madVR - high quality video renderer (GPU assisted)

[nsnhd]

I use a Full HD screen so it's all NGU up and SSIM2D down for me. I barely bother with additional sharpening.

I'm using gtx 1060 on FHD also. What chroma are you using? And what actually does activate SuperRes filter on chroma upscaling ? I select it on Jinc and it doesn't cost much on performance, maybe 2ms.

[ryrynz]

I went with NGU AA medium/High depending on what profile activated.

Super Res is is an image enhancement algorithm which does the following.

Calculate an initial guess (using a different upscaler)
Downscale and calculate differences with original image.
Scale those differences to the final size
Improve guess by:
Softening the image
Subtracting differences with the original image
Sharpening (optional)
Removing aliasing (optional)
Removing ringing (optional)
Repeat steps 2-4 several times.

Judge for yourself if you like the results, I've never felt I've required it, especially for chroma.

[QBhd]

I'm pretty sure that most if not all users thought Super Res was detrimental when used with NGU. Correct me if I'm wrong please...

QB

[chros]

I'm pretty sure that most if not all users thought Super Res was detrimental when used with NGU. Correct me if I'm wrong please...
Since we are here, can I ask you guys what settings (luma, chroma, enhancement, etc) you use to double 1080p / 720p content? Thanks

[ryrynz]

I'm pretty sure that most if not all users thought Super Res was detrimental when used with NGU. Correct me if I'm wrong please...


Most users don't use it, can't say I ever would either. Not required IMO.

[nsnhd]

I went with NGU AA medium/High depending on what profile activated.

Super Res is is an image enhancement algorithm which does the following.

Thank you for infos. With NGU it looks not good, but I'll compare between Jinc with and NGU without it.

[ryrynz]

Please do your research into chroma before saying it's "not good" the standard and AA variants of NGU are known to be the best chroma upscalers along side Jinc and a few others. Your preference may be sharp scalers.. Personally I prefer natural results.

[nsnhd]

Please do your research into chroma before saying it's "not good" the standard and AA variants of NGU are known to be the best chroma upscalers along side Jinc and a few others. Your preference may be sharp scalers.. Personally I prefer natural results.
What I mean is that using SuperRes with NGU does not look good. I just want to compare SuperRes with Jinc chroma and the NGU chroma (AA or Sharp).

[huhn]

it is highly recommended to not use super res with NGU. it's just how these filter work they don't like each other.

[Warner306]

WHICH Nvidia Driver is the Recommended, that Madshi is developing on ??

There are dozens of drivers that "work." Avoid any that don't work for your needs.

[Warner306]

Considering a used 1070 or 1080 right now.

Is the additional power much greater than a 1060, (FOR MADVR, I don't play games.)

I already own a 1060 for the kitchen pc.

1060 priced ~$160 used
1070 ~$250
1080 ~$400


What extra performance setting (MADVR) can be reached by the 1070/ 1080..

I would only bother if you are using the HDR -> SDR tone mapping. A 1070 or better can help there.

[tp4tissue]

I would only bother if you are using the HDR -> SDR tone mapping. A 1070 or better can help there.

What if it's HDR tonemap to HDR output on a FakeDR TV ?

Does a 1070 help with that ?

[Warner306]

Do have all of the trade quality for performance options related to tone mapping unchecked and highlight recovery strength enabled? If yes, then the GTX 1060 can still do the job (I would guess barely). If not, a GTX 1070 or better would help.

[griffind]

I need help evolving from 'poking at settings' to 'knows what poking at settings does' :D

Been using MadVR for a while now for HDR>SDR tone mapping on my Projector and I'm wondering if I am missing out on picture quality because I don't understand the settings properly.

Windows 10 - 64-bit
i7 - 16gb ram - NVIDIA 1070ti
madVR v0.92.17
LAV - 0.70.2.81-git

- My Calibrated Screen/PJ combo is hitting approx 70 nits with a rec.709.3dlut for Gamma 2.2 and I have this chosen in the calibration section of MadVR. I don't have any other 3DLut files generated.

- For HDR>SDR I am using shaders with a target peak of 200 nits, BT.2390 curve, balanced color tweaks and high highlight recovery.
I have 'measure each frame's peak luminance' checked but not sure if it works with NVIDIA CUVID decoding.

I wondering about how much I 'dont know' and if I'm getting less than stellar results with this config.

I 'think' my picture looks good, but without having any other point of reference I feel a bit in the dark.

- For my 4k HDR BT2020 content, should I be using a different LUT?
- Should I have Gamma processing enabled?
- How do I know if I am crushing blacks or if my highlights are right?
- How do I enable dynamic peak luminance adjustments?
- Why does my cat drink out of the toilet not her water bowl?

So many questions.

[Soxbrother]

Hello,

I've installed MadVR (latest version)
and this is my system :

HP Pavilion 300-030NB
Windows 10 - 64-bit
i3 - 4gb ram - intel HD 4400 graphics
Latest mediaportal
Latest LAV codec - set to dvxa2 copy back

Connected to my Samsung UEJS9000 tv ( maximum refresh rate 30hz ( according to the "windows settings page" )

I had setup MadVR, according to what different users of an intel HD graphics 4000 (I have 4400) said, which was to use dvxa2 settings and such.

But nothing seemed to prevent my video from not playing smooth/stuttering.

I've tried numerous settings and different codecs ( Lav, CoreAvc Pro ), but nothing seemed to give me smooth playback.

Until I tried the setting under "Devices - Generic PnP Monitor - display modes"
"list all display modes madVR may switch to"

The default setting had nothing in it, so I've now set it to "1080p23" and it worked.

When I try to use 2160p23, it gives a black screen with audio only ( Lav codec - dvxa2 copy-back )
I guess the 2160p23 gives the same result as when you leave the field blank.

So I'll use 1080p23 for now.

But that leaves me thinking, now when a video is played, it will play in 1080p on a 2160p screen.
So now it's not upscaling anything, right ?

So is there any difference in quality now with this setting to "1080p23"
between MadVR and the EVR renderer option in mediaportal settings ?

I also had almost everything checked on the settings page of "Trade quality for performance".

I'm going to uncheck one at a time now and see if quality can improve without performance.

Is there any way to upscale my 1080p video's to 4k resolution with this intel 4400 HD Graphics and have smooth video playback?
Or am I stuck with the 1080p23 setting for now ?

Thanks in advance.

Greetings,
Soxbrother

[tp4tissue]

Hello,

I've installed MadVR (latest version)
and this is my system :

HP Pavilion 300-030NB
Windows 10 - 64-bit
i3 - 4gb ram - intel HD 4400 graphics
Latest mediaportal
Latest LAV codec - set to dvxa2 copy back



Soxbrother

HD 4400 benches 567 on passmark.

You will have to turn off many of the madvr's features in order to scale 1920x1080 into 3840x2160.

Madvr will still give you more color accuracy, but to use higher quality filters like Lanczos, there's simply not enough processing power to do it.

If you scale 1080p to 1080p, then that is chroma scaling only, 1:1 luma. the 4400 will do this perfectly fine. You can enable Jinc or Lanczos for chroma.


I believe you also need displayport to hdmi 2.0 adapter, because 4400's native hdmi is probably not HDMI 2.0, which is required for 3840x2160 60hz

Also, run CPU decode on Lav, do not use dxva, because you want to save as much GPU time as possible for Madvr.

Letting the CPU decode the first part is fine.

[Asmodian]

- For my 4k HDR BT2020 content, should I be using a different LUT?

Unneeded, especially since your projector is SDR.

- Should I have Gamma processing enabled

No
- How do I know if I am crushing blacks or if my highlights are right?

Use the AVS HD 709 Test patterns (https://www.avsforum.com/forum/139-display-calibration/948496-avs-hd-709-blu-ray-mp4-calibration.html)

- How do I enable dynamic peak luminance adjustments?

Check the "measure each frame's peak luminance" box in the hdr settings page.

I would also recommend not using CUDA acceleration anymore, DXVA2 copy-back or D3D11 copy-back is just as good or better and as universal solutions they are getting more development.

[tp4tissue]

So... if there are large patches of very bright data, using highlight recovery, this can cause clipping right ?

I noticed some clipping on the helmets in film -interstellar- and near the end black hole escape scene.

[chros]

Also, run CPU decode on Lav, do not use dxva, because you want to save as much GPU time as possible for Madvr.
This is not how GPU hardware acceleration works: it's a completely separate pipeline. Just use GPU hardware acceleration all the time if you can.

[Axelpowa]

Hi guys,

I'm thinking about replacing my gtx1080 with a Rtx2070 which could be quite affordable money wise.

The reason is I use my gpu almost to the limit with all the hdr new stuff and with the "reduce random noise" filter which I enjoy a lot, but uses lots of resources.

Would it be a smart move or better get a gtx1080TI?

Regards!

[nevcairiel]

A 1080 and a 2070 are quite comparable in performance, that wouldn't be a huge "upgrade" to speak of. If you can get a cheap used 1080Ti, that might be a better idea if you really need more performance.

[tp4tissue]

This is not how GPU hardware acceleration works: it's a completely separate pipeline. Just use GPU hardware acceleration all the time if you can.


I'm not sure how the pipeline works, I'll take your word for it.

But in my own testing, if I use Dxva or Dx11, the render time is always slightly higher than if I use CPU decode.

This was especially true on my iGPU and slower graphics cards.

On my HD3000 for example it will stall out /drop frames if I use Dxva decode, whereas if i let it Cpu decode, and gpu render, it will play smoothly.

[Axelpowa]

A 1080 and a 2070 are quite comparable in performance, that wouldn't be a huge "upgrade" to speak of. If you can get a cheap used 1080Ti, that might be a better idea if you really need more performance.

Thx!!!

[chros]

This was especially true on my iGPU ...
I'm not sure about iGPUs (never used them for madVR, I even used the discrete GPU in an optimus system): whether they have small amount of dedicated memory or they use the system wide RAM.

[Alexkral]

I would also recommend not using CUDA acceleration anymore, DXVA2 copy-back or D3D11 copy-back is just as good or better and as universal solutions they are getting more development.

I keep seeing this, and while I'm sure it's true for most cases I don't think it is for some. At least in my system (Windows 7, GTX 1080) I get framedrops with DXVA and D3D11 when playing 4k and up @ 60 fps, while CUVID runs smooth. This means that at least for this content CUVID is faster (it was even faster with old drivers but they produced some ugly artifacts with HDR). I don't know if this could be related with some particular rendering or scaling setting, but I don't think so because it's the same with other renderers and players. Maybe for VR in Windows 10 this is solved with that ATW thing.

[chros]

I'm thinking about replacing my gtx1080 with a Rtx2070 which could be quite affordable money wise.
According to the specs, a 1080 can be faster than a 2070 :)
https://youtu.be/CT2o_FpNM4g?t=916

[nevcairiel]

This means that at least for this content CUVID is faster (it was even faster with old drivers but they produced some ugly artifacts with HDR). I don't know if this could be related with some particular rendering or scaling setting, but I don't think so because it's the same with other renderers and players. Maybe for VR in Windows 10 this is solved with that ATW thing.

Since you mention HDR, you should know that HDR playback with CUVID will be much worse (in quality) then with any other mode, since CUVID does not actually export all the necessary metadata for full HDR reproduction.
D3D11 in Native mode should be the best you can do for performance (ie. without selecting a graphics device in LAV Video).

[huhn]

cuvid forces your GPU into high performance mode while dxva and d3d11 use the NCP power setting which is by default optimal and which is very well known to be problematic with madVR.

no matter which hardware decoder you select on nvidia you use the same decoder in the end the only thing that changes is how this data is transferred.

so go to the NCP and change the power setting to adaptive or maximum performance and try again.

[Warner306]

I need help evolving from 'poking at settings' to 'knows what poking at settings does' :D

Been using MadVR for a while now for HDR>SDR tone mapping on my Projector and I'm wondering if I am missing out on picture quality because I don't understand the settings properly.

- How do I know if I am crushing blacks or if my highlights are right

Try using these HDR10 black clipping patterns to test black clipping (white clipping is not overly important):

https://drive.google.com/file/d/1PYY37NqZLPEo1ZOwFmbUZ3T8JqyuGH7-/view

The 8-bit/10-bit combined black clipping video should be paused when making adjustments because it is short. The 8-bit values are listed along the bottom of the pattern as thick bars. The faint, grey gradient should extend as close as possible to the edge of Bar 16 from right-to-left.

The target nits is tradeoff between brightness and contrast. Increasing the target nits will give you more highlight detail, but the entire image will become progressively darker to create necessary contrast.

The other options have been simplified in the latest test build. Hopefully, it comes out soon because I find it is hard to get a bad picture with the latest build.

[chros]

so go to the NCP and change the power setting to adaptive or maximum performance and try again.
That's what I thought as well, until 2 weeks ago I had to set Optimal back in NCP because I had massive drops with Adaptive. :) Now it seem to work fine. (Setup is in my signature.)

[huhn]

so your card get's slower with more aggressive performance?

[chros]

so your card get's slower with more aggressive performance?
For whatever reason with Adaptive setting it often dropped the performance state and quickly raised it back (monitored with nvidiainspector), hence it produced bunch of dropped frames. Why? That's a good question, because it worked before it with the same setting. :)

[madjock]

Try using these HDR10 black clipping patterns to test black clipping .

These are not HDR ? Think they are just 8-10bit panel checkers ?

[Asmodian]

At least in my system (Windows 7, GTX 1080) I get framedrops with DXVA and D3D11 when playing 4k and up @ 60 fps, while CUVID runs smooth.

I had totally forgotten to mention setting Nvidia's power management mode to adaptive (in 3D settings). Optimal power often causes issues.

That's what I thought as well, until 2 weeks ago I had to set Optimal back in NCP because I had massive drops with Adaptive. :) Now it seem to work fine. (Setup is in my signature.)

I would redo that test... because setting optimal power definitely causes the GPU to clock lower more often. Did you have temperature issues or something?

[Soxbrother]

HD 4400 benches 567 on passmark.

You will have to turn off many of the madvr's features in order to scale 1920x1080 into 3840x2160.

Madvr will still give you more color accuracy, but to use higher quality filters like Lanczos, there's simply not enough processing power to do it.

If you scale 1080p to 1080p, then that is chroma scaling only, 1:1 luma. the 4400 will do this perfectly fine. You can enable Jinc or Lanczos for chroma.


What about image doubling, would that require less juice than upscaling ?


I believe you also need displayport to hdmi 2.0 adapter, because 4400's native hdmi is probably not HDMI 2.0, which is required for 3840x2160 60hz

So you can convert the DisplayPort version 1.2 to Hdmi 2.0 or 2.0a ?

If this works, how about my Dolby Atmos audio ?
Would my Dolby Atmos audio work through such an adapter ?


Also, run CPU decode on Lav, do not use dxva, because you want to save as much GPU time as possible for Madvr.

Letting the CPU decode the first part is fine.
This is not how GPU hardware acceleration works: it's a completely separate pipeline. Just use GPU hardware acceleration all the time if you can.

Thanks guys.

At the moment I believe the screen is set at 24hz.
I can go up to 30hz.
The higher the better ?
Should I set it to 30hz ?

Would 60hz with the adapter give me better quality and smoother playback ?

Thanks in advance.

[Asmodian]

What about image doubling, would that require less juice than upscaling ?

No, it needs more.

So you can convert the DisplayPort version 1.2 to Hdmi 2.0 or 2.0a ?

This would need to be an active adapter because DP and HDMI don't use the same signaling. They are expensive and I expect that most (all?) don't support Atmos.

At the moment I believe the screen is set at 24hz.
I can go up to 30hz.
The higher the better ?
Should I set it to 30hz ?

Would 60hz with the adapter give me better quality and smoother playback ?

This is somewhat complicated and depends on the source frame rate. If you are using smooth motion the higher the better (I would say 60 Hz is the minimum for decent smooth motion) but even better than smooth motion is using a refresh rate that matches the source frame rate. This means that you need to tune your refresh rate so you never get dropped or repeated frames. This can be tricky but madVR has a custom resolution tool that helps tune the refresh rate.

[Warner306]

These are not HDR ? Think they are just 8-10bit panel checkers ?

They are from a free HDR10 test pattern set. The OSD should confirm. Black clipping is not really HDR because the peak of the one mentioned is less than 1 nit, but that is intended.

[Alexkral]

so go to the NCP and change the power setting to adaptive or maximum performance and try again.

Thanks huhn, I think this helped a bit but didn't fixed it. Anyway I think I found the problem. I noticed that this only happened with SVP and very high bitrate videos (and with VR), so I downloaded a high bitrate 4k 60fps video and got framedrops with both CUVID an DXVA2. Then I blocked ffdshow raw and this fixed it for both modes. So the problem is to have ffdshow raw in the chain, but obviously it's needed for SVP. Also I noticed that you can select D3D11 but it doesn't work on Windows 7, I'll have to check what I'm missing.

@nevcairiel

Yes, I discovered this recently, but anyway I'm currently not using madVR's HDR to SDR because SVP drops the needed metadata. I managed to find some alternatives however, still using both SVP and madVR, but without dynamic tonemapping (though I have some ideas also for this). I wouldn't say that normal tonemapping is much worse than dynamic, not so long ago it was the only thing available and everybody was more than happy with it. Anyway I'm not very sure how madVR is doing dynamic tonemapping with BT.2390, because only changing Mastering display white level gives me different results, and I don't see other configurable parameters for this.

[Dogway]

Just realised that NGU causes quite a lot of coil whine, even when using it only for chroma. I bought a 1070 in black friday and didn't notice until I enabled NGU for my monitor the other day. What puzzled me is that I do GPU intense work and never had coil whine until this, I did a comparison rendering with GPU with Redshift and GPU-Z report is the same than with NGU, with further testings I found that the trigger was Memory Controller Load, at 25% it makes the buzz noise below or above doesn't, the problem is that NGU very high (above 25%) heats my card like a toaster (running it in a mATX case)

[tp4tissue]

Just realised that NGU causes quite a lot of coil whine, even when using it only for chroma. I bought a 1070 in black friday and didn't notice until I enabled NGU for my monitor the other day. What puzzled me is that I do GPU intense work and never had coil whine until this, I did a comparison rendering with GPU with Redshift and GPU-Z report is the same than with NGU, with further testings I found that the trigger was Memory Controller Load, at 25% it makes the buzz noise below or above doesn't, the problem is that NGU very high (above 25%) heats my card like a toaster (running it in a mATX case)

That's how you know it's Lookin' Guud :D

[tp4tissue]

Guys, For HDR 10bit, on nvidia 1060, Does setting 4:2:2 10bit/ 12bit on the NVIDIA control panel make a difference ? OR, does it auto pop to the correct output setting when NVHDR is engaged..


And also, does setting 10bit in the Madvr control panel cause issues for 8bit outputs ?

[Asmodian]

madVR's dithering is better than the GPU's, and you don't want to double dither if you can avoid it, so it is better to set madVR to the same bitdepth as the GPU (or lower).

For HDR I much prefer sending a display 8 bit RGB over 10 bit 4:2:2, madVR will not change the GPU output at all.

[tp4tissue]

madVR's dithering is better than the GPU's, and you don't want to double dither if you can avoid it, so it is better to set madVR to the same bitdepth as the GPU (or lower).

For HDR I much prefer sending a display 8 bit RGB over 10 bit 4:2:2, madVR will not change the GPU output at all.

Does that mean if I have 4:2:2 12bit selected in Nvidia control panel,

Madvr will be sending 4:2:2 12bit ?


I'm confused as to how this works, because it's suppose to output -Limited- range, but the desktop looks exactly the same as full range.

And in madvr control panel, if i set the display to limited range, it's actually too bright.

So, How does it all line up. ?

[huhn]

this is the 3rd time i tell you that madVR will always output RGB to the GPU driver with either 10 or 8 bit no matter what if the GPU can send it is nothing for madVR to be bothered with.

the GPU driver always expects the input RGB to be full range.
and sending YCbCr is always limited range only RGB has too ranges.

so yes the GPU driver is doing the conversation not madVR.

[tp4tissue]

this is the 3rd time i tell you that madVR will always output RGB to the GPU driver with either 10 or 8 bit no matter what if the GPU can send it is nothing for madVR to be bothered with.

the GPU driver always expects the input RGB to be full range.
and sending YCbCr is always limited range only RGB has too ranges.

so yes the GPU driver is doing the conversation not madVR.


Huhn, I still do not understand this entire chain clearly.

Is there a diagram for where the conversions happen end to end. :scared:


Not so much that I don't understand what you're saying, but I can't seem to confirm the difference in end result visually.


For example, if I set the NvidiaCP to RGB 8bit, FULL range, When NVHDR kicks in, my Display only tells me it's HDR10.. Madvr top line says RGB 8bit

But how could RGB be 10 bit if it's set to 8.


How do I know for sure, what the gpu is outputing.

[madjock]

Huhn, I still do not understand this entire chain clearly.


https://forum.kodi.tv/showthread.php?tid=259188

Lots of info here.

[Asmodian]

Is there a diagram for where the conversions happen end to end. :scared:

There are three places where conversions can happen before the display and they do not interact. The display can be doing conversions internally too, but we cannot control these.

1) LAV video can do a conversion if the render does not support the source format. When using madVR this should never be needed and it is not ideal.
2) madVR will output RGB at whatever bitdepth is set.
3) The GPU driver will do a conversion to whatever you set in its control panel.

madVR will never change what it outputs based on what the GPU is set to and the GPU will never change its output based on what madVR sends it. All these steps simply convert any input to their configured output format without a lot of smarts involved. For example the GPU is happy to convert madVR's 8 bit output to 10 bit. You know for sure what the GPU is outputting based on what it is set to in its drivers.

Also, I can send my TV 8 bit RGB HDR and it will say HDR10, that is just a brand label for the metadata standard, it works with 8 bit RGB too. :)

[madjock]

They are from a free HDR10 test pattern set. The OSD should confirm. Black clipping is not really HDR because the peak of the one mentioned is less than 1 nit, but that is intended.

Yes but there are no numbers on this ?

I did try the HDR set you mention, but I could not get the clipping ones to work, or rather get my TV to adjust once in HDR mode, unsure if this was me, or that adjusting for black crush is broken for me on my TV as I read is common in HDR mode.

[HillieSan]

Thanks huhn, I think this helped a bit but didn't fixed it. Anyway I think I found the problem. I noticed that this only happened with SVP and very high bitrate videos (and with VR), so I downloaded a high bitrate 4k 60fps video and got framedrops with both CUVID an DXVA2. Then I blocked ffdshow raw and this fixed it for both modes. So the problem is to have ffdshow raw in the chain, but obviously it's needed for SVP. Also I noticed that you can select D3D11 but it doesn't work on Windows 7, I'll have to check what I'm missing.

@nevcairiel

Yes, I discovered this recently, but anyway I'm currently not using madVR's HDR to SDR because SVP drops the needed metadata. I managed to find some alternatives however, still using both SVP and madVR, but without dynamic tonemapping (though I have some ideas also for this). I wouldn't say that normal tonemapping is much worse than dynamic, not so long ago it was the only thing available and everybody was more than happy with it. Anyway I'm not very sure how madVR is doing dynamic tonemapping with BT.2390, because only changing Mastering display white level gives me different results, and I don't see other configurable parameters for this.

Stop using SVP and set the right frequency of your display and your graphics card. This is my experience and it works well.

[el Filou]

For whatever reason with Adaptive setting it often dropped the performance state and quickly raised it back (monitored with nvidiainspector), hence it produced bunch of dropped frames. Why? That's a good question, because it worked before it with the same setting. :)I can say that's not the experience I have; on my system Adaptive always sets the clocks sufficiently high that the GPU load never reaches too high and frames are dropped, and it keeps them stable.
I see you are using 385.28 so I presume you have specific reasons for that, but maybe another driver version would help with this?
Also, it may be an option to use the High Performance power setting if you have a dedicated HTPC that doesn't see long idle times.