madVR - high quality video renderer (GPU assisted)

[huhn]

Originally Posted by madshi

I think there should always be a (small) performance benefit. Scaling factor should not be affected. The media player might change the window size, though (if "notify media player" is activated), which *may* affect scaling factor, once again. <sigh> Complicated...

a 1080p 4/3 file with some letterbox 16/9 on a 1080p 16/9 screen is for sure not faster with black boarder detection.

and in the end the file should run without black boarder detection.
for a simple reason what if the file stops using black boarders?

that's why i would never recommend someone to use that potential extra processing power that is available when black border detection is cropping the image.

[leeperry]

Originally Posted by 6233638

I don't know if something like a 10% or 15% option might be possible, but if I'm not looking at it with a critical eye, 1D 25% looks pretty good most of the time, with the exception of very detailed/busy shots, or bad sources which already contain some degree of ringing.

Fully agreed, a lower strength than 25% would be most welcome and eventually ±1 increments would be the shiznit.

Originally Posted by madshi

I find 1D 25% already so low that it barely makes a difference for me. Do you really want/need less than that? It would not be a big problem to add, but I somewhat question if you'll see much of a detail improvement at all with such low values...

Usual story that SR@1 is already very sharp so the goal would be to benefit from SSIM without the utter sharpness that comes with 25% in order to make the picture look more detailed than CR AR LL and yet not that sharp. I would myself either need softer SR strengths and/or softer SSIM.

[omarank]

Originally Posted by madshi

When releasing v0.90.11 about a week ago I asked you guys for feedback about the following things:

1) Agressive downscaling anti-ringing filter: Is it working well for you, or do you see any problems?
2) New sharpening anti-ringing filter: Is it working well for you, or do you see any problems?
3) crispen edges: Do you prefer the linear light option to be enabled or disabled?

1) & 2) I have only tested with movie content and didn’t see any problem with the modified anti-ringing filters.

3) I prefer LL enabled with crispen edges. Previously too, I preferred finesharp with LL enabled before the LL option was dropped.

Originally Posted by 6233638

I've just spent a good few hours doing testing with downsampling

Have you tested Jinc downscaling? I would be interested to know your observations about Catmull Rom vs Jinc vs SSIM downscaling.

[Sparktank]

Originally Posted by madshi

Thanks guys. Did you have a chance to compare Crispen edges with Linear Light and/or AR turned on vs off? Did you like LL on better? Or was turning on LL more of a "random" choice?

For me it was more of a random choice.
I'll have to check with diffrent types of movies to see if I can notice a difference with LL on or off.

So far, I don't see a difference if it's on or off.

Regardless of LL, madVR enchancements are making the horrid bluray transfer of Face/Off (1997) look better than a VHS upscale.

Once I get coffee going, I'll run some quick tests between the two to see if I see any real world differences.

[ang3l]

hi madshi

I'm experiencing an annoying bug with the latest version of madVR (MPC-HC.1.7.10.101.x86.VS2015 + MADV 0.90.12)

to reproduce the problem starting with the defaults settings (DX9 fullscreen exclusive)

you have to set image double and SSIM1D + antiringing as downscaler

then closing mpc with ctrl+c the program remains frozen and then I can just reboot my system

I tried on two pc with nvidia and amd and I have the same result so I think that system configuration isn't so important

it not happens with dx11 fullscreen exclusive, using SSIM1D without AR, or using any others downscaler with or without AR

[6233638]

Originally Posted by madshi

Wonderful, very helpful feedback! And of course I very much like your findings! So you did not find any aliasing caused by the new AR?

Well all of my testing was done with Blu-rays, and with that, things seemed to be ever so slightly sharper than they were before, but not necessarily more aliased.
When looking at very detailed/busy images, I'd almost say that you're best to use SoftCubic if your priority is to eliminate aliasing because it basically acts as a low-pass filter for the video. I've had a quick look over the sharpening tools and it seems that you can then sharpen this result to bring back some of the detail without introducing aliasing.
However, with a focus on sharpening tools for upscaling and SSIM for downscaling to make the image look very "detailed", I'm guessing that this is not the look most people want. I know that I was very unhappy with how strongly low-pass filtered and then oversharpened so many DVD releases were for example - though you can avoid the over-sharpening with madVR.
If you want that pixel-level detail which looks good in a static image, you're going to see aliasing when it moves in a video. You can't really get around that fact.

Here's an example from Oblivion which I think shows this off reasonably well. (note: my testing was at 50% scale)
Looking at the top-down shot at the chapter marker, if you compare SSIM at 25% to something like SoftCubic80+Sharpen Edges 0.5+Enhance Detail 1.5, there is dramatically lower aliasing in motion with the latter.
Of course the image is also lacking that really sharp and detailed look because the high frequency detail has been filtered out.
To be clear, those settings are not a recommendation for anyone, just a quick example of what I mean about the balance between aliasing/detail and the need for low-pass filtering (or similar) if you want to eliminate aliasing.
I'm not saying that there's anything wrong with SSIM or preferring that image though - it looks good for what it's trying to do, and for the type of content that I actually watch downscaled, it's very possible that I'll switch from Catmull-Rom over to SSIM.

I've had a look at the anti-ringing filter using proper test patterns now (such as a high-res star chart) and I'd say that anti-ringing looks great for the cubic scalers.
It maybe adds a tiny bit of roughness at the very high frequency details, and Bicubic >60 could look a bit better, but I'd also not be recommending that people use the higher levels of Bicubic anyway.
For Lanczos/Spline/Jinc, it could stand to be a little bit more aggressive with linear light enabled. There are still some remnants of the dark ringing that you get with LL downscaling at some angles - though I never found it to be a problem with video.
With SSIM, anti-ringing is actually adding aliasing at all levels so that probably needs looked at. I thought I had seen this to some degree with some video clips, but assumed it was just because SSIM was so sharp - which was why I wondered if something like 10% SSIM might be an option.

Originally Posted by madshi

What is your opinion about LL scaling now? I'm considering moving both LL and AR options for downscaling into the "trade quality for performance" section. Which would mean I'd consider highest quality to be achieved with both LL and AR turned on. Would you agree with that assessment?

Absolutely. If you can make the tweaks that I mentioned above, other than the performance hit for enabling them, I see no reason why you wouldn't want to use them now that AR+LL works well with every scaling algorithm.

Originally Posted by madshi

Do you like sharpening algos? If not, I don't want to bother you with tests. But if you do, I'd like your feedback about Crispen edges linear light on vs off, and whether the sharpening AR filter works well for you or not.

Sorry, I've not really been using sharpening at all.
I did some tests when it was first introduced but my issue with sharpening has always been that it's not something I would ever want to apply globally.
It can help a bad source if you're tweaking it to be specific for that source, but always hurts a good source in my opinion.
And generally I find that a disc is either well mastered, or if it's not then it probably has a ton of sharpening already encoded in the video. There aren't many films in my library which just have a soft image.

If it was like debanding where I could hit a key to either toggle it on/off or perhaps switch between say three separate sharpening profiles I might be more inclined to use it.
That's not a feature request though, as I just don't think that sharpening is something I'd use much even if that were an option.

Originally Posted by madshi

Ouch, I find 1D 25% already so low that it barely makes a difference for me. Do you really want/need less than that? It would not be a big problem to add, but I somewhat question if you'll see much of a detail improvement at all with such low values...

If the aliasing with AR can be addressed, that may actually solve the problem instead of it needing to be less sharp - though if it's a small change, I'd still like to see how that looks.
At the same time, it's possible that SSIM just isn't what I'm looking for, so I may just be better off using another scaling algorithm instead of a reduced-strength SSIM.

Originally Posted by omarank

Have you tested Jinc downscaling? I would be interested to know your observations about Catmull Rom vs Jinc vs SSIM downscaling.

I didn't see much benefit to it compared against the other scalers to be honest. The difference may have been larger before the new anti-ringing filter was implemented?
I'd like to see examples if anyone has found cases where Jinc has made a difference though.

[Cudo]

Originally Posted by madshi

When releasing v0.90.11 about a week ago I asked you guys for feedback about the following things:

1) Agressive downscaling anti-ringing filter: Is it working well for you, or do you see any problems?
2) New sharpening anti-ringing filter: Is it working well for you, or do you see any problems?
3) crispen edges: Do you prefer the linear light option to be enabled or disabled?

So far from what I remember I've received zero feedback about any of that. Of course you don't *have* to provide feedback, but if you don't, then I get the impression that I develop too fast for you guys to keep up. Which means I'd have to slow down development accordingly to allow you to catch up. Your choice...

Seeing as my computer didn't blow up, probably everything is fine.

[Uoppi]

Originally Posted by madshi

There are variations of the profile vars available if you want profiles to be selected according to the original resolution (before cropping). E.g. you can use "uncroppedSrcWidth" instead of "srcWidth".

Cool! Sounds like "uncroppedSrcWidth" would do exactly what I need so I'll start using that instead.

Again I'm not sure what you mean. E.g. are you just talking about some numbers in the OSD, or are you talking about the actual refresh rate the GPU is outputting?

The numbers in the OSD as well as actual refresh rate the TV switched to. I've just never encountered a situation before where the refresh rate is doubled to 60 Hz (or 50 Hz for that matter) when it "shouldn't". But I guess that's a non-issue if frames simply get duplicated (30 Hz -> 60 Hz). Maybe it had something to do with a badly done variable frame rate encode with fluctuating fps (?). The VFR videos shot with my Galaxy Note 4 never produce this issue though, even if the fps is sometimes all over the place.

[madshi]

Originally Posted by Sunset1982

Just did some testing with the crop black bars feature. Tested a 720p File which is upscaled to 2160p by my GTX960. Activating the crop black bars option brings me a better performance by 12ms rendering time.

Originally Posted by huhn

a 1080p 4/3 file with some letterbox 16/9 on a 1080p 16/9 screen is for sure not faster with black boarder detection.

Black bar detection itself does not improve performance. But black bar detection with the additional option "crop black bars" activated should improve performance a bit - if there are black bars of noticeable size. Of course if you use cheap algos everywhere you might not notice. But if you do heavy lifting like NNEDI3 image doubling, you should notice a difference.

Originally Posted by huhn

and in the end the file should run without black boarder detection.
for a simple reason what if the file stops using black boarders?

that's why i would never recommend someone to use that potential extra processing power that is available when black border detection is cropping the image.

99% of all Cinemascope Blu-Rays stay Cinemascope throughout the whole movie.

Anyway, whether it makes sense to setup a special profile to make use of the saved power is a question I don't want to answer. If all else fails, at least the GPU will consume less power, which is better for the environment, for your power bill and for fan noise.

Originally Posted by omarank

1) & 2) I have only tested with movie content and didn’t see any problem with the modified anti-ringing filters.

3) I prefer LL enabled with crispen edges. Previously too, I preferred finesharp with LL enabled before the LL option was dropped.

K, thanks.

Originally Posted by Sparktank

For me it was more of a random choice.
I'll have to check with diffrent types of movies to see if I can notice a difference with LL on or off.

So far, I don't see a difference if it's on or off.

If you can see a difference I'd like to hear which you prefer. Thanks.

Originally Posted by ang3l

I'm experiencing an annoying bug with the latest version of madVR (MPC-HC.1.7.10.101.x86.VS2015 + MADV 0.90.12)

to reproduce the problem starting with the defaults settings (DX9 fullscreen exclusive)

you have to set image double and SSIM1D + antiringing as downscaler

then closing mpc with ctrl+c the program remains frozen and then I can just reboot my system

FWIW, simply pressing Alt+F4 would have closed MPC-HC in this situation (at least it did for me). But thanks for letting me know, it was a resource leak and should be fixed in the next build.

Originally Posted by 6233638

Well all of my testing was done with Blu-rays, and with that, things seemed to be ever so slightly sharper than they were before, but not necessarily more aliased.

Ok, I like to hear that!

Originally Posted by 6233638

When looking at very detailed/busy images, I'd almost say that you're best to use SoftCubic if your priority is to eliminate aliasing because it basically acts as a low-pass filter for the video. I've had a quick look over the sharpening tools and it seems that you can then sharpen this result to bring back some of the detail without introducing aliasing.

Fair enough. But my main worry was that the AR algo would add introducing by itself, which would be quite bad. If that's not the case then at least I know that the AR algo itself works fine. Which algo combination the user might prefer for best sharpness vs lowest aliasing is something I'm worried about at this point. Just want to make sure that the more aggressive AR algo isn't harmful in any way.

Originally Posted by 6233638

I've had a look at the anti-ringing filter using proper test patterns now (such as a high-res star chart) and I'd say that anti-ringing looks great for the cubic scalers.
It maybe adds a tiny bit of roughness at the very high frequency details, and Bicubic >60 could look a bit better, but I'd also not be recommending that people use the higher levels of Bicubic anyway.
For Lanczos/Spline/Jinc, it could stand to be a little bit more aggressive with linear light enabled. There are still some remnants of the dark ringing that you get with LL downscaling at some angles - though I never found it to be a problem with video.

The algo is currently very "logical" in its aggressiveness. I don't want to make it any more aggressive (for Lanczos/Spline/Jinc) because that would most probably introduce new problems. It's as aggressive as I feel reasonable safe to do.

Originally Posted by 6233638

With SSIM, anti-ringing is actually adding aliasing at all levels so that probably needs looked at.

Oh, that's exactly the kind of feedback I need!

I've been able to reproduce that somewhat with your star test pattern. Have added that to my to do list for maybe next weekend.

Originally Posted by Cudo

Seeing as my computer didn't blow up, probably everything is fine.

My question was not about performance but about image quality.

Originally Posted by Uoppi

The numbers in the OSD as well as actual refresh rate the TV switched to. I've just never encountered a situation before where the refresh rate is doubled to 60 Hz (or 50 Hz for that matter) when it "shouldn't".

This all heavily depends on which frame rate the decoder/splitter reports to madVR. Usually this is out of my control. madVR makes its refresh rate decisions based on that information.

[madshi]

madVR v0.90.13 released

http://madshi.net/madVR.zip

* added support for finding the right 3D depth for the active subtitle track
* added "use alternative glitch handling mode" option for D3D11 presentation
* added a workaround for glitch problems with shaky 3D GPU drivers
* fixed: downscaling sometimes crashed, when not using SSIM
* fixed: Intel GPU always used DXVA scaling when DXVA deint/decode was active
* fixed: SSIM AR eventually left some resources open

3D users please make sure you use the very latest LAV nightly build, once more.

[Georgel]

This should be left on the first page as all Optimus laptops will have problems with mpc-hc + madVR.

The solution is quite simple.

First, you have to go to where mpc-hc is installed for example C:\Program Files (x86)\MPC-HC and rename your mpc-hc.exe to mpc-hc1.exe.

After this you have to go to Nvidia control panel -> Manage 3D settings -> Add, then browse to C:\Program Files (x86)\MPC-HC\mpc-hc1.exe and set to to use High performance Nvidia Processor.

I will add photos in here.

https://www.dropbox.com/s/khfwwz86cwf3k6v/12.png?dl=0

https://www.dropbox.com/s/7rrf8thbdjnwx9k/10.png?dl=0

https://www.dropbox.com/s/oog8l8150u47uti/11.png?dl=0

This should do the trick.

If power is anywhere under maximum performance, I experience -drops in performance. I am running an Acer VN7-791G with GTX860M and the latest Nvidia driver, 361.91.

[hu1kamania]

Originally Posted by madshi

Thanks guys. Did you have a chance to compare Crispen edges with Linear Light and/or AR turned on vs off? Did you like LL on better? Or was turning on LL more of a "random" choice?

In my limited testing, I couldn't see much of a difference in Cripen Edges with Linear Light on and off. So I guess keeping it checked was more of a random choice.

With the AR filter from Refinements + Crispen Edges, I was able to detect a difference, and it was an improvement. Very subtle in my test though.

Edit: I'll have to take back what I said about Crispen edges and linear light. Depending on SR strength, I do see a difference and prefer it to be on with SuperSampling up until I hit SR4.

[mindz]

This has probably been asked before, but i couldnt find the answer.

When choosing super-xbr for chroma, why dont the sliders move when applying AR or choosing different sharpness? Do the changes have effect or dont they?

[mogli]

Before v0.90.13 madVR used DXVA image scaling while deinterlacing, now it uses the user chosen image scaler. Is this intentional or caused be the next to last fix (NVIDIA here, not Intel)?

[GCRaistlin]

Originally Posted by madshi

99% of all Cinemascope Blu-Rays stay Cinemascope throughout the whole movie.

In my experience, there was at least one movie ("Gone Girl", 2014) that is 2,35:1 but switching for a couple of secs to 16:9. If set to crop black borders, madVR would simply cut such "non-standard" frames to 2,35:1 (unfortunately I don't have this movie anymore so I can't check myself)?

[6233638]

Originally Posted by madshi

The algo is currently very "logical" in its aggressiveness. I don't want to make it any more aggressive (for Lanczos/Spline/Jinc) because that would most probably introduce new problems. It's as aggressive as I feel reasonable safe to do.

That's fair. Though most of my testing was focused on cubic/SSIM downscaling, I did not see it as being a problem with the Blu-rays that I tested, so it's probably more of a theoretical issue.
As I said before, they're massively improved now - to the point that people should be able to use linear light downscaling with any of them based on their preference.

[Warner306]

I've only used crispen edges. The new anti-ringing may be a small improvement with a slight decrease in performance. But I wasn't overly bothered by crispen edges before. I haven't checked sharpen edges.

I don't see any improvement using linear light.

[Warner306]

I was browing this: Dolby Vision Whitepaper.

I am wondering what the CPU/GPU cost is of decoding the enhancement/HDR layer (HDR10)? My system can't handle high-bitrate HEVC, so it is difficult to tell what is going on.

[Cudo]

Originally Posted by Georgel

This should be left on the first page as all Optimus laptops will have problems with mpc-hc + madVR.

The solution is quite simple.

First, you have to go to where mpc-hc is installed for example C:\Program Files (x86)\MPC-HC and rename your mpc-hc.exe to mpc-hc1.exe.

After this you have to go to Nvidia control panel -> Manage 3D settings -> Add, then browse to C:\Program Files (x86)\MPC-HC\mpc-hc1.exe and set to to use High performance Nvidia Processor.

I will add photos in here.

https://www.dropbox.com/s/khfwwz86cwf3k6v/12.png?dl=0

https://www.dropbox.com/s/7rrf8thbdjnwx9k/10.png?dl=0

https://www.dropbox.com/s/oog8l8150u47uti/11.png?dl=0

This should do the trick.

If power is anywhere under maximum performance, I experience -drops in performance. I am running an Acer VN7-791G with GTX860M and the latest Nvidia driver, 361.91.

Here's a better way to get to what you have done.

Download Nvidia Inspector:
http://www.techspot.com/downloads/50...inspector.html
Click the Profile Settings button:

In the Profiles box, select Media Player Classic
Scroll down to "Common" (Blue text with a horizontal line)
Under Common you will see "Power management mode"
To the right of that you will see "Adaptive"
Click in that box and select "Prefer maximum performance"
Scroll down to "Other" (Blue text with a horizontal line)
Under Other you will see "Enable application for Optimus"
To the right of that you will see a bunch of capital letters
Click in that box and select SHIM_RENDERING_MODE_ENABLE
Click Apply Changes

[madshi]

Originally Posted by Georgel

First, you have to go to where mpc-hc is installed for example C:\Program Files (x86)\MPC-HC and rename your mpc-hc.exe to mpc-hc1.exe.

After this you have to go to Nvidia control panel -> Manage 3D settings -> Add, then browse to C:\Program Files (x86)\MPC-HC\mpc-hc1.exe and set to to use High performance Nvidia Processor.

If power is anywhere under maximum performance, I experience -drops in performance. I am running an Acer VN7-791G with GTX860M and the latest Nvidia driver, 361.91.

Originally Posted by Cudo

Here's a better way to get to what you have done.

Download Nvidia Inspector:
http://www.techspot.com/downloads/50...inspector.html
Click the Profile Settings button: [...]

In the Profiles box, select Media Player Classic
Scroll down to "Common" (Blue text with a horizontal line)
Under Common you will see "Power management mode"
To the right of that you will see "Adaptive"
Click in that box and select "Prefer maximum performance"
Scroll down to "Other" (Blue text with a horizontal line)
Under Other you will see "Enable application for Optimus"
To the right of that you will see a bunch of capital letters
Click in that box and select SHIM_RENDERING_MODE_ENABLE
Click Apply Changes

Ok, thanks!

Originally Posted by hu1kamania

In my limited testing, I couldn't see much of a difference in Cripen Edges with Linear Light on and off. So I guess keeping it checked was more of a random choice.

With the AR filter from Refinements + Crispen Edges, I was able to detect a difference, and it was an improvement. Very subtle in my test though.

Edit: I'll have to take back what I said about Crispen edges and linear light. Depending on SR strength, I do see a difference and prefer it to be on with SuperSampling up until I hit SR4.

Originally Posted by Warner306

I've only used crispen edges. The new anti-ringing may be a small improvement with a slight decrease in performance. But I wasn't overly bothered by crispen edges before. I haven't checked sharpen edges.

I don't see any improvement using linear light.

Ok, thanks for your feedback!

Originally Posted by 6233638

That's fair. Though most of my testing was focused on cubic/SSIM downscaling, I did not see it as being a problem with the Blu-rays that I tested, so it's probably more of a theoretical issue.
As I said before, they're massively improved now - to the point that people should be able to use linear light downscaling with any of them based on their preference.

Good to know, thanks!

Originally Posted by mindz

This has probably been asked before, but i couldnt find the answer.

When choosing super-xbr for chroma, why dont the sliders move when applying AR or choosing different sharpness? Do the changes have effect or dont they?

They do, but all the settings are changing all the time, and results are different in different test patterns, so I don't always get around updating the sliders, and they're a VERY rough estimate, anyway. Don't put too much weight into them.

Originally Posted by mogli

Before v0.90.13 madVR used DXVA image scaling while deinterlacing, now it uses the user chosen image scaler. Is this intentional or caused be the next to last fix (NVIDIA here, not Intel)?

The changelog only mentions Intel, but it also applies to NVidia, and yes, it's an intentional change. Actually it was a bug that the chosen image scaler was *not* used. If you like it to be used, you can always manually select it. I'll change the changelog.

Originally Posted by GCRaistlin

In my experience, there was at least one movie ("Gone Girl", 2014) that is 2,35:1 but switching for a couple of secs to 16:9.

There are a couple of IMAX movies, which switch between Cinemascope and 16:9, but those are not too common.

Originally Posted by GCRaistlin

If set to crop black borders, madVR would simply cut such "non-standard" frames to 2,35:1 (unfortunately I don't have this movie anymore so I can't check myself)?

The exact behaviour depends on the "zoom control" settings. You can define exactly what you want madVR to do in such a situation. There are dedicated options for exactly this sort of thing.

Originally Posted by Warner306

I was browing this: Dolby Vision Whitepaper.

I am wondering what the CPU/GPU cost is of decoding the enhancement/HDR layer (HDR10)?

How should anybody know? There's no real technical information available. Whitepapers are mostly useless.