madVR - high quality video renderer (GPU assisted)

[DigitalLF]

i just upgraded my GPU from a 5750 to a 280x but my motherboard is a PCI-E 2.0 and when i tried NNEDI3 "64 neurons" it cant handle it but its ok at "32 neurons", material is 1080p24 High-Bitrate and ffdshow is set to output 16bit P016. Could that be that my PCI-E is to slow? Do i need to upgrade? I'm very bad at all this but i really try to learn.. :(

lost frames seems to be way better with P016 then at RGB32 if that makes any sense..... but i get artifacts in windows when i use P016 and NNEDI13.

[StinDaWg]

madshi, what are your thoughts on using sharpening filters in combination with nnedi doubling? nnedi gives a very nice, sharp, artifact and ringing/aliasing free picture, but I still prefer a bit more sharpness for 720p->1080p scaling. I'm using LumaSharpen at very low values pre-resize, just enough to take out the little bit of softness that is left over from upscaling, nothing more. I think the picture looks pretty much perfect, but I'm wondering if this works against nnedi at all or its just one of those things "do what you think looks best".

[iSunrise]

@StinDaWg:
When you sharpen something, you basically add ringing. Other than that, it's all personal preference.

Personally, I always use luma doubling in combination with Lanczos 8 AR, which gives me an incredibly sharp and still mostly artifact-free result. I'm also using 1080p as an output device.

[StinDaWg]

@StinDaWg:
When you sharpen something, you basically add ringing. Other than that, it's all personal preference.

Personally, I always use luma doubling in combination with Lanczos 8 AR, which gives me an incredibly sharp and still mostly artifact-free result. I'm also using 1080p as an output device.
So nnedi just uses whatever algorithm you choose and doubles it? I guess I was under the impression that nnedi was it's own upscaling algorithm itself. I'll try Lanczos 8 AR.

[huhn]

the downscaler sharp it.

[iSunrise]

So nnedi just uses whatever algorithm you choose and doubles it? I guess I was under the impression that nnedi was it's own upscaling algorithm itself. I'll try Lanczos 8 AR.
NNEEDI is it's own algorithm, but in combination with Lanczos 8 AR, you get the sharpest results, because Lanczos 8 AR is very sharp to begin with. That's only for upscaling purposes though, I still use Catmull-Rom linear light for downscaling.

[StinDaWg]

I tried L8AR/CRAR and L8AR/L8AR and still prefer using LumaSharpen pre-resize. It takes out the softness better than Lanczos scaling alone. One thing is for certain though. nnedi is way better than Jinc when it comes to sharpness/ringing/artifacts.

I'm using
#define sharp_strength 0.5
#define sharp_clamp 0.05
#define pattern 3

[iSunrise]

I tried L8AR/CRAR and L8AR/L8AR and still prefer using LumaSharpen pre-resize. It takes out the softness better than Lanczos scaling alone. One thing is for certain though. nnedi is way better than Jinc when it comes to sharpness/ringing/artifacts.

I'm using
#define sharp_strength 0.5
#define sharp_clamp 0.05
#define pattern 3
madshi still has some tricks up his sleeve. I expect madshi to provide an even more sophisticated (or rather, more optimized for good results) sharpener, once he gets to it. Same thing with an "add grain" algorithm.

With madshi's AR algorithm, possibly even NNEDI could be improved, which I suggested before. But we would need to test that.

And yes, I also prefer NNEDI (but NNEDI only works for powers of 2) > Lanczos AR > Jinc AR. But I expect that to be highly dependent on the content and your display setup. And of course, your personal preference.

[seiyafan]

How does NNEDI in combo with Jinc3AR compare to NNEDI with Lanczos ?

[dansrfe]

Just to make sure, there's no current workaround for using NNEDI on nvidia other than downgrading the driver?

[nevcairiel]

Just to make sure, there's no current workaround for using NNEDI on nvidia unless I downgrade the driver?

Downgrading driver IS the workaround. :p

[tp4tissue]

Hey guys.. how does this nnedi thing work

does it double the resolution of the source material, then use your selected (downscaler) to downscale?

So if I have nnedi selected for luma, it takes 1080p and makes it 2160p ?

then use the Catmul-r I have selected for downscaling, and outputs?

I'm confused.. :sly:

Also I'm outputing to 2560x1600... and getting like super slow fps.. is there something to configure, or is my card just not strong enough? 7970 crossfire, but i know cross fire doesn't do jack here, LOL

using 13.12 cat.. is there some opencl stuff i gotta install?

what magic is Madshi using to get his vids to play nnedi on his 7770 ?

[mikecks]

Downgrading driver IS the workaround. :p
and which driver version would this be?
regards.

[Asmodian]

Also I'm outputing to 2560x1600... and getting like super slow fps.. is there something to configure, or is my card just not strong enough? 7970 crossfire, but i know cross fire doesn't do jack here, LOL


Oh a crossfire user!

If you hit control-J you get a madVR OSD that will tell you the average rendering times (along with a lot of other information).

What are your average rendering times with and without crossfire enabled? I am wondering if performance is worse with crossfire enabled like it is with SLI enabled.

And you have it right, NNEDI3 is an algorythim that can only double the resolution. Image down scaling is used to get to the display resolution if needed and image upscaling to bring chroma up if it wasn't doubled.

and which driver version would this be?
regards.

327.23 are the newest that work.

[StinDaWg]

New madVR levels :D

http://www.avsforum.com/t/1477339/so-youve-built-your-htpc-now-what-is-next-how-to-get-the-ultimate-picture-and-sound-quality-from-your-htpc-madvr-svp-xbmc-mediabrowser-jriver/400_100#post_24521720

http://cdn.avsforum.com/5/55/553cafaf_madVRlevels.PNG

[ryrynz]

I find the other graphs more interesting.. No NNedi doubling for Chroma at the highest levels? That's one of the main reasons I'd want a better graphics card. Guess that's level 6.

[nevcairiel]

I think such "levels" are not such a great idea when it comes to NNEDI, there is just so many different options that spreading such pre-defined levels can cause more confusion then it helps to solve.
Before, defining the performance levels of the scalers alone was OK, as most people agreed on the order of quality/performance, and if they didn't they usually knew why, but now..

There is so many cases where you might be able to use NNEDI on SD with a "worse" image scaler, but not on 720p, where you want a better image scaler. "Levels" are not going to cut it.
Not to mention the huge differences between NVIDIA and AMD at this time (in performance, and with NVIDIA even only working with certain drivers)

[StinDaWg]

I find the other graphs more interesting.. No NNedi doubling for Chroma at the highest levels? That's one of the main reasons I'd want a better graphics card. Guess that's level 6.
I don't really find nnedi doubling to be anything special for chroma. The biggest improvement by far is with luma. I actually don't see any difference between nnedi chroma double and bicubic 75AR, but it's way more taxing on the gpu.

[ryrynz]

I don't really find nnedi doubling to be anything special for chroma. The biggest improvement by far is with luma. I actually don't see any difference between nnedi chroma double and bicubic 75AR, but it's way more taxing on the gpu.

6233638 did some testing on it and with the results I saw, I would gladly dedicate the GPU.. I think levels can be a good guideline for the lazy but you really need to open up those MadVR profiles when working with NNEDI.

[iSunrise]

6233638 did some testing on it and with the results I saw, I would gladly dedicate the GPU.. I think levels can be a good guideline for the lazy but you really need to open up those MadVR profiles when working with NNEDI.
I think you're confusing NNEDI 4:2:0 -> 4:4:4 chroma upscaling with NNEDI chroma doubling. NNEDI chroma doubling hardly makes a difference when not going to extreme scaling factors, whereas NNEDI chroma upscaling is definitely giving a real good benefit all around. And if you're upscaling just from 720p to 1080p, even in that case the differences are quite hard to spot without actually zooming in. Where NNEDI chroma upscaling makes the most difference is with lower resolution content or, more generally, way higher scaling factors.

The reason for that is that the 4:2:0 -> 4:4:4 chroma upscaling happens before doubling in the madVR pipeline/chain, so it's also at a more critical place in the chain, which to a rather large part determines the end result.

[ryrynz]

Where NNEDI chroma upscaling makes the most difference is with lower resolution content or, more generally, way higher scaling factors.


I actually meant full NNEDI chroma upscaling not doubling, hard to remember what's what when you don't have enough GPU power to even enable it on any of your systems. Looks like Jinc 3AR is the nest logical step without requiring insane GPU performance.

[QBhd]

I agree that these charts are not something that should be pushed as gospel... there is far to many variations in content, target resolutions, gpu performance and neuron count to recommend hard "levels" like this.

QB

[Asmodian]

I also agree, no way to make a reasonable chart that includes NNEDI3.

Tuning for maximum image quality (only IMO of course) for 720p24 -> 2560x1440 I ended up with NNEDI3 128 Luma doubling which left room for Jin3AR chroma + image resizing. Going to 64 neuron luma doubling for 32 neuron chroma scaling or 16 neuron chroma doubling is not worth it. I do like NNEDI3 chroma; I think if I could run 128 luma doubling with NNEDI3 for the chroma too I would but then again I haven't tested 256 neuron luma doubling. ;)

Of course this is a better option in the winter. :p

[seiyafan]

How come I cannot see any difference between NNEDI3 chroma upscaling and nearest neighbor chroma upscaling? Should my eyes be literally inches away from my 27' 1440p monitor to notice a difference?

[qduaty]

I have nnedi3 working on GTX660/GK106 and Nvidia 335.23 drivers (and, it's 20% faster). I blocked deleting of OpenCL/D3D interop images between frames (madvr gets existing instances mapped to IDirect3DTexture9*) and it seems to fix the problem entirely.

[Vyral]

New madVR levels :D

http://www.avsforum.com/t/1477339/so-youve-built-your-htpc-now-what-is-next-how-to-get-the-ultimate-picture-and-sound-quality-from-your-htpc-madvr-svp-xbmc-mediabrowser-jriver/400_100#post_24521720

http://cdn.avsforum.com/5/55/553cafaf_madVRlevels.PNG

What about debanding, dithering and smooth motion ?

Right now, my GPU have no problem to handle Jinc 3AR, Jinc 3AR, Catmull-Rom AR/LL + debanding + ordered dithering + smooth motion with everything disabled in trade quality for performances. I can't use NNEDI due to interopt cost, and error diffusion increase the rendering time by 10ms (without dropped/delayed frame though).
If AMD fix the interopt problem in their drivers, I should be able to use the level 4b settings.

[seiyafan]

I recommend 270x as the minimum graphics for NNEDI3 doubling.

[GREG1292]

I have nnedi3 working on GTX660/GK106 and Nvidia 335.23 drivers (and, it's 20% faster). I blocked deleting of OpenCL/D3D interop images between frames (madvr gets existing instances mapped to IDirect3DTexture9*) and it seems to fix the problem entirely.
How and where do we do this.

[SUPERBIF]

Newbie questions:

I have the following setup:
i3-2100 CPU 3.10 Ghz
8 GB memory
HD6450 Saphire

1. Will this setup be okay for madvr and how high should I expect settings wise? Can I use perhaps jinc3?
2. If I should get better madvr result should I then replace my graphics card with HD6670 or should I upgrade CPU? Does madvr more rely on CPU or GPU?

[seiyafan]

Newbie questions:
2. If I should get better madvr result should I then replace my graphics card with HD6670 or should I upgrade CPU? Does madvr more rely on CPU or GPU?

GPU only.

[qduaty]

How and where do we do this.I did it with a library hack and only madshi can fix it permanently in madVR.

[markanini]

I recommend 270x as the minimum graphics for NNEDI3 doubling.

Seems overkill for common sources. I'm surprised at what my HD7770 can do myself, 1.2x and 2.4x doubling is smooth on all >30p sources.

[SUPERBIF]

GPU only.
Thanks, but what about my current setup. Will that be good enough and what level of settings should I hope to reach?

[seiyafan]

Thanks, but what about my current setup. Will that be good enough and what level of settings should I hope to reach?

I can't say for sure as I am not familiar with your setup but you should be able to tell by looking at the rendering time.

[Vyral]

Thanks, but what about my current setup. Will that be good enough and what level of settings should I hope to reach?

I've more or less the same specs (same CPU, 4GB memory, HD6850)
On Niyawa's madVR scaling algorithms chart (http://i.imgur.com/HLv1qmj.png), I can use the highest preset. IMO, with your GPU, you can try High or Mid.

[SUPERBIF]

I've more or less the same specs (same CPU, 4GB memory, HD6850)
On Niyawa's madVR scaling algorithms chart (http://i.imgur.com/HLv1qmj.png), I can use the highest preset. IMO, with your GPU, you can try High or Mid.
Thanks for the answer but isn't hd6850 more powerful than hd6450?

[daglax]

I posted this a few pages ago and i do it again:

madvr's fps estimation is broken on any live sources. Madvr estimates the framerate to what the source says and not to what the actual decoder/post-processor says.
the problem is, that first smooth motion isn't working correctly on movies/tv-shows which sometimes (not very often either) switch framerate. That's a minor thing, because it's pretty rare.

the real problem is, that you can't use SVP or any other frame interpolation method correctly with madvr. first smooth motion's estimations are wrong, second NNEDI is way waaay to demanding on 60 fps (or more) videos.

is there any chance, this could be fixed?

[seiyafan]

It's all about the interop bug. I remember with 290x if I set the number of neurons to 16, the GPU utilization rate was never more than 30%, however the rendering time was still not short enough to not cause frame drop for 60p videos.

[huhn]

I posted this a few pages ago and i do it again:

madvr's fps estimation is broken on any live sources. Madvr estimates the framerate to what the source says and not to what the actual decoder/post-processor says.

are you sure the decoder/post-processor is even trying this? if svp is not giving this information how can madvr fix this?

the real problem is, that you can't use SVP or any other frame interpolation method correctly with madvr. first smooth motion's estimations are wrong, second NNEDI is way waaay to demanding on 60 fps (or more) videos.

smooth motion works fine with VFR files so it used timestamps not fps to do this work if the timestamps of svp are right smooth motion should work.

and about nnedi, if you have to use is 2.5 time is normal that it needs a lot of gpu power what is wrong with this?

[daglax]

I don't know what you mean by VFR files.
But i do know that the decoder sure sets the fps output right. It's madvr that doesn't get it. If i hit ctrl+4 in MPC-HC, it says the correct fps. And ffdshow raw also shows to right fps.

About NNEDI: There's nothing wrong with high gpu usage, you missed my point :). I was saying that 60 fps with high NNEDI-settings don't work for me, so i need a profile, that chooses the settings according to the source video frame rate - which madvr can't display right.

[huhn]

the statistics in mpc-hc doesn't mean madvr gets the info. it just show how many frame are displayed per sec. these numbers are calculated over time but madvr need them be for it starts playback. these number can change the hole time did you want madvr to change the display mode the hole time? and these numbers are inaccurate too.

this can may be get fixed in ffdshow. ffdshow just has to override the movie frame interval. but i'm pretty sure this will create a lot of problems...

if i'm not miss informed SVP starts delayed

VFR =variable frame rate

[daglax]

ok thanks for helping. For now i live with the hotkey option. As i'm watching not that much in 60fps, this isn't a huge problem anyway.
Maybe it'll be working in future releases, but i doubt it.

[rurifan]

madshi, any thoughts about the impact of dithering on the motion interpolation algorithms in modern HDTVs? The new "change dither for every frame" option caught my attention considering that such algorithms rely on tracking things across frames, and introducing temporal noise might reduce their success.

[*Touche*]

Ugh, why would you ever want to use those?

[e-t172]

So, umm...

I know I might be getting annoying with my frame drop issue, but this time it's clearly gone into paranormal voodoo magic head-banging territory.

I just got a GTX 770 (reminder: I have a 780 Ti) from a colleague for testing. Also note that for a while now I've been using a clean Windows 7 install on a separate drive dedicated to madVR testing.

Also bear in mind that I already changed motherboard, CPU, RAM, and even PSU (and the Windows install is on a different drive).

I'm still getting frame drops (one frame drop typically every 30 minutes or so). On the 770. With drivers 332.21, 335.23, and 327.33.

W T F.

*Every* single hardware and software component has been swapped. Completely fresh hardware with fresh Windows install and testing multiple driver versions. These frame drops keep haunting me everywhere I go. At this point I'm starting to wonder if I'm not becoming insane.

So, to the people who are reporting 0 frame drops even on long playbacks with smooth motion on, here's the one million dollar question: are you doing any special configuration? If you are changing power management settings, how exactly are you doing it? Hand-picked obscure driver version? Do you have non-standard flush settings, queue sizes, "use a separate device for presentation", "disable composition" settings? Are you using specific settings in the NVidia control panel? Are you using ReClock? Weird BIOS settings? Or any other weird trick that I don't know about that might be necessary to achieve perfect playback? Because, clearly, from my tests it is crystal clear that if some people manage to get it working, some default setting had to be changed somewhere and I don't know about it.

At this point I realize that there is one, and only one, hardware component that has not been swapped: the monitor. Oh well, I'll try that tomorrow. At this point I might just as well try black magic incantations.

[rurifan]

Ugh, why would you ever want to use those?

Unfortunately, content producers have left us stuck with 24fps which is insufficient motion resolution to avoid wretched judder and strobing (http://www.projectorcentral.com/judder_24p.htm) everywhere. Motion interpolation algorithms are not perfect (some better than others), but we're left picking the least-bad option: constant judder/strobing vs occasional interpolation artifacts.

Hopefully we can at least keep these algorithms in mind so that madvr doesn't do things which reduces their effectiveness.

[Audionut]

Don't use the option. Problem solved.

[DarkSpace]

madshi, any thoughts about the impact of dithering on the motion interpolation algorithms in modern HDTVs? The new "change dither for every frame" option caught my attention considering that such algorithms rely on tracking things across frames, and introducing temporal noise might reduce their success.
That sounds like your best option would be to disable dithering altogether (use simple rounding), as even static dithering introduces some patterns. And that sounds just wrong.
So, I'd propose that if you really need Motion Interpolation, you experiment some and just use what works best for you...

[seiyafan]

I'm still getting frame drops (one frame drop typically every 30 minutes or so). On the 770. With drivers 332.21, 335.23, and 327.33.



Considering movies are usually two hours long, so you drop on the average 4-5 frames for one entire movie? :p That's about 0.2 seconds of frame dropping per ENTIRE movie.

On the other hand, blink of an eye takes 0.2 to 0.3 seconds. The average person blinks some 15-20 times per minute, or 1800-2400 times per a movie. According to the theory of probability, you chance of see a frame drop is pretty low.

[*Touche*]

*Every* single hardware and software component has been swapped. Completely fresh hardware with fresh Windows install and testing multiple driver versions. These frame drops keep haunting me everywhere I go. At this point I'm starting to wonder if I'm not becoming insane.

Not sure if you've already tried or if it could make any difference whatsoever, but can you try with another display?