madVR - high quality video renderer (GPU assisted)

[leeperry]

N3 and NL2 are close cousins but NL6 blows them both out of the water, the former are like watching through a window but the latter is opening the darn window and getting your head out

It's even more striking on 720p60 content than 1080p24, yay!

[Shiandow]

Originally Posted by madshi

Yeah, quite interesting. Unfortunately it doesn't work this way for me. My code got more complicated and used more registers. Actually I'm a bit surprised that the performance hit is as small as it is with my code.

Anyway, does the change improve image quality in your tests, too?

The differences in image quality between the various versions are getting a bit too small for me to actually convincingly say which is better. I'm actually having trouble to tell them apart on some of the 'enhanced' images that were posted.

FWIW the new version does cause a 2%~3% increase in total GPU usage. Previously enabling error diffusion increased GPU usage by about 13%~14%, but with the new version this is closer to 16% so in my case the new error diffusion uses about 18% more GPU than the old one. This is on a Nvidia GTX560Ti.

[Ver Greeneyes]

Originally Posted by leeperry

N3 and NL2 are close cousins but NL6 blows both them both out of the water

Glad to see you have a strong preference again (I was wondering if the addition of the limiter would bring them closer together)

[leeperry]

Me too, who needs 3D when you got NL6

[madshi]

Originally Posted by Shiandow

newerror = (f(x) + error) - result;

Shouldn't it be: "newerror = (f(x) + error) - f(result);"?

Originally Posted by nevcairiel

high noise is random dithering.

Exactly!

Originally Posted by Shiandow

The differences in image quality between the various versions are getting a bit too small for me to actually convincingly say which is better. I'm actually having trouble to tell them apart on some of the 'enhanced' images that were posted.

Same here. The "limiter" did bring a visible improvement for me, though. At least it reduces/removes stray dots (from gray bars and pure black surfaces etc).

[Shiandow]

Yeah that should indeed be f(result), thanks.

[bacondither]

So many builds...
I used the 16-bit "gradient-perceptual.mkv" at the specific frame of "59, 29.970" and clipped the histogram at 4.

Just to give a little perspective i tested the "old" DC3 build and the new NoisyLimited. The improvement is doubleplus good!

DirectCompute3
NoisyLimited

I'm perfectly satisfied with the NoisyLimited build but will test the new randomised builds sometime in the weekend.

[madshi]

Originally Posted by bacondither

So many builds...

Yeah, I know. And these are just the medium-noise builds. I still might offer a lower-noise option. But I'll work on that only after we've decided on the medium-noise builds.

Where can I get that MKV?

[bacondither]

Originally Posted by madshi

Where can I get that MKV?

Ver Greeneyes posted it earlier, here is is:
gradient-perceptual.mkv

[Ver Greeneyes]

Originally Posted by madshi

Where can I get that MKV?

It's one I made a couple of days ago, link in my signature

By the way madshi, could you try the colored gradient with debanding? Looking at the horizontal bars adjacent to black (blue and white), even with debanding set to low I see the bars noticeably bleeding over into black. I'm using NL6 from your latest package, although I don't expect that to matter - if I turn debanding off the 'mist' on the black regions goes away.

Yey! I like the latest "Noisy 2" ! I find it better than the old DC3, old ED2, and old ED4. I do see just a small amount of black dithering, but its acceptable.

I still dislike ED7 though...

I still see a lot of 3DLUT banding though - I guess that won't ever be helped...

[6233638]

Originally Posted by XMonarchY

I still see a lot of 3DLUT banding though - I guess that won't ever be helped...

I wonder if an option to apply debanding after the 3DLUT is applied rather than beforehand might be useful for this.

[TheElix]

Harry Potter and the Deathly Hallows Part 1 on Optoma GT760 720p DLP projector (around 10:00). Sitting close to a 84" screen so I see every little detail. +_+
My finalists are: NL1 and NL6. NL1 produces slightly cleaner image here in dynamic. I really can't tell be screenshots. But NL1 screenshot weights more. Does it mean it contains more information?

[iSunrise]

Originally Posted by madshi

I was surprised, too. The key processing loop just consists of a lot of math stuff, like floating point multiplications, exp(), pow() etc instructions. I don't really understand why it's so much slower with DirectCompute(), either. Could be that DirectCompute doesn't understand the way the kernel uses shared memory. Anyway, I've not finished this yet. Maybe I'll find a way to bring this up to speed. But I'm not hopeful atm...

Would it make sense to write to AMD developer support and ask for help, providing them with the critical parts of your code (code snippets?)? Never done this myself, because I am still in the NV camp, but I´ve heard that some developers had success with it and AMD gave them some advices. Not sure if that would also help on NV hardware, though.

But there clearly is something either extremely non-optimized or unnecessarily slow (if you say it´s 10 times slower, that´s a bit too much IMHO) within their DirectCompute compiler vs. OpenCL.

[DarkSpace]

Originally Posted by madshi

I'm not sure what to say or do about this. madVR tries to use OpenCL, and if it works, it works. If the OpenCL APIs error out, there's not much madVR can do about it.

I see, okay. It would be nice if there was a notice at least in the Debug OSD, though, so one can know without having to specifically test whether or not it fails.

[nevcairiel]

Originally Posted by Shiandow

Debanding should probably only act on the source, and should in theory remove all kinds of banding before the 3DLUT is applied. It seems likely that the banding is caused by applying the 3DLUT after rounding the result to 8 bits, this could be solved by applying the 3DLUT before rounding to 8 bits but then you either need a Massive LUT (256 TiB) or you need to interpolate the 3DLUT, which is slow (and sort of defeats the point of using a LookUp Table in the first place).

IIRC, the 3DLUT is 8bit input and 16bit output through interpolation.

[Ver Greeneyes]

Originally Posted by Shiandow

or you need to interpolate the 3DLUT, which is slow (and sort of defeats the point of using a LookUp Table in the first place).

Eight lookups and seven calls to lerp() I guess (assuming a HLSL shader)? I wonder how much of a difference it would make visually.

[Shiandow]

Originally Posted by 6233638

I wonder if an option to apply debanding after the 3DLUT is applied rather than beforehand might be useful for this.

I'm sort of confused on how this is done now, doesn't debanding increase the bitdepth of the original image? And if so how would you then use the 3DLUT? You'd either need a 16bit LUT which would be MASSIVE (256 TiB) or you'd need to interpolate the LUT, which sort of defeats the point of using a LookUp Table, or you'd need to round the image down to 8bits again which would make the debanding useless.

[Shiandow]

Originally Posted by nevcairiel

IIRC, the 3DLUT is 8bit input and 16bit output through interpolation.

Originally Posted by Ver Greeneyes

Eight lookups and seven calls to lerp() I guess (assuming a HLSL shader)? I wonder how much of a difference it would make visually.

Sorry about that, I realized that I was probably mistaken about how the 3DLUT was used so I deleted that post, seems that I wasn't fast enough. Luckily your answers still make sense in the context of my other post.

[Anime Viewer]

Originally Posted by leeperry

Me too, who need 3D when you got NL6

I have an LG tv that can convert 2d to 3d, and it appears with nl6 it may have more 3d enhancement compared to when I originally watched the same 2d->3d converted video using MPC-HC with MadVR back in December.

I'm all for anything that enhances 2d -> 3d given my TV, and if that is NL6 then it gets my vote.