madVR - high quality video renderer (GPU assisted)

[StinDaWg]

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.

[flashmozzg]

Originally Posted by StinDaWg

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.

Have you updated madVR itself? Is interop hack enabled?
Have you not installed anything into your PC PCIe slot? NNEDI3 on amd is very PCIe speed dependent. (You can try to see the pcie speed for your gpu in gpu-z)

[Anime Viewer]

Originally Posted by StinDaWg

I don't know what the heck happened but I'm getting ridiculous rendering times using NNEDI3 after updating to the latest version of madVR. 20ms higher render times on my AMD 7850.

I'm going to restart my computer to see if something got messed up.

Edit: Nope. I'm at a loss for words. I haven't changed anything on my end except to continue to update Windows 8.1.

My system is different from yours in quite a few ways, but I haven't seen any drastic increase between versions.

Did you uninstall the previous version before you installed the new version? If you did uninstall are you sure you re-added all the settings? (Like display resolution and refresh rates?) You used to have have display modes set, and as you know they can be a factor. You didn't put any checks in the new features that you might not have used previously (up-scaling refinement: SuperRes, FineSharp, or image enhancement) did you? You didn't uncheck any trade quality for performance check marks you previously had checked, did you?

[StinDaWg]

I rolled back to an old version of madVR and still have high render times so that's not the issue. This started a few days ago after Windows did a system update and rebooted. GPU usage is very high now. 91% with the same NNEDI3 settings where it was around 50% before.

[Anime Viewer]

Originally Posted by StinDaWg

I rolled back to an old version of madVR and still have high render times so that's not the issue. This started a few days ago after Windows did a system update and rebooted. GPU usage is very high now. 91% with the same NNEDI3 settings where it was around 50% before.

Perhaps it is NNEDI3 related since I haven't used NNEDI3 in quite a while, and I didn't have that problem (I've been using NEDI). Are you using NNEDI3 for only image doubling, or something else? (I know NNEDI3 for Chroma upscaling can increase render times x4.

[StinDaWg]

Only doubling.

[Werewolfy]

@madshi

I tested the latest version of Shiandow's debanding and I still prefer the high preset. Most of the time, there is no clear winner but I found an example where the high preset is clearly superior, you have to set very high values in Shiandow's debanding to be on par with madVR's debanding and by doing that you lose a lot of details.

High preset : [img=http://s5.postimg.org/6enu7he9z/high1.png]

Shiandow - power : 0.80 - margin : 0.00 : [img=http://s5.postimg.org/qnb7t7dl3/shiandow085000.png]

Shiandow - power : 0.85 - margin 0.30 : [img=http://s5.postimg.org/9ckt16lxj/shiandow085030.png]

Shiandow - power : 0.90 - margin : 0.50 : [img=http://s5.postimg.org/ushj1dlzb/shiandow090040.png]



Sometimes Shiandow's algo is better, sometimes it's weaker. MadVR's debanding smoothes the banding and Shiandow's algo blurs it. It also blurs compression artefacts which could be usefull but it's not really the role of a debanding algo and I think that a denoise or a deblocker will be a lot more efficient at this task.

I also played with the high preset and I found a way to improve it a little bit by setting Angleboost at 1.4 and maxAngle at 0.10.
You don't lose more details than with the original high preset and it can help sometimes.

[img=http://s5.postimg.org/gnbpzkcxz/highc1.png]
(don't pay attention to avgDif, it's a mistake and it dosen't change anything)

I know that the difference is not huge but you don't lose more details by doing that so I don't really see why not doing it With these settings, I find hard to find an example where Shiandow's algo is better.

[InstantAli3n]

I must agree with Werewolfy. After switching between the algo's for a few days I've come to the conclusion that madVR's high preset does a better job with less blurring.

n00b question: Is it normal to get frame drops above about 85% GPU usage? I don't reach 90% at all but if I use any options that get me between 82-87% I get frame drops.

GTX 770, 347.88

[Asmodian]

I have wanted to provide feedback on debanding but my experiments have failed to provide useful distinctions (unless a report of inconclusive is useful).

Depending on the source and the exact banding and details I am paying attention to I can sometimes notice differences but I have not found anything that would allow me to feel comfortable saying generally one is better than the other.

I generally use no or light debanding, conceptually I prefer viewing banding to losing details. I haven't done any blind tests to determine what I actually prefer.

Originally Posted by InstantAli3n

n00b question: Is it normal to get frame drops above about 85% GPU usage? I don't reach 90% at all but if I use any options that get me between 82-87% I get frame drops.

GTX 770, 347.88

I don't start dropping frames but I get a lot of presentation glitches when using settings that get above the high 80% GPU usage. I think this is normal. I seem to max out at ~95% usage and start dropping frames with settings that would use above that.

GTX Titan, 352.86

[ryrynz]

Originally Posted by Werewolfy

@madshi

I still prefer the high preset. Most of the time, there is no clear winner but I found an example where the high preset is clearly superior, you have to set very high values in Shiandow's debanding to be on par with madVR's debanding and by doing that you lose a lot of details.

Easy win there for f3kdb.

[Sunset1982]

I got some quesion for you guys:

I'm using MadVR, LAV and MPC-HC on an i5 4590 CPU + AMD R9 270 GPU.

Next week i'll get a 4K TV. So now I want to upscale my movies to 4K resolution.

Is this possible with good quality on my current hardware or do I need a better GPU?

What settings should I use to get best quality? Or should i try to get the best quality 1080p image out and let the tv do the upscale? (LG OLED)

[huhn]

Originally Posted by Sunset1982

I got some quesion for you guys:

I'm using MadVR, LAV and MPC-HC on an i5 4590 CPU + AMD R9 270 GPU.

Next week i'll get a 4K TV. So now I want to upscale my movies to 4K resolution.

Is this possible with good quality on my current hardware or do I need a better GPU?

What settings should I use to get best quality? Or should i try to get the best quality 1080p image out and let the tv do the upscale? (LG OLED)

24 hz 1080p can be upscaled with nnedi3 16 and nedi using a r9 270. i highly doubt a TV scaler can get close to this.

and don't forget forum rule 12: http://forum.doom9.org/forum-rules.htm

[ashlar42]

I seem to be unable to activate the D3D11 presentation mode.

I am on Windows 7 x64, madVR 0.88.8, LAV Filters 0.65.0 and XySubFilter 3.1.0.705.
dxdiag reports DirectX 11 as correctly functioning, GPU is GTX 660 with Nvidia drivers 352.86.

No matter the settings combinations I try on Rendering settings, I always get D3D9 with CTRL+J.
But if I choose the DX11 only dithering methods, they appear to be working just fine.

Anything worth checking?

I tested low-mid 720p SMPTE 170M content I had with debanading algorithms and I'd have to agree with Werewolfy also. Shiandow's makes some places with low dark levels look smoother, but at the cost of removing detail. It also makes those smoothed-out areas stand out against the rest of the picture. So, for example, madVR's mid/high de-banding ends up showing noisy bands / compression squares in dark shadows, without taking out much detail and without making these noisy bands stand out. There is a rather smooth gradient transition from dark shadowy parts to brighter parts of the image. It feels natural - uniform. If you add or use only Shiandow's algorithm, that same dark shadow becomes really nice and smooth, but it loses detail and that part stands out against the rest of the image. It makes the dark shadow part look like 2-3 giant bands in my tests and on Werevolfy's posted image comparisons. It kind of makes it seems like the image uses / is of lower bit-depth...

I do think more tweaking could make it useful though. It needs to get better at 2 things - 1. preserving more detail at the cost of decreased smoothness and 2. merging smoothed out "banded" content with "non-banded" content better. Otherwise it simply creates fewer, smoother, but more obvious bands with less detail. If I had to pick what to fix - I'd rather fix #2 problem before #1 because loss of detail is not as much of a big deal for lower-quality content, but the look of smoothed out bands standing out against the rest of the image is quite annoying.

P.S. I feel SO bad critiquing other people's hard work...

[ryrynz]

Originally Posted by huhn

24 hz 1080p can be upscaled with nnedi3 16 and nedi using a r9 270. i highly doubt a TV scaler can get close to this.

Actually LG's 4K sets do something very similar to NNEDI3.

[Shiandow]

Originally Posted by XMonarchY

I do think more tweaking could make it useful though. It needs to get better at 2 things - 1. preserving more detail at the cost of decreased smoothness and 2. merging smoothed out "banded" content with "non-banded" content better. Otherwise it simply creates fewer, smoother, but more obvious bands with less detail. If I had to pick what to fix - I'd rather fix #2 problem before #1 because loss of detail is not as much of a big deal for lower-quality content, but the look of smoothed out bands standing out against the rest of the image is quite annoying.

P.S. I feel SO bad critiquing other people's hard work...

I think the main cause for #2 is that MadVR's banding adds dithering, which tends to hide those "larger" bands a bit more. In theory what my debanding algorithm does is a bit closer to the source, but that's not much help if the source isn't good. Might be worth investigating if adding dithering helps, fortunately that's quite easy. It might also be possible to remove those larger bands entirely, but that would require reverse engineering how they were created which could be difficult, and might even be different for each file / frame, so that's not really something you can do with a general purpose algorithm.

Thanks for the criticism by the way, you were very clear in describing how and why it looked better / worse, which is very helpful, so there's no need to feel bad about it.

[Sunset1982]

Who else is using madvr with an amd card?

what settings should i use in Madvr and catalyst to get best results?

currently I'm using:

madvr: 16-235

AMD Controlpanel: RGB 4:4:4


is that ok or should I use something else?

[huhn]

Originally Posted by Sunset1982

Who else is using madvr with an amd card?

what settings should i use in Madvr and catalyst to get best results?

currently I'm using:

madvr: 16-235

AMD Controlpanel: RGB 4:4:4


is that ok or should I use something else?

i use a r9 270 my self.

all these settings depend on the screen and what the screen does with it.

in theory sending unlimited RGB to the GPU and outputting unlimited is the best way.

[iSunrise]

Originally Posted by ryrynz

Actually LG's 4K sets do something very similar to NNEDI3.

I highly doubt it. How do you know? Did you do side-by-side comparisons?

Something along the lines of NNEDI3 is pretty wasteful of your GPU resources. And upscaling anything to 4K with NNEDI3 requires a lot of performance that usually only discrete desktop GPUs can handle.

This would be WAY too expensive for LG to put into their scaler unit, even for an ASIC that has largely fixed-function units, I really highly doubt they would go that far.

I have not seen real-world comparisons though, but if it's another variant of the so-called "Super Resolution" algorithms, I bet that not even Sony can deliver what madVR can do.

Feeding a 4K panel with very low-resolution content should be a good way to judge, but when you feed at least 720p or 1080p to it, differences will probably be very hard to see. Even the use of a very mild sharpener and basic upscaling suddenly makes Marketing-Departments scream "Super Resolution".

I have never owned a 4K display though and I haven't done comparisons with madVR on any of them, so consider this just an opinion, based on the way the monitor industry works, meaning they usually only implement small improvements, if at all.

[dbcooper]

Originally Posted by huhn

i use a r9 270 my self.

all these settings depend on the screen and what the screen does with it.

in theory sending unlimited RGB to the GPU and outputting unlimited is the best way.

Do you have a TV that accepts RGB 0-255 at 24p? How about 4:4:4?