fft3dGPU 0.8.2

[Didée]

Originally posted by Blue_MiSfit
(gives 6of9 nightmares when barely filtered!!)

No, it doesn't give nighmares to 6of9.

It gives you nightmares because of the achieved high bitrates ...

For pressing very noisy sources into tiny files, 6of9 is not suited, by intention.

[Backwoods]

Originally posted by tsp
Backwoods: I'm a little curious why you only get 8-12 fps when I get about 18-21 fps on a 720x576 clip using bt=3. Maybe it's the 4 extra pipelines in the geforce 6800 GT.

Just installed 71.84 and fft3dgpu 0.31 and now:

12~22 fps maintained 16-18

720x480

Xvid Q2

[vinetu]

I did some tests and here is the results.
The CPU is P4 1.8GHz overclocked to 2.9GHz,The VGA card is Radeon 9600 Non Pro.
The source avi is an uncompressed, progressive 720x576 YV12.avi ,273 frames,almost static natural video .
The avs script is :
-------------
LoadPlugin("fft3dGPU.dll")
Avisource("X:\YV12.avi")
fft3dGPU(bla,bla)
-------------
The "X:" drive is a 500 Mb RAMDiSK drive.
Processed in VirtualDubMod in direct stream mode and the filtered avi is saved on the same RAM drive -no HDDs involved...

It's really impossible to me to see the differences between original and filtered images,
so I decide to compress the filtered avi files at later point to XviD (single pass,quant 2,no B-frames) to "visualize" the filtering.

The chain is "uncompressed->fft_filter->uncompressed".

|__ settings __________________ | fft3dfilter (fps) / XviD avi Size _____ | fft3dGPU (fps) / XviD avi Size

|__ no filtering ________________ | _______________________________________________ 9,861,120 bytes

| (sigma=2.0, bt=1) ____________ | _ 10.11 fps / 8,026,112 bytes _____ | 18.20 fps / 9,439,232 bytes

| (sigma=3.0, bt=3) ____________ | __ 7.58 fps / 6,307,840 bytes _____ | 16.05 fps / 7,495,680 bytes ___ |

| (sigma=3.0, bt=3, bh=16 ,bw=16) | __ 5.46 fps / 6,408,192 bytes _____ | 19.50 fps / 8,501,248 bytes ___ |

| (sigma=3.0, bt=3, bh=48 ,bw=48) | __ 7.00 fps / 6,305,792 bytes _____ | 13.00 fps / 6,516,736 bytes ___ |
| (sigma=3.0, bt=3, bh=48 ,bw=48) | ___ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ | 14.37 fps (R9600 overclocked from 325/202 to 425/225 core/memory)

This test is my fisrt try with fft based filtering -I'm VERY impressed by the compressibility results without distorting original.

Bow to the ground to both of you Fizick and Tsp!!!

Edit: text formating

[tsp]

thanks for the feedback. Wonder if someone with a Geforce 6800 Ultra SLI could test this filter

vinetu: Please note that if bw or bh is not a power of two (4,8,16,32,64,128,256,512) it's rounded up to the next power of two 48 -> 64. This is not the case when using fft3dfilter because it uses fftw. Also it should be faster to compress directly to XviD when using fft3dGPU because the waisted CPU cycles when waiting on the graphics card are used the encode the last processed frame (if the program is multithreaded (virtualdubmod is)).

[vinetu]

Thank You tsp!!!

In addition to speed tests - there is no difference in speed between AGPx8 and AGPx4 modes here,
so I guess PCI-E cards should have equal performance to AGP ones (if core chip/memory is the same)

Best Regards

[Soulhunter]

Uhm, the latest version throws a error... :\

[tsp]

Soulhunter: You should copy the file ps.hlsl from the zip file into the same directory as fft3dgpu.dll (in this case c:\programme\Avisynth 2.5\plugins\)

[Soulhunter]

Ouch, I feel very stupid now...

Guess it was one of this "drag&drop" errors !?!


Bye

[Leo 69]

With this filter I get resized picture (i.e reduced one) with bunch of big black artifacts all over the place. I use GeForce FX 5900 NU @ 71.84 official drivers.

[tsp]

Leo 69: Damn I hoped the new version would work on a Geforce Fx. Would you try this version and see if it works with bt=1 or bt=2 or both. In this test version the filtering is disabled so it is just to see where the error is.

[Blue_MiSfit]

@ Didee

No, it doesn't give nighmares to 6of9.

It gives you nightmares because of the achieved high bitrates ...

For pressing very noisy sources into tiny files, 6of9 is not suited, by intention.

I was actually doing a cq2 compressibilty test for 6of9

[Leo 69]

Originally posted by tsp
Leo 69: Damn I hoped the new version would work on a Geforce Fx. Would you try this version and see if it works with bt=1 or bt=2 or both. In this test version the filtering is disabled so it is just to see where the error is.

Yes, the test version works fine

[tsp]

Leo 69: Also with fft3dGPU(bt=1)?? If that is the case it shouldn't be to hard fixing the error.

[Leo 69]

Originally posted by tsp
Leo 69: Also with fft3dGPU(bt=1)?? If that is the case it shouldn't be to hard fixing the error.

Yes, everything's OK with bt=1 too, tsp

[tsp]

Leo 69: Good then try version 0.31 again but change the following passage in the end of ps.hlsl(should be in the same directory as fft3dgpu.dll. It's an ordinary text file so use notepad to open it)

//****************************************************************
#ifdef BETA
float4 WFilter( PS_INPUT In) : COLOR
{
float4 src=tex2D(Src,In.texCoord);
float2 PSD=float2(length(src.xz),length(src.yw));
float4 MulFac=float4(BETA.x,BETA.x,BETA.x,BETA.x);
if(SIGMA.x<PSD.x)
	MulFac.xz=((PSD.x-SIGMA.y)/PSD.x);
if(SIGMA.x<PSD.y)
	MulFac.yw=((PSD.y-SIGMA.y)/PSD.y);
return MulFac*src;
}
#endif
//*******************************************************************

to this:

//*******************************************************************
#ifdef BETA
float4 WFilter( PS_INPUT In) : COLOR
{
float4 src=tex2D(Src,In.texCoord);
float2 PSD=float2(length(src.xz),length(src.yw));
float4 MulFac;
float4 dst;
MulFac.xz=((PSD.x-SIGMA.y)/PSD.x)*(SIGMA.x<PSD.x)+(SIGMA.x>=PSD.x)*float2(BETA.x,BETA.x);
MulFac.yw=((PSD.y-SIGMA.y)/PSD.y)*(SIGMA.x<PSD.y)+(SIGMA.x>=PSD.y)*float2(BETA.x,BETA.x);
dst=MulFac*src;
return dst;
}
#endif
//*******************************************************************

if that doesn't work try this version

//*******************************************************************
#ifdef BETA
float4 WFilter( PS_INPUT In) : COLOR
{
float4 src=tex2D(Src,In.texCoord);
float2 PSD=float2(length(src.xz),length(src.yw));
float4 dst;
if(SIGMA.x<PSD.x)
	dst.xz=src.xz*((PSD.x-SIGMA.y)/PSD.x);
else
	dst.xz=src.xz*float2(BETA.x,BETA.x);
if(SIGMA.x<PSD.y)
	dst.yw=src.yw*((PSD.y-SIGMA.y)/PSD.y);
else
	dst.xz=src.xz*float2(BETA.x,BETA.x);
return dst;
}
#endif
//*******************************************************************

[Leo 69]

None of the script versions work (properly), tsp. Overall Bt=2 mode gives largest amount of artifacts and by the way my mouse constantly stops responding for very short periods of time during playback (~0.2 sec or so).With test version too.

[tsp]

hmm could you post a sceneshoot of the artifacts? Also does this version produce artifacts(This disables the filtering but it's mainly to confirm that the bug lies in the if statements).

//*******************************************************************
#ifdef BETA
float4 WFilter( PS_INPUT In) : COLOR
{
float4 src=tex2D(Src,In.texCoord);
float4 dst;
dst=src;
return dst;
}
#endif
//*******************************************************************

Also I haven't heared about the mouse problem before. But I will see what I can do about it later when I get this filter working in a geforce FX (even if I have to buy a geforce FX 5200 to test on. Anyone has a spare one?)

[LordIntruder]

Hi,


I tested you work TSP and here is what I get:

I encoded a 10000 video frames (720 x 528) with this parameter for FFT:
'FFT3DFilter(sigma=3, bt=3, measure=true)'

For FFT3dGPU I used:
'fft3dGPU(sigma=3, bt=3)'

On a 2400+, 1Gb Ram, AGP Radeon 9600 Pro with latest drivers, Windows XP SP2, DirectX 9.0c. Neither my CPU nor my GPU are overcloked.

Without FFT:

1st Pass = 13 min
2nd Pass = 37 min

FFT Measure True:

1st Pass = 53 min
2nd Pass = 80 min

FFT Measure OFF:

1st Pass = 59 min
2nd Pass = 82 min

FFT3dGPU

1st Pass = 17 min
2nd Pass = 40 min

In my example I selected the heaviest solution with latest XviD beta: Qpel, GMC, VHQ4, VHQ for bframes, Chroma motion, etc... So this is why the encode is so slow. You untick Qpel, you put VHQ1, no chroma motion and the speed rise up to the roof

The GPU version is amazingly fast !!!! I can't believe it!!!

As far as my eyes can see, the quality seems the same between your filter and the original by Fizick. Can you confirm the only thing is about the 16 bits float (useFloat16)? Except that option (which we still can enable to 32 bits), we are suppose to get the same quality that the original filter right?

A last thing I don't understand, I quote you

--
"usecache: if enabled the frames are saved in the GPU after the 2d FFT to avoid calculating them again the next frame if bt=2 or 3.
It can be necessary to disable this internal cache if using motion compensation. Default = true"
--

What do you mean by motion compensation? GMC option in DivX or XviD? What drawbacks are we suppose to get? Artifacts I suppose? And this option Off slow down the encode a lot?

A great thanks for your work, Fizick's filter is very good but so slow. A good idea you've got here. I wouldn't imagine that my video card helps me to encode faster. Another reason to buy a fast new card

[tsp]

Originally posted by LordIntruder

As far as my eyes can see, the quality seems the same between your filter and the original by Fizick. Can you confirm the only thing is about the 16 bits float (useFloat16)? Except that option (which we still can enable to 32 bits), we are suppose to get the same quality that the original filter right?

Ahemm I just discovered that my filter cheats a little more than just using 16 bit float. I'm only using a 1:1 overlap instead of a 2:1 this means that the filter only does half as many calculations as fft3dfilter (explaining in part why it's 4 times as fast). This results in border artifacts when using high sigmavalues (about 2.5-10 depending on bw/h that is higher bw lower sigma before artifacts appears ). That is a 1-2 pixel width dark border. Like this image:


I will do two things about that:
1) Implement the 2:1 overlap(This will cut the speed in half ).
2) and as an option for the speed hungry people. Just use a slightly larger blocksize and then crop the borders.

Until that is implementet be a little extra carefull with high sigma values.

A last thing I don't understand, I quote you

--
"usecache: if enabled the frames are saved in the GPU after the 2d FFT to avoid calculating them again the next frame if bt=2 or 3.
It can be necessary to disable this internal cache if using motion compensation. Default = true"
--

What do you mean by motion compensation? GMC option in DivX or XviD? What drawbacks are we suppose to get? Artifacts I suppose? And this option Off slow down the encode a lot?

No it was mainly aimed at MVTools but I don't think it will cause artifact anyway so I will disable this option in the next version.

[tsp]

released version 0.40. Now includes sharpening, 2:1 overlap and 1:1 overlap with border. Also better optimized for multitasking (the filters before fft3dgpu are processed at the same time as fft3dgpu)