fft3dGPU 0.8.2

[tsp]

Uploaded version 0.5a. It fix a bug with bt=2. Only file changed is fft3dgpu.hlsl.

[Revgen]

Originally Posted by Fizick

What is degrid sharpening?

Originally Posted by tsp

sharpening with degrid.

LOL

I'll eventually try it out and see how it works, once I find some time.

[Kopernikus]

@tsp: Is there somewhere more information about shader programming available? Perhaps a sort of SDK?

[AssassiNBG]

Umm ... am I blind or is there no new version on http://www.avisynth.org/tsp/ ?

[tsp]

Sorry forgot to update the index page. It should be fixed now. Also I uploaded a new version 0.51 that fixes a bug where the parameters after NVPerf was shifted one place so degrid=scutoff,scutoff=svr, etc. Improved download speed from GPU and Kalman should work with geforce fx 5xxx.

I created a special version of fft3dGPU that reports the time it takes to download the final image from the gpu. You can get it here. Just run the included download speed.avs after the included fft3dgpu.dll has been extracted to the plugin directory. On my computer with a AGP Geforce 6800GT it takes ~4.3 mikrosec to download it. That's about 100 MBytes/sec [EDIT]92MBytes/sec (the other is million bytes/sec)[/EDIT]. AGPx8 speed upload speed is about 2100 MBytes/sec. So if anyone with a PCI-express GPU would run the test to compare the result.

Kopernikus: There are the DirectX SDK that contains some sample. Also both NVidia and ATI have SDK available. Most of the sample is game orientatet but there are also some image/video and general purpose GPU (GPGPU) shader examples. The sample chapters from the ShaderX bookseries contains some nice sample for image manipulations on the GPU.
www.gpgpu.org is also a good site although mostly OpenGL.

[Kopernikus]

Thank you

[ariga]

Originally Posted by tsp

ariga: I will post the next version shortly but please post your script else it is very difficult to figure out that is wrong.

It's a simple
AviSource("dv.avi")
LeakKernelDeint(order=0)
FFT3DGPU() # no params

Doesn't matter what params i pass, the error is the same.

BTW, I tried 0.47 and it complained about missing d3dx9_27.dll

Just d/l 0.51. Will see if it works.

[acrespo]

I am trying a compare between FFT3Dfilter and FFT3DGPU.

FFT3Dfilter is more efficient than FFT3DGPU in my anime captures. The parameters I used:

FFT3DFilter(sigma=5, sharpen=1.0) << version 1.8.3
FFT3DGPU(sigma=5, bh=48, bw=48, mode=1, sharpen=1.0) << version 0.5a

I still have some grid lines with fft3dgpu in some frames and fft3dfilter don't have any grid in all video. I guess the only thing is different in the parameters above is the overlap. The default overlap in fft3dfilter is bw,bh/3 and I don't know the default overlap of fft3gpu.

[tsp]

all: I made an installer. It might be more userfrindly

acrespo:the default overlap is ow=bw/2 oh=bh/2 currently that can't be changed. Also there where a bug in 0.5a that assigned the value from svr to degrid. So the default value was 0.3. That might explain it. That is fixed in 0.51. Another difference is the precision used. By default fft3dgpu uses 16 bit floating point precission while fft3dfilter uses 32 bit precision. You can change that by setting useFloat16=false. This slows down the filter and uses more memory on the GPU.

ariga: I don't get any errors with 0.51, fft3dgpu() and a Geforce fx 5200.

[AI]

Originally Posted by tsp

I created a special version of fft3dGPU that reports the time it takes to download the final image from the gpu. You can get it here. Just run the included download speed.avs after the included fft3dgpu.dll has been extracted to the plugin directory. On my computer with a AGP Geforce 6800GT it takes ~4.3 mikrosec to download it. That's about 100 MBytes/sec. AGPx8 speed upload speed is about 2100 MBytes/sec. So if anyone with a PCI-express GPU would run the test to compare the result.

ATI x700 DDR(I) 128bit, A64 3000+ S939

if core ratio = 4 (i.e. 800Mhz - min) = 7,5e-4 (sec?)
if core ratio = 9 (i.e. 1800Mhz) 1,2e-3 (sec?)

I can test in PIII800 fx5200, but later.

[tsp]

Ai : thanks for the report. The time is reported in seconds. Strange that increasing the core ratio slows the download down. The x700 is pci-express?

[AI]

Originally Posted by tsp

The x700 is pci-express?

Yes, of course... (buy special for fft3dGPU and upgrate CPU special for FFT3DFilter)

Originally Posted by tsp

Strange that increasing the core ratio slows the download down

Becouse "CPU more busy memory" (or "memory more busy from CPU" - my English is very bad)

in A64 memory controller integreted in core (more exactly in CPU chip)

PS What about speed? (How many Mb/s on my timings(?))

PPS Why "bt=2 uses the current and next frame",
but "bt=4 uses the two previous frames, the current and next frame"?

[ariga]

Originally Posted by tsp

ariga: I don't get any errors with 0.51, fft3dgpu() and a Geforce fx 5200.

Still no change with 0.51 and 5600 Ultra. I'll try installing the latest DirectX update and drivers. Just 9.0c with the d3dx9_27.dll may not suffice.

[acrespo]

fft3dfilter 1.8.4 include muti-plane option (luma + all chroma planes > plane=4). Can be develop in fft3dgpu too?

[acrespo]

Originally Posted by tsp

acrespo:the default overlap is ow=bw/2 oh=bh/2 currently that can't be changed. Also there where a bug in 0.5a that assigned the value from svr to degrid. So the default value was 0.3. That might explain it. That is fixed in 0.51. Another difference is the precision used. By default fft3dgpu uses 16 bit floating point precission while fft3dfilter uses 32 bit precision. You can change that by setting useFloat16=false. This slows down the filter and uses more memory on the GPU.

The new 0.51 version degrid is function correctly now. Also I think bw and bh defaults is producing grid. When I try 48 (fft3dfilter defaults) I don't have grid, so I think that change bh,bw defaults to 48 is good.

[AI]

Originally Posted by acrespo

The new 0.51 version degrid is function correctly now. Also I think bw and bh defaults is producing grid. When I try 48 (fft3dfilter defaults) I don't have grid, so I think that change bh,bw defaults to 48 is good.

RTFM

Originally Posted by fft3dgpu.txt

bw,bh: blockwide and block height. It should be a power of 2 ie valid values is 4,8,16,32,64,128,256,512. Default=32

IMHO

i.e. then you use bw=bh=48 - realy bw=bh=64 (possible)

[ariga]

Originally Posted by ariga

Still no change with 0.51 and 5600 Ultra. I'll try installing the latest DirectX update and drivers. Just 9.0c with the d3dx9_27.dll may not suffice.

Updated to the latest Nvidia drivers and it works ! Sorry for the trouble. However it's still slow (~4fps). Will update DirectX too and see if it makes any difference.

[tsp]

Originally Posted by AI

PS What about speed? (How many Mb/s on my timings(?))

to find the speed divide 0.39551 (=720*576/1024/1024) with the time i took. So 7.5*10^-4 sec becomes ~530MBytes/sec and 1.2*10^-3 sec is ~330MBytes/sec.

Originally Posted by AI

PPS Why "bt=2 uses the current and next frame",
but "bt=4 uses the two previous frames, the current and next frame"?

sorry bt=2 does use the current and the previous frame. An error in the documentation (maybe it did use the next frame in version 0.1?)

Originally Posted by ariga

Updated to the latest Nvidia drivers and it works ! Sorry for the trouble. However it's still slow (~4fps). Will update DirectX too and see if it makes any difference.

Good to hear. Just curious what version did you use before? Also I think the main reason you are seeing such a low speed is because a Geforce FX 5600 Ultra is very slow. It does only have 2 pixel pipeline(Meaning it can only process two pixel at a time) running at 350 MHz. It doesn't support MRT(Multiple Render Targets) meaning it needs more passes to do the fft (and kalman filtering). 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

Originally Posted by acrespo

fft3dfilter 1.8.4 include muti-plane option (luma + all chroma planes > plane=4). Can be develop in fft3dgpu too?

sure no problem (but it will become plane=3 because plane=1 process both the U and V chroma plane)

Originally Posted by acrespo

The new 0.51 version degrid is function correctly now. Also I think bw and bh defaults is producing grid. When I try 48 (fft3dfilter defaults) I don't have grid, so I think that change bh,bw defaults to 48 is good.

As AI said fft3dGPU uses bw=64 and bh=64 instead. Also the default setting is mode=0 that does produce some nasty griding(but is faster). When I implement the variable overlap I think I will use mode=1 as the default. If you use:
FFT3DGPU(sigma=5, bh=32, bw=32, mode=1, sharpen=1.0)
Does the griding appear?

[acrespo]

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).

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?

For your information, I have forceware 81.89 and the speed is great.

[Fizick]

tsp,
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).