fft3dGPU 0.8.2

[tsp]

I need some more tests but I think mode=1 and useFloat16=false can avoid grids with bw=bh=64. I will take some screenshots to compare side by side with bw=bh=32, but with useFloat16=true I see grids with all bw/bh and modes (less than v0.47 and 0.5a).

So it seems like it is more the Float16 that are causing it. Could you post a screenshot with the grid artifacts? [EDIT]I made a special version for you that might produce better result with useFloat16=true. Get it here[/EDIT]

I have a doubt about other planes. In readme.txt you wrote about fft3dgpu in plane=0 and other instance in plane=1 can decrease speed more than fft3dfilter. Is it correct to all situation/video cards?

I think it mainly was a problem with some of the older versions. It shouldn't cause any problem with the current version. You can also mix fft3dfilter and fft3dgpu and let one of them process the luma plane and the other the chroma plane (just remember to place fft3dgpu after fft3dfilter). BTW what GPU do you use?

Originally Posted by Fizick

IMHO, for FFt3dFilter plane=4 you may consider using plane=4 too,
and for plane=3 consider plane=3 (as alias of your 1 and 2).

Might be a good idea.

[acrespo]

Well here is some screens using bh=bw=32 and bh=bw=64 and the original without fft3dgpu.
fft3dgpu_screens.rar

edit:Your special version is slight better than v0.51.

[tsp]

acrespo: I'm not sure that the grid you are seeing is the same grid artifact that I'm talking about. For me it looks like the gridsize is about 4x4 pixels when using bh=bh=32 and if you look closely at the original the 4x4 grid is ever sligtly present. If you want to see how the grid artifacts looks like with fft3dgpu try this (the grid that then appears has a gridsize of bw/2, bh/2):
fft3dgpu(mode=1,sigma=1000,degrid=0.0)
But you are right about the grid disappears when using bw=bh=64. It looks like the the image is more blurry with that setting. What codec is the capture file saved in? It could be that the grid is caused by enhancing the artifacts from the compression.

[acrespo]

I always capture with PIC MJPEG 3 quality 19 (or decrease some chrominance and luminance to between quality 18 and 19).

I understood the grid you explain now.

Well I don't understand why bh=bw=64 resolved my problem, because this setting have closely to fft3dfilter with bh=bw=48 (defaults). Can be something wrong with 32 setting?

[ariga]

Originally Posted by tsp

Good to hear. Just curious what version did you use before?

Don't know, since I don't have the original CDs right now. The Nvidia tab in graphics panel, didn't display the version earlier. The UI seems to have changed a little now and the version is also being displayed.

The good thing about having a slow videocard is the cpu-utilization should also be very low so it is possible to run some very slow filters before fft3dgpu(like a deinterlacer) or use a slower codec

You make even the bad sound so good! (Marketing depts anywhere would love to have you )
Slower codec ? Hmm... DivX at "Insane quality" ?
Until an upgrade, it's fft3dfilter for me...

[Mug Funky]

slower codec like snow or dirac...

x264 is always good (it's pretty fast too).

definitely not divx at insane quality... it's not worth it (the insanity lies in wondering why something could take so long while not looking as nice as an xvid encode...)

[puffpio]

x264 w/ Sharktooth's HQ Insane profile...

[AI]

tsp, I have guestions for you... (best phrase on my bad English )
1) I can't found reduceCPU on last docs
2) I found about you comment "reduceCPU"

3) CPU busy 100% -
--------------
FFT3DFilter(bt=4,plane=0)
FT3DGPU(bt=4,mode=1,plane=1)
--------------
my CPU busy 80% -
--------------
FT3DGPU(bt=4,mode=1,plane=1)
FFT3DFilter(bt=4,plane=0)
--------------
I undestend about it... (You sad about it already)
but earlier 20% busy kernel (in taskmanager)

4) --------------------
i=DePanEstimate(Last,dxmax=24,dymax=0,fftw=True)
Pre2=DePan(Last,i,offset=+2,subpixel=1,matchfields=false,mirror=3).Crop(80,0,-80,0)
Prev=DePan(Last,i,offset=+1,subpixel=1,matchfields=false,mirror=3).Crop(80,0,-80,0)
Next=DePan(Last,ir,offset=-1,subpixel=1,matchfields=false,mirror=3).Crop(80,0,-80,0)

Clip

Interleave(Pre2,Prev,Last,Next)
FFT3DFilter(sigma=5,sigma2=4,sigma3=3,sigma4=2,bt=4,plane=0)
SelectEvery(4,2)

Interleave(Pre2,Prev,Last,Next) # a) Busy 80%
#Interleave(Last,Pre2,Prev,Last,Next) # b) Busy 80%
#Interleave(Last,Pre2,Prev,Last) # c) Busy 100% - Good!
#Interleave(Pre2,Prev,Last,MergeLuma(Next,Trim(1,0))) # d) Busy 100% - Good! - My Idea about a+c (or b+c)

FFT3DGPU(sigma=7,bt=4,mode=1,plane=1)
SelectEvery(4,2) # a)
#SelectEvery(5,3) # b)
#SelectEvery(4,3) # c)
#SelectEvery(4,3) # b)
-------------------

[tsp]

AI:
1 and 2: I removed reduceCPU in version 0.5 so that is why you can't find it.
3: That is because fft3dGPU is about 6-7 times faster than fft3dfilter in that script so the reason you see a 80% cpu utilization is that fft3dfilter uses most of the time. In this case it might be worth it to set UseFloat16=false to increase the quality. This shouldn't affect the process time very much but try and see:

FTT3DGPU(bt=4,mode=1,plane=1,usefloat16=false)
FFT3DFilter(bt=4,plane=0)

or compare it to this(try useFloat16=false if you have 256 or 512 MB ram on the graphics card):

FTT3DGPU(bt=4,mode=1,plane=1)
FFT3DGPU(bt=4,mode=1,plane=0)

4:
a) look right to me. Might be a little faster like this:

i=DePanEstimate(Last,dxmax=24,dymax=0,fftw=True)
Pre2=DePan(Last,i,offset=+2,subpixel=1,matchfields=false,mirror=3).Crop(80,0,-80,0)
Prev=DePan(Last,i,offset=+1,subpixel=1,matchfields=false,mirror=3).Crop(80,0,-80,0)
Next=DePan(Last,ir,offset=-1,subpixel=1,matchfields=false,mirror=3).Crop(80,0,-80,0)

Clip

Il=Interleave(Pre2,Prev,Last,Next)
MergeChroma(il.FFT3dFilter(sigma=5,sigma2=4,sigma3=3,sigma4=2,bt=4,plane=0),il.
FFT3DGPU(sigma=7,bt=4,mode=1,plane=1))
SelectEvery(4,2)

acrespo: I think some of the grid might come from interlacing.

[AI]

next question:

FFT3DGPU().FFT3DFilter() - GPU first, CPU secont (in script, realy on the contrary)

Why v0.47 Busy Kernel = 20% (i.e. wait when work GPU, and CPU busy = 100%)
now (v0.51) CPU idle = 20% (Kernel Busy =1-2%, i.e. I/O only)

[tsp]

AI: I change the inner working of this filter somehow in version 0.5. The frame is now uploaded to the GPU while the GPU is working on the frame and the number of passes needed to process the frame is also reduced(that might save a switch to the kernel to process the commands to the GPU). But I think it is a good thing that the kernel and cpu utilization is lower now.

[AI]

ask again: (possible error on your optimize algoritm)

usual script:

i=DePanEstimate(Last,fftw=True)
Pre2=DePan(Last,i,offset=+2,subpixel=1,matchfields=false,mirror=3)
Prev=DePan(Last,i,offset=+1,subpixel=1,matchfields=false,mirror=3)
Next=DePan(Last,i,offset=-1,subpixel=1,matchfields=false,mirror=3)

Interleave(Pre2,Prev,Last,Next)
MergeChroma(FFT3DFilter(bt=4,plane=0),\
FFT3DGPU(bt=4,mode=1,plane=1))
SelectEvery(4,2)

(then FFT3DGPU ask from AVISynth next frame, work only DePan)

optimize script:

i=DePanEstimate(Last,fftw=True)
Pre2=DePan(Last,i,offset=+2,subpixel=1,matchfields=false,mirror=3)
Prev=DePan(Last,i,offset=+1,subpixel=1,matchfields=false,mirror=3)
Next=DePan(Last,i,offset=-1,subpixel=1,matchfields=false,mirror=3)

Interleave(Pre2,Prev,Last,Next)
FFT3DFilter(bt=4,plane=0)
SelectEvery(4,2)

Interleave(Last,Pre2,Prev,Last,Next)
FFT3DGPU(bt=4,mode=1,plane=1)
SelectEvery(5,3)

When GPU work in VRAM n-2,n-1,n and n+1 frames...
FFT3DGPU ask from AVISynth frame n+2, (on this script n+1 after Interleave().FFT3DFilter().SelectEvery())
But this script not use 100% CPU

not right,but realy optimize script:

i=DePanEstimate(Last,fftw=True)
Pre2=DePan(Last,i,offset=+2,subpixel=1,matchfields=false,mirror=3)
Prev=DePan(Last,i,offset=+1,subpixel=1,matchfields=false,mirror=3)
Next=DePan(Last,i,offset=-1,subpixel=1,matchfields=false,mirror=3)

Interleave(Pre2,Prev,Last,Next)
FFT3DFilter(bt=4,plane=0)
SelectEvery(4,2)

Interleave(Last,Pre2,Prev,Last)
FFT3DGPU(bt=4,mode=1,plane=1)
SelectEvery(4,3)

now CPU busy 100%...

And I fount decision:
First script string
"Interleave(Pre2,Prev,Last,Next)" change to
"Interleave(Pre2,Prev,Last,MergeLuma(Next,Trim(1,0)))"
now CPU use 100% and all frames (n+1 too) is compensation.

PS very difficult write on not native language

[tsp]

What about the framerate? Also how does my script above perform?

[tsp]

from my latest development. Guess what it is (texture is missing).

[puffpio]

Originally Posted by tsp

from my latest development. Guess what it is (texture is missing).

hmm the new plane in fft3dfilter where it process YUV all at once?

[tsp]

Nope. Compare it to the overlap picture in the fft3dfilter documentation. Yellow= 4 block overlap. red=2 block overlap and turquoise=no overlap.
the YUV option will be implemented as it is in fft3dfilter so the filter converts:
fft3dgpu(plane=4)
to
fft3dgpu(plane=0)
last.fft3dgpu(plane=1)

[puffpio]

Originally Posted by tsp

the YUV option will be implemented as it is in fft3dfilter so the filter converts:
fft3dgpu(plane=4)
to
fft3dgpu(plane=0)
last.fft3dgpu(plane=1)

Is that really how fft3dfilter does it? because in the fft3dfilter doc it has an example for how to do it w/o using plane=4 like this:

YToUV(fft3dfilter(sigma=3, plane=1).UToY,\
fft3dfilter(sigma=3, plane=2).VToY,\
fft3dfilter(sigma=2, plane=0))

Is that equiv to what you are saying is the way to process all planes?

[tsp]

sorry didn't look closely at the fft3dfilter sourcecode. Fizick implemented it like you said but I don't understand why because it produces the same result as my script just with more copying from one plane to another. Try this script. It should be grey if the result is the same:

src=avisource("f:\20051210-143802.avi")
motion=src.DepanEstimate(trust=2.5, fftw=true)

src.DepanInterleave(data=motion)
YToUV(fft3dfilter(sigma=3, plane=1).UToY,\
fft3dfilter(sigma=3, plane=2).VToY,\
fft3dfilter(sigma=2, plane=0))
a=SelectEvery(3,1)


src.DepanInterleave(data=motion)
fft3dfilter(sigma=3, plane=1)
fft3dfilter(sigma=3, plane=2)
fft3dfilter(sigma=2, plane=0)
b=SelectEvery(3,1)

subtract(a,b).Levels(127, 1, 129, 0, 255)

[AI]

tsp
What about speed?
first script faster!

Give "filter" with matrix 121 (as Convolution3D but only temporal)
filter.filter = "filter" with matrix 14641
filter.filter.filter = "filter" with matrix [1 6 15 20 15 6 1]

if pre filter = matrix [1] then
1
1 2 1
1 4 6 4 1
1 6 15 20 15 6 1
i.e. triangle Pascal (SelectEvenString )

i.e. if use several temopral filters, it more temporal

i.e. need

DePanInterleave(...)
filtre1()
selectevery(3,1)
DePanInterleave(...)
filtre2()
selectevery(3,1)
DePanInterleave(...)
filtre3()
selectevery(3,1)

or

DePanInterleave(prev=3, next=3)
filtre1()
filtre2()
filtre3()
selectevery(7,3)

or
if filters "independent"

DePanInterleave(prev=1, next=1)
MergeChroma(filtre1(),filtre2())
selectevery(3,1)

PS my English is bad up to this date

[tsp]

AI: that is what the build in cache is for. So in my script the filterchain really look like this:

src.DepanInterleave(data=motion)
InternalCache()
fft3dfilter(sigma=3, plane=1)
InternalCache()
fft3dfilter(sigma=3, plane=2)
InternalCache()
fft3dfilter(sigma=2, plane=0)
InternalCache()
b=SelectEvery(3,1)

Most of the pascal triangle is in the cache so the same amount of work is done by fft3dfilter compaired to the YToUV approach just with less copying from one colorspace to another=second script is faster.
[EDIT]
sorry you're right. The InternalCache reduces the number of extra calls to fft3dfilter(plane=2) and fft3dfilter(plane=1) to 4. But the YtoUV is still inefficient with the copying from one plane to another. I will implement it better in fft3dgpu so they write to a common frame.