fft3dGPU 0.8.2

[modsoul]

see first message in this thread

PS manual (http://avisynth.org.ru/docs/english/externalfilters/fft3dgpu.htm)

i've read that post and dl'ed every file it list. the newer files just don't have ps.hlsl (but i use the ps.hlsl from the older versions it does work). and the only referewnce to ps.hlsl in the first post is a changelog. same with the manual you linked.

[tsp]

modsoul: what version did you install?

Not related to your problem. This is what you got if you overclock the GPU to much with fft3dgpu(upper left corner)
http://img101.imageshack.us/img101/8528/ocgpu1il.th.jpg (http://img101.imageshack.us/my.php?image=ocgpu1il.jpg)

[modsoul]

modsoul: what version did you install?

Not related to your problem. This is what you got if you overclock the GPU to much with fft3dgpu(upper left corner)
http://img101.imageshack.us/img101/8528/ocgpu1il.th.jpg (http://img101.imageshack.us/my.php?image=ocgpu1il.jpg)

the latest version. 0.6.2.9 beta. then since that did not work 0.6.2.
then since that did not work i took the ps.hlsl file from one on the much older (0.4) versions.

[tsp]

ok I tried to recreate the problem but couldn't. I even search the memory used by ftt3dgpu.dll for the string ps.hlsl and it doesn't show up. Maybe you got an old installation of fft3dgpu laying around somewhere (try searching for fft3dgpu.dll and see if there is more than one on your computer). Try this srcipt to see what version of fft3dgpu is loaded:

blackness().subtitle(LoadPlugin("fft3dgpu.dll"))

it should report version 0.6.2.9 if you use the latest beta.

Ps: Is there really a whitespace in front of avisynth 2.5 in the path "C:\program file\ avi synth 2.5\"?

[AI]

May be version DirectX isn't right?

[Boulder]

Doesn't the installer automatically install the needed DirectX dll file as well?

[tsp]

yes but it requeres directx 9.0c to be installed. Anyway the errormessage looks different if it was a missing directx file. The missing ps.hlsl file message should only popup on the older versions of my filter so I suspect that an old version is loaded instead.

[tsp]

Released version 0.6.3. It includes the new fft code that should give a performanceboost for larger blocksize(>16) and when using precision>0 also includes a html doc

[Adub]

Great! Thanks for the update tsp. Keep on coding man, we all need you!

[Alain2]

This new version doesn't seem to work on my computer (with a X1600pro gpu):
unfiltered:
http://img140.imageshack.us/img140/3041/temp9bq.jpg

fft3dgpu(sigma=3,bt=4,plane=4) (v0.6.3)
http://img140.imageshack.us/img140/8801/temp1ui.jpg

fft3dgpu(sigma=3,bt=4,plane=4) (v0.6.2.1)
http://img516.imageshack.us/img516/294/temp6hf.jpg

[Boulder]

Doesn't work on my GeForce 6600GT either. v0.6.2.9 does work but v0.6.3 doesn't.

I tried FFT3DGPU(sigma=3,plane=1,mode=1,degrid=1,bt=3,bw=16,bh=16) and also without bw and bh.

[tsp]

Boulder: What driver version are you using? I tested it using 84.21 on a geforce 7800GT.

Alain2: Does version 0.6.2.9 (http://www.avisynth.org/tsp/fft3dgpu0.6.2.9.exe) produces the same result?

All: If you got this version working could you post what card and driver version you are using?

[Warpman]

Released version 0.6.3. It includes the new fft code that should give a performanceboost for larger blocksize(>16) and when using precision>0 also includes a html doc
do you plan to make it working on geforce 5XXX again?
would be nice though ^^'

[tsp]

Warpman: It should work on geforce fx 5xxx (at least I added the code. I first have my comp with geforce fx 5200 on next monday so I can first test it there)

[Warpman]

okay it works on my 5900 (just tested it). Great improvment in Speed here. ^^

But i have Croma shifting :/ i have a smple clip here-->
http://rapidshare.de/files/21713411/conan3dgpu2.mkv.html

[tsp]

Warpman: what setting did you use?

[Warpman]

Warpman: what setting did you use?
fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8)

[Boulder]

Boulder: What driver version are you using? I tested it using 84.21 on a geforce 7800GT.

I've got 78.01 installed, haven't had the need to upgrade yet. I could try upgrading to the latest official version tomorrow and see if that fixes things. I actually thought that the newer versions were slower what comes to FFT3DGPU than the old ones, there's a post about that somewhere in this thread.

[tsp]

Boulder: Revgen tested it with a dualcore athlon X2 processors and found it to be slower. Don't know if it still applies. I don't know what should cause it if it is not the driver as fare is I know only transparency AA was added to the Geforce 7xxx series and it isn't used by fft3dgpu.

Warpman: I can't reproduce it here. Is the chroma lagging 1 frame or is the luma 1 frame forward? If the chroma is lagging try this script:

MergeLuma(fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8))

Just to be sure is the filter working?

[Warpman]

Boulder: Revgen tested it with a dualcore athlon X2 processors and found it to be slower. Don't know if it still applies. I don't know what should cause it if it is not the driver as fare is I know only transparency AA was added to the Geforce 7xxx series and it isn't used by fft3dgpu.

Warpman: I can't reproduce it here. Is the chroma lagging 1 frame or is the luma 1 frame forward? If the chroma is lagging try this script:

MergeLuma(fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8))

Just to be sure is the filter working?
Your right Luma is 1 frame ahead ... sorry my bad
MergeLuma sadly did not help

about the speed i did upgrade from a old version of ff3gpu to 0.6.3 so this explains the speed gain...

http://img304.imageshack.us/img304/5278/shift9ep.png (http://imageshack.us)

[tsp]

MergeLuma(trim(1,0).fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8))

When I get my geforce fx back I will make a proper fix.

[Warpman]

thx ^^

but with this workaround luma is now 2 frames ahead (sorry to bother you with this)

Edit: MergeChroma(fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8)) fixed it....

Edit2: im a bit confused now ... it´s defenitly the luma which is 1 frame "to ealy" ... but MergeChroma Fixed it...

[tsp]

that is because you take the unfiltered Chroma from fft3dgpu and merge it into the untoched luma so basicly you skip the filtering so try this:

t=last
t.trim(1,0).MergeLuma(t.fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8))

easy to mess up the trims. Sorry for that.

[Chainmax]

Would FFT3DGPU have any issues on a 64-bit, 128MB, AGP (i.é: no turbocache) GeForce 6200?

[Alain2]

Alain2: Does version 0.6.2.9 (http://www.avisynth.org/tsp/fft3dgpu0.6.2.9.exe) produces the same result?

All: If you got this version working could you post what card and driver version you are using?

Yes, same problem. Versions up to 0.6.2.1 are ok, 0.6.2.9 and 0.6.3 are not on my config (ATI X1600pro pci-e)
http://img415.imageshack.us/img415/6295/temp0oz.jpg
(I think it's the catalyst drivers from january 2006, 6-1_xp-2k_dd_29602.exe)

[Warpman]

that is because you take the unfiltered Chroma from fft3dgpu and merge it into the untoched luma so basicly you skip the filtering so try this:

t=last
t.trim(1,0).MergeLuma(t.fft3dGPU(sigma=3, plane=0, bt=3, bw=16, bh=16, sharpen=0.8))

easy to mess up the trims. Sorry for that.
no problem ;) works... thx again ^^


Would FFT3DGPU have any issues on a 64-bit, 128MB, AGP (i.é: no turbocache) GeForce 6200?

if it works with my 5900 (agp, 128Mb, 128bit) a 6200 should work too.

[AI]

tsp
May be easy add aviable GPUFFTW (http://gamma.cs.unc.edu/GPUFFTW/download.html) in fft3dfilter? ;)

[tsp]

AI: not that easy because:
1) It's OpenGL and I use directx
2) It doesn't work on ATI cards
3) It's only a 1d transform (and I need a 2d transform)
4) It doesn't support multiple strides(meaning it can only transform one block)
5) It returns the output in main memory and I need a texture to work with
6) I'm not sure it's released under the GPL license

But I might be able to use some of the code in the fft but the new fft algorithm I also use a Stockhamn autosort fft like gpufftw

most of the above points also applies to fft3dfilter

[AI]

tsp,Thanks for detailed answer my question.

I hope my link help you :)

[tsp]

AI: sure. BTW does your x700 work with the new version?

For those of you that have problem with the new version could you try this (http://www.avisynth.org/tsp/fft3dgpu0.6.3D.exe) version that reports back error(only thing new) if any in the fft code. Also try precision=2 and if it helps.

[Alain2]

no, not better for me. Can't see any error report either

It's strange, reloading the same script several times will bring a different first image everytime, but it's getting messier, starting with something like the picture I posted, and now it's more like:
http://img171.imageshack.us/img171/6474/temp8er.jpg

Hope it's not clogging my cards memory (256Mo, forgot to mention) ^^

I reduced my script to a simple one:
mpeg2source("01.d2v",idct=7)

tfm(d2v="01.d2v")
tdecimate(mode=1)

crop(4,2,-4,-2)
BicubicResize(704,528,0,.5)

fft3dgpu(precision=2)

No apparent change with/without precision=2

[Lemonzest]

i'm getting the above also with 0.6.3

[Boulder]

Upgrading to the latest beta ForceWare drivers (v91.28) fixed the problem. I guess it was time to upgrade :p

[AI]

v0.6.3 (w/o "d") not work.

tsp, I can't download 1Mb every time, becose I use GPRS :(
do you can give me link do archive dll and hlsl only?

[tsp]

So until now 3 ATI users that can't get version 0.6.3 to work. That's enough for me. Just ordered a X1300 pci card to test the code (and add support for multiple cards for the SLI freaks out there). And please tell me you used the latest drivers right (I think I will add that to the FAQ part of the manual)

AI: You can be glad that I didn't add the full 52 MB directx installation to the installer. But I can make a light version that only includes the dll, html doc and hlsl for those people that doesn't have flatrate or slow connections(Do you use a cellphone for your internet connection?)

[Lemonzest]

Yes, ATi User here, using the latest Catlyst Version 6.5 garbled display.

[G_M_C]

Boy, am i glad i bought myself a 7900GTX in AGP deguise (aka 7800GS GS+) :p

But to make shure all is fine; I'll test the new version the next few days, and will report back afterwards.

[hartford]

Having the same problem as Alain2.

Version 0.6.2 works, 0.6.2.9 does not.

Specs:

Graphics: Ati 9700

DirectX: 9.0c with December Update

Motherboard: Gigabyte GA-7DPXDWP (dual processor)

OS: Windows 2000 sp4 with updates.

[blubberbirne]

Hi, same problem here.

Still using Version 0.6.1.9 and with this one it works perfekt :)
ATI is CAT6.5 with X850XT (Maybe Powerfull enough)

Use the Filter to sharpen my DVB Caps. The Result looks Amazing :)

Nice work for the Filter

[tsp]

I just got my x1300 card. Unforfunatly it didn't fit in my second PCI-express slot because of my watercooling kit. So I replaced my Geforce with the ati card and will try to trackdown what's going wrong.

[Wilbert]

@tsp,

Just curious. Did you turn it into a closed source filter?

[tsp]

Wilbert: No the source is included in the installer. You can select only to extract the source or install both the source and filter or only the filter(the default). The sourcecode are extracted to a directory called fft3dgpu_src in the install directory.

All: I found the bug that was causing problem with the ati card (and reference render but not on nvidia cards) so a new version should be released shortly.

[tsp]

version 0.6.4 ready. Should now work on ati card. (also contains limited support for interlaced source. use interlace=true to enable it. currently on works correctly with ow=bw/2 and oh=bh/2(the default value).

[blubberbirne]

*hm* Ok its work, but performance is bad :(

Going back to 6.2 an waiting for new release ;)

[InsulinJunkie]

0.6.4 does now work on an ATI 9700, but it is a bit slower than 0.6.2 using a fairly simple:

FFT3dGPU(sigma=2,bt=3,mode=1,sharpen=0.62)

On my XP2400, running FFT3DGPU alone - no other filters - and going from MJPEG to MJPEG, it runs at about ~12-13FPS, vs ~16FPS using 0.6.2.

Not a big deal, but just wanted to pass that along...

[tsp]

hmm yes performance seems to be worse on lower end hardware using the default value for precision(0 = 16 bit float) with the new code. I will enable the old fft in the next release (that will include support for interlaced frame processing and sigma2,sigma3 and sigma4).

[blubberbirne]

hmm yes performance seems to be worse on lower end hardware using the default value for precision(0 = 16 bit float) with the new code. I will enable the old fft in the next release (that will include support for interlaced frame processing and sigma2,sigma3 and sigma4).

lower end hardware? i got a X850XT and a X2 3800+, don't think this is lower end Hardware :devil:

[Alain2]

tsp, is it faster on your X1300 card than the old fft code of 0.6.2?

[tsp]

average time(2000 frames average) to calculate a frame in fft3dgpu with a geforce 7800GT

FFT3dGPU(sigma=1,bt=-1,sharpen=1.0,degrid=1.0,precision=0)
New fft code 0.0271 sec per frame(spf)
Old fft code 0.0258 spf
new code 5% slower

FFT3dGPU(sigma=1,bt=-1,sharpen=1.0,degrid=1.0,precision=1)
new fft:0.0390 spf
old fft:0.0463 spf
new code 16% faster

FFT3dGPU(sigma=1,bt=-1,sharpen=1.0,degrid=1.0,precision=2)
old fft: 0.0472
new fft: 0.0404
new code 14% faster

FFT3dGPU(sigma=1,bt=-1,bw=256,bh=256,sharpen=1.0,degrid=1.0,precision=2)
new fft: 0.0904
old fft: 0.114
new code 21% faster

FFT3dGPU(sigma=1,bt=-1,bw=256,bh=256,sharpen=1.0,degrid=1.0,precision=0)
old fft: 0.0556
new fft: 0.0521
new code 6% faster

Higher precision or blocksize favours the new fft code. The old code is faster at the default blocksize and precision (but use more GPU memory). This is the same result I got on my X1300. I will see if I can make a hybrid version between the new and old code. Anyway I will add an option to use the old code instead of the new code and if not specified fft3dgpu will time what code is fastest and use that.

[Chainmax]

Is FFT3DGPU v0.6.4 missing anything from FFT3DFilter v1.8.5 regarding features or stability?