madVR - high quality video renderer (GPU assisted)

[eero94]

Any idea what causes my stuttering with x264 10bit content, same clip with x265 10bit plays fine

Render stats:

x264 10bit www.imgur.com/LMdpoQl.png

x265 10bit www.imgur.com/lrlqcBo.png

[madshi]

Originally Posted by leeperry

So at the end of the day, when ppl claimed that 10bit off mVR looked better than 8 for all we know they were gazing at dithering in their graphic card drivers or TV? There's no hard-proof test pattern that can validate how things are being sent?

As I said:

The only reliable way I know is if your receiver or TV report the incoming bitdepth.

Originally Posted by eero94

Any idea what causes my stuttering with x264 10bit content, same clip with x265 10bit plays fine

Render stats:

x264 10bit www.imgur.com/LMdpoQl.png

x265 10bit www.imgur.com/lrlqcBo.png

Could the difference be native DXVA decoding vs software decoding? Not sure why software decoding would result in dramatically slower rendering times, though. That seems very weird!

[huhn]

Originally Posted by madshi

Is that with "chroma quality" set to "normal"? I get a weird colored image that way. But with every other chroma quality setting the problem doesn't seem to occur. Can you confirm that finding?

yes only happens with chroma normal

Then some other part of the configuration must have changed because there should be no difference in NGU AA speed compared to NGU pix art 3.

One thing coming to mind is NGU pixart 3 "high" preset didn't exist. It actually used "medium". Could that be an explanation?

i'm talking about very high. high and lower are pretty much "free" for 1080p23 -> 2160p23.

Originally Posted by leeperry

So at the end of the day, when ppl claimed that 10bit off mVR looked better than 8 for all we know they were gazing at dithering in their graphic card drivers or TV? There's no hard-proof test pattern that can validate how things are being sent? And worst case scenario if it's still entirely subjective, I guess A/B'ing as quickly as possible a 10bit gray ramp might help? Considering how much of a crap shot a potential upgrade to RX480 & NGU + HDMI2.0b UHD TV would be for me, right now my only motive to upgrade to W8.1 would be 10bit 1080p60.

the selected bit deep in the AMD driver is the bit deep that is send to the display and the madVR setting doesn't matter for this. this can be confirmed on AMD with some advanced tests.
and the default is 10 bit if possible.
if you are using a 1080p >TV< it is currently getting 10 bit.
this is different with nvidia the option in the nvidia control panel doesn't really mean a lot. if the driver get's 10 bit from any software like madVR it is output 12 bit and 8 bit with 8 bit input. but i haven't tested this with nvidia in a long time.


all this works on with windows 7...
and be aware that only the some flagship screens are using true 10 bit panels.

[madshi]

Originally Posted by huhn

i'm talking about very high. high and lower are pretty much "free" for 1080p23 -> 2160p23.

It should have the same speed. I've revisited the "let madVR decide" decisions, though. Could that explain the speed difference?

How big is the difference in render times?

Can you see a difference in the listed algorithms in the OSD?

[Neo-XP]

Originally Posted by madshi

Your comparison shows a very very small difference, and you've zoomed the image up to like 1000% to make it more easily visible. I don't find that very convincing, to be honest. How does the comparison look like when upscaling just 2x instead of 10x?

The way the chroma doubling setting is designed atm limits the number of algos I can offer. If I replace Bicubic60 AR with Lanczos3 AR, performance *will* go down for some users with older GPUs. It might not be a big speed difference, but it will be there, and it might be the difference between smooth and stuttering playback. So I don't like the idea to replace Bicubic60 with Lanczos3 at all.

Of course it's always possible to allow both options, but doing so in an intuitive way is difficult. E.g. currently "normal" quality is Bicubic60 AR. So how would I name Lanczos3 AR? Would that be "normal + 0.01% quality"? I don't think it makes sense for the majority of users.

Since they are very similar, I agree that it doesn't make much sense to propose both from this list box in a different quality setting. But for users who want the best quality without NGU, you could propose a "same as upscaling algo" choice. This way "Lanczos3 AR" and "Jinc AR" would be available for chroma doubling. New users can always choose "let madVR decide" anyway.

Originally Posted by madshi

Looking at my source code, the NGU algo was tweaked between v0.91.1 and v0.91.6. The quality presets "medium" and "very high" seem to be identical, but the quality preset "high" was modified slightly. I don't remember the exact reason, probably I found some artifacts somewhere and saw the need to improve "high". Sometimes improving quality in one aspect costs a little bit quality somewhere else. Looking at your image comparison, I have a hard time seeing any sharpness difference. Ok, the reflection in the right eye is a tiny bit sharper with v0.91.1, but that's really a *very* small difference. Probably when comparing with a different image, it might look better with v0.91.6.

I couldn't find any case where NGU Sharp (high) performs better than NGU-med in v0.91.1.

Here is another example :

original : http://www.mediafire.com/file/z27qsd.../original2.png
v0.91.1 : http://www.mediafire.com/file/1h6f9b17dvkt0v2/old2.png
v0.91.6 : http://www.mediafire.com/file/jfvxw096d0fk60f/new2.png

You can see in DiCaprio's eyes and all around them the gain in details and sharpness.
You can also see it on the green trees behind him (at the right side of his face in the image).

Originally Posted by huhn

i kind of miss "use image downscaling" under image upscaling for the downscaling algorithm.

I also miss this. Even if I don't need it anymore, others may want it back.

[huhn]

Originally Posted by madshi

It should have the same speed. I've revisited the "let madVR decide" decisions, though. Could that explain the speed difference?

no i checked this.

How big is the difference in render times?

about 5 ms
i'm used to about 36 ms with this version it is currently 40-53 ms

Can you see a difference in the listed algorithms in the OSD?

no. the name is now Anti-Alias but that's it.

i will double check it with the old version.
i found a file where it still works for some time so something is odd here and the resolution is the same.

[kasper93]

@madshi: Downscaling after doubling settings has one item offset. If you select Lanczos3 AR, Bicubic150 AR LL will be used. And this is valid for each item. Actual selection is -1

[aufkrawall]

Originally Posted by madshi

Great to hear that! I'm a bit surprised, though. I thought that the deringing filter made problems with both NGU and NNEDI3. But now you say it works fine with NGU AA? It's surprising because NGU AA is supposed to behave very similar to NNEDI3. So I'm not sure why NGU AA would work better with deringing than NNEDI3 does. But if that's really the case, of course I won't complain!

In my opinion NNEDI3 works fine with your deringing filter, that's why I once voted to keep it in.
Will do some further NGU AA tests tomorrow, but I'm very positive so far.

Would it be possible to take a look at edge smoothing for NGU VH 4x at some time? A stronger smoothing effect would be helpful for low-res content that isn't extremely aliased and thus doesn't require NGU AA. The current smoothing solution for NGU VH 4x has a too weak effect.

[eero94]

Originally Posted by madshi

As I said:


Could the difference be native DXVA decoding vs software decoding? Not sure why software decoding would result in dramatically slower rendering times, though. That seems very weird!

Yeah, it seems pretty weird and stuttering only starts after I put fullscreen, fully maximized windowed with all the scaling activated plays fine too.

Edit: Disabling D3D11 solved this, something wrong with s/w decoding + D3D11 fullscreen for me!

[Backflash]

Originally Posted by madshi

It should be identical. However, if you downscale after NGU Sharp, the final result should be sharper now, because the downscaling algorithm uses a less aggressive anti-ringing filter now.

Nope, direct double NGU Sharp - High to 1440 from 720. Looks sharper than before, didn't change anything, I think. And not going to touch anything, I like how it looks.

[madshi]

Originally Posted by Neo-XP

Since they are very similar, it doesn't make much sense to propose both from this list box in a different quality setting.

Agreed.

Originally Posted by Neo-XP

But for users who want the best quality without NGU, you could propose a "same as upscaling algo" choice.

I don't understand how that would work. If you select one of the doublers in the "image upscaling" settings dialog, you can't select Lanczos 3, anymore. It's either Lanczos 3 or e.g. NGU in the current settings dialog. So how does "same as upscaling algo" make any sense? It would then be NGU again for chroma doubling. Or am I missing something?

I couldn't find any case where NGU Sharp (high) performs better than NGU-med in v0.91.1.

It doesn't have to be sharpness, it can be ringing or aliasing artifacts in certain images. I'm also not only testing with Blu-Ray film, but also with other sources, e.g. anime, pixart, photos etc.

Anyway, it doesn't matter too much, NGU Soft/Standard/Sharp will all change again soon.

Originally Posted by huhn

no i checked this.

about 5 ms
i'm used to about 36 ms with this version it is currently 40-53 ms

no. the name is now Anti-Alias but that's it.

i will double check it with the old version.
i found a file where it still works for some time so something is odd here and the resolution is the same.

I've no explanation right now.

Originally Posted by kasper93

@madshi: Downscaling after doubling settings has one item offset. If you select Lanczos3 AR, Bicubic150 AR LL will be used. And this is valid for each item. Actual selection is -1

Oh, thanks. That's probably the missing "use the downscaling settings" item, which got accidently lost.

Originally Posted by aufkrawall

In my opinion NNEDI3 works fine with your deringing filter, that's why I once voted to keep it in.

Ah, ok! I do remember some users complaining about deringing + NNEDI3 making some problems, though. Well, anyway...

Originally Posted by aufkrawall

Would it be possible to take a look at edge smoothing for NGU VH 4x at some time? A stronger smoothing effect would be helpful for low-res content that isn't extremely aliased and thus doesn't require NGU AA. The current smoothing solution for NGU VH 4x has a too weak effect.

You mean you want regular NGU to have a strong anti-aliasing effect? I'm not sure if I can implement that, to be honest.

Originally Posted by eero94

Yeah, it seems pretty weird and stuttering only starts after I put fullscreen, fully maximized windowed with all the scaling activated plays fine too.

Edit: Disabling D3D11 solved this, something wrong with s/w decoding + D3D11 fullscreen for me!

Strange. Could be a GPU driver thing.

[aufkrawall]

Originally Posted by madshi

You mean you want regular NGU to have a strong anti-aliasing effect? I'm not sure if I can implement that, to be honest.

Nope, it would just be nice if the edge smoothing strength for direct quadrupling NGU would be as strong as for 2x doubling.

[Georgel]

Hi @madshi!

I like the new upscaling refinement page!

Not entirely sure if NGU AA low is equal with NNEDI 16 neurons quality wise, and I'm not entirely sure if it looks exactly the same so it might still be okay to keep it around if it's not too much effort. I mean, NNEDI looks good by itself, but NGU looks pretty good as well.

This new version doesn't have PixArt installed within, no?

[pose]

Originally Posted by madshi

That's probably the missing "use the downscaling settings" item, which got accidently lost.

Can we expect hotfix today or tomorrow? I wont be upgrading for now then.

[Warner306]

I think the new image upscaling/doubling settings are a mess and difficult to interpret if you are a noob...way too many drop-down menus.

What is the difference between upscaling and doubling? Why am I able to double and quadruple the same image?

[Telion]

Originally Posted by pose

Can we expect hotfix today or tomorrow? I wont be upgrading for now then.

And please add a promised hotkey for the settings dialog, madshi. I hope it won't take much efforts.

[nevcairiel]

Originally Posted by Warner306

I think the new image upscaling/doubling settings are a mess and difficult to interpret if you are a noob...way too many drop-down menus.

What is the difference between upscaling and doubling? Why am I able to double and quadruple the same image?

No layout in the world can suddenly make noobs not be noobs.
If they don't even understand the basic concepts being offered there (ie. what doubling means), it doesn't matter how the options look. The "let madVR decide" modes are already much easier then before.

[jkauff]

Originally Posted by madshi

Hope you guys are happy with the upscaling settings redesign? I'm not fully happy with that there are so many options now, but I suppose defaulting them all to "let madVR decide" should help newbies, so I hope it will be ok for a while.

madshi, I've been a professional UI designer for 25 years. If you ever want to bounce ideas off me, or get an opinion on a design, feel free to reach out. I'd be glad to help.

[dioxholster]

I dont know what madvr settings to use. example, super-xbr ? Using amd 470 card.

Originally Posted by nevcairiel

No layout in the world can suddenly make noobs not be noobs.
If they don't even understand the basic concepts being offered there (ie. what doubling means), it doesn't matter how the options look.

I dont know why we need to choose between two extremes (win10 vs linux) when we can have both with a neat UI.

[ABDO]

wow NGU AA need more performance than NGU pix art 3. but it looks better than NNEDI3 256-taps. madshi great job here.