt on XP that chokes.. :(

~MiSfit

lexor

14th August 2008, 22:53

@ Lexor:
Are you complaining about the banding on the wall, or the blocking below the lamp shade? I think the former is a result of using aggressive defaults and not dithering afterwards. Try rolling off sigma values like this:
No I'm talking about the giant squares to the left and right of Fry's head.

If it's the latter (blocking), then that's probably in your source. Deblock if you must ;)

The source is not blocky, and deblocking does nothing to affect this problem. Notice how the walls are smooth, the headboard should be that smooth.

fft3dgpu creates those squares all by itself, it's not enhancing a preexisting condition.