fft3dGPU 0.8.2

[G_M_C]

Originally Posted by tsp

Is that speed with fft3dgpu(sigma=2,plane=4,precision=2,bw=64,bh=64,ow=24,oh=24)?
The Geforce 7900 is a rather fast card.
I mean Graphics memory. The fft version I work on now should be less memory bandwidth dependent as I have cut down the number of texture lookup

Yes, I've modified my script exactly as you specified in your posting, with fft3dgpu(sigma=2,plane=4,precision=2,bw=64,bh=64,ow=24,oh=24). I was amazed too at the very big speed gain.

And speaking of the GPU; I've bought my card as a 7800GS GoldenSamplePlus, BLISS edition. I allways thought the 7800 series had a G70 class GPU.

But when i removed the cooler today (to change the heatpaste&heatpads to Arctic Silver 5) i saw that my card actually had a G71 on it. I mention this because its indeed a 7900GTX in AGP format. The difference between my version and the "real" 7900GTX is slower mem (512 Mb GDDR3 @ 1250 MHz). So im sort a lucky I guess buying a 7800 and getting a 7900 (BLISS indeed )

I never would have guessed it would be this fast, even with my AviSynth scripts this new Graphics board speeds things up (guess what it does to my 3DMark scores :P).

But as i said; Hope you keep developing filters that use GPU, and I hope others follow suit. Because results are absolutely amazing. Many complements on your work, and i'll keep using your filter (and hopefully more GPU filters in the future)

[swaaye]

So would SM2b add anything useful for FFT3DGPU & ATI's Xx00 series? Or do you think that would be a waste of your time (TSP) to implement?

I'd also like to see some of the favorite settings of the forumgoers here. I've settled on:

FFT3dGPU(sigma=2,bt=2,sharpen=0.5)

for speed and quality. My overclocked X800GTO2 at 510/610 can pull about 55fps with this. I run this with DIVX set at quality=4 on the slider and use DIVX's internal resizer that's multithreaded. Gives me about 90% CPU so I don't slow down because of either GPU or CPU.

[tsp]

So would SM2b add anything useful for FFT3DGPU & ATI's Xx00 series? Or do you think that would be a waste of your time (TSP) to implement?

well it is allready implemented. The code in fft3dgpu.hlsl is compiled and optimized by directx using the best shader model your card support (that is0 SM 2.0 for Radeon 9xxx,SM 2a for geforce fx 5xxx, SM 2b for Radeon Xxxx, SM 3 for Geforce 6xxx,7xxx and Radeon X1xxx)

[G_M_C]

[offtopic & Brainstorm]
Reading one of you previous postings, i understood that you treat the AviSynth fram an a texture for processing with FFT3DGPU.

That would mean you've implemented a system to transfer the avisynth fram to the texture-buffer and back.

But now the brainstorm: Since you've got a frame into the texture buffer, it would be "kind of easy" toe implement a rezise to it, wouldn't it ? Rezising textures is one of the main functions of a graphics card after all.

This would be a usefull function, because a Graphics card has much more advanced functions for resizing than AviSynth. All GPU's support trilinear resizing for instance, and AviSynth doesn't up till now. And modern GPU's even have more advanced methods than trilinear allone ("called advanced" or "optimal" or such). And added to the advanced methods, most modern GPU's can do a AntiAliasing stage when resizing, if i'm not mistaking.

And this all would be bazingly fast, since it is one of the most optimized functions of any GPU.

So: What would your thoughts be on a "GPUresize ()" function ?
[/offtopic & Brainstorm]

[Egh]

Originally Posted by G_M_C

[offtopic & Brainstorm]

So: What would your thoughts be on a "GPUresize ()" function ?
[/offtopic & Brainstorm]

Called "Haali renderer" As for advanced resizing, I dont' think cards natively support Lanczos. And even bicubic resizing is not straightforward and require quite a lot of GPU resources.

[Fizick]

http://www.avisynth.org/JapanesePlugins

[G_M_C]

Originally Posted by Fizick

http://www.avisynth.org/JapanesePlugins

Thx,

never thought of looking in that section ... because my Japanese isn't up to spec atm

Guess i missed something

Originally Posted by Egh

Called "Haali renderer" As for advanced resizing, I dont' think cards natively support Lanczos. And even bicubic resizing is not straightforward and require quite a lot of GPU resources.

GPU resources I have enough (I think) since my new card arrived

But what i meant is that there are some "standard" filtering / resizing methodes build into the hardware of any GPU that have a equivelant command in AviSynth (bilinear). Those could be "transferred" to the GPU to gain speed, at least that was my thought. And doing so could gain at least one extra resizing method; Trilinear, wich isnt in AviSynth, but is standard in all GPU's.

[foxyshadis]

Trilinear isn't made for the totally planar video frames, though. It'd be interesting if someone came up with a good way to use it, but the way trilinear works is that it uses prerendered larger and smaller textures, interpolates both to the desired size, and averages them, weighting based on how close the result is to each prerendered size. (The gpu has certain efficiency requirements, like each step up being a power of two.) Thing is, how often do you have two videos, one twice the size of the other, that you can pass to the gpu?

I can only name one: When using eedi2 reconstruction.

You'd be better off implementing spline or gaussian, or even some crazy vector/fractal resize, in the gpu if you want the top quality, or bilinear/bicubic for hardware asic speed.

[G_M_C]

Ive been watching GPU programmabillity for some time now, and awaiting the first atual uses it would have for me personally. FFT3DGPU is one of the first more "gereral" apps designed to take advantage of the computational power of the GPU.

Stanford univerety has been doing a lot of research in to the usage of GPU's in geberal progamming. Much of their research is open sourced, and might be of some use.

So i give the link: http://graphics.stanford.edu/project...gpu/index.html

[Egh]

Originally Posted by G_M_C

Ive been watching GPU programmabillity for some time now, and awaiting the first atual uses it would have for me personally. FFT3DGPU is one of the first more "gereral" apps designed to take advantage of the computational power of the GPU.

I would call Haali renderer much more "general app" than FFT3DGPU. Besides, even older builds of mpc used bicubic resizing implemented with PS 2.0 (i.e. GPU was used).

[foxyshadis]

3D renderers, like Maya, and video software like Combustion and AE, have been able to use them for years too, although mostly limited to the Quadro/FireGL lines of pro cards (because the GeForce/Radeon drivers are slightly crippled, though hardware is nearly identical, a bit of a scam). AGP also wasn't really suited to GPU computation at all - returning to main memory is painfully slow - so the PCIe transition has started the evolution now.

[tsp]

Another GPU filter by antitorgo: http://forum.doom9.org/showthread.php?t=86793
As foxyshadis stated trilinear filtering wouldn't be all that usefull even if non power of two textures can be used (as on geforce 6xxx/7xxx and maybe radeon X1xxx) more advanced resizers can be used as postprocesing using shaders (and is used in virtualdub) but currently it is not one of my highest priority.
G_M_C: The first two versions of fft3dgpu was created with brook.

[swaaye]

Well honestly the GPU is tapped out pretty much just doing fft3DGPU. If you're running a dual core setup the GPU quickly becomes the bottleneck, unless of course you were using fft3dfilter or some such. I didn't even bother with noise filtering until this GPU filter. Speed is fairly critical for me. The normal resizers don't seem to have that much overhead relatively, especially DIVX's multithreaded resizers in their codec.

I'd like to hear from someone endowed with a SLI/Crossfire setup.

[tsp]

You can use another GPU filter after fft3dgpu without interfering with it. Unfortunatly SLI/Crossfire doesn't work with fft3dgpu as there would be excesive texture transfers from one GPU to anhother when running in Split frame rendering and Alternative frame rendering wouldn't work because fft3dgpu needs the result from the previous frame.

[tsp]

new test version 0.6.2.9 that includes new fft code. Should improve speed for larger bw, bh size(>32). Doesn't work with geforce fx 5xxx card currently. Please try it and report back if you experience slowdown or other problem with.
Link:
http://www.avisynth.org/tsp/fft3dgpu0.6.2.9.exe

[modsoul]

i ried using this simple scriupt as a beginnig test

directshowsource("C:\anime\black

cat\[Ani-Kraze]_Black_Cat_-_01_[C93856CD].avi")
ConvertToYV12()
fft3dgpu (sharpen=3)

but i get this error
error X1507 failed to open source file C:\program file\ avi synth 2.5\plugins\ps.hlsl

what am i doing wrong.

[Boulder]

Do you have the ps.hlsl file in your Avisynth plugins directory?

[modsoul]

NOPE. its not there. where can i find. it.

[modsoul]

ok i renamed fft3dgpu.hlsl to ps.hlsl ( i did not remove fft3dgpu.hlsl)
is that what i am supposed to do
forgot to mention. now it gives error
"sharpen entry point not found.
is my script ok ?
ok i downloaded version 0.47 and that seems to have the pd.hlsl file. usingthat file and the dlls from the latest version i can run it. am i doing this all wrong ?

[AI]

Originally Posted by modsoul

where can i find. it.

see first message in this thread

PS manual