madVR - high quality video renderer (GPU assisted)

[madshi]

I still can't choose "Lanczos3 AR" for chroma doubling.
I've yet to see a screenshot comparison where Lanczos 3 AR is visibly better than Bicubic60 AR for chroma doubling.

"NGU-Sharp (high)" is still not as sharp as "NGU-med" in v0.91.1. Is that normal ?
Using exactly the same downscaling algorithm? Which image are you testing with? I need it in original (unscaled) size.

[tobindac]

1) Is the new upscaling settings page fine now?
It seems fine for power users. You might want to create an alternative/default/first impressions interface for very novice users (with basic options like "Sharp Upscalers (<type1>)", "Sharp upscalers (<type2>, "Soft Upscalers" etc). madVR is getting popular and people start using it en masse and they are not going to like having to read much about it.
Heck even other programs start copying it that they don't know much what they're doing *cough*mpv*cough*.

[Neo-XP]

I've yet to see a screenshot comparison where Lanczos 3 AR is visibly better than Bicubic60 AR for chroma doubling.

It is slightly sharper and the closest to NGU (NGU is way sharper, but demands too much resources). Also, Bicubic60 AR and Lanczos3 AR are identical in term of performances.
See https://forum.doom9.org/showthread.php?p=1790283#post1790283 for original image
and https://forum.doom9.org/showthread.php?p=1790272#post1790272 for a zoomed comparison.

With Cr rendered as Y (by huhn) :
Bicubic60 AR : https://abload.de/img/1dup4r.png
Lanczos3 AR : https://abload.de/img/24arvq.png

I am sure that you can see it better with some well-picked patterns...

Using exactly the same downscaling algorithm? Which image are you testing with? I need it in original (unscaled) size.

There is no downscaling algorithm used, only upscaling with NGU. All the other settings are the same.

original : http://www.mediafire.com/file/ce28j9340ad5t0h/original.png
v0.91.1 : http://www.mediafire.com/file/gv5wl45d67daawf/v0.91.1.png
v0.91.6 : http://www.mediafire.com/file/2m6n7bse99s2a5h/v0.91.6.png

Once again, you need to zoom quite a lot to see the difference, but there should not be one.

[aufkrawall]

Great job on NGU AA, to my eyes it looks better than NNEDI3.
Good thing is that it apparently can also be combined with deringing filter without nasty side effects. I think that's the way you want to watch bad quality sources, also a good fit for chroma.

[tobindac]

PS. For the interface, big tooltips after hovering over algorithms/methods/settings, that explain what each of the setting/algorithm/method is might help. It could be kinda cryptic but perhaps helpful and not distracting if it pops up only after 2seconds. A half-decent solution I had devised was to have a "help" indicator next to each option that would automatically hover a very decriptive tooltip about it which makes it more explicit that that documentation is available so the work for it doesn't go to waste.

[leeperry]

Is the new upscaling settings page fine now?

Thanks a lot for the new build but could you please bring back the "use image downscaling" option for downscaling doubling? My fav combo on many noisy videos is sxbr75+SR2 AR LL+SSIM 2D 100% LL without AR. Everything else looks great, only this option for OCD'ed video nerds and it'll be fine I guess ^^

[huhn]

i get image corruption with NGU sharp very high X2 -> NGU sharp very high X2.
or with other words when i use NGU sharp 2 times.

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

but downscaling got a lot less important for me now that i can control the supersampling again.

EDIT: for me NGU AA needs a "lot more" performance than NGU pix art 3. this is working as intended?
i can't use it in real time anymore.

[Backflash]

1) Is the new upscaling settings page fine now?
Yes. In my opinion it's very intuitive now.

Is NGU Sharp sharper than 0.91.5? Looks that way. I love that, i'd crank it up even more like one of the first versions though.
Can't comment on pixart changes, I really do not understand them.

[madshi]

It seems fine for power users. You might want to create an alternative/default/first impressions interface for very novice users
PS. For the interface, big tooltips after hovering over algorithms/methods/settings, that explain what each of the setting/algorithm/method is might help.
I've already said this many times before: There will be a proper documentation, and a redesigned settings dialog for madVR v1.0. Until then I try to keep the changes to the settings dialog as small as possible. So any dramatic changes will have to wait.

It is slightly sharper and the closest to NGU (NGU is way sharper, but demands too much resources). Also, Bicubic60 AR and Lanczos3 AR are identical in term of performances.
See https://forum.doom9.org/showthread.php?p=1790283#post1790283 for original image
and https://forum.doom9.org/showthread.php?p=1790272#post1790272 for a zoomed comparison.
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.

There is no downscaling algorithm used, only upscaling with NGU. All the other settings are the same.

original : http://www.mediafire.com/file/ce28j9340ad5t0h/original.png
v0.91.1 : http://www.mediafire.com/file/gv5wl45d67daawf/v0.91.1.png
v0.91.6 : http://www.mediafire.com/file/2m6n7bse99s2a5h/v0.91.6.png

Once again, you need to zoom quite a lot to see the difference, but there should not be one.
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.

Great job on NGU AA, to my eyes it looks better than NNEDI3.
Good thing is that it apparently can also be combined with deringing filter without nasty side effects. I think that's the way you want to watch bad quality sources, also a good fit for chroma.
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!

could you please bring back the "use image downscaling" option for downscaling doubling?
i kind of miss "use image downscaling" under image upscaling for the downscaling algorithm.
Ooops. Removing that option wasn't intentional.

i get image corruption with NGU sharp very high X2 -> NGU sharp very high X2.
or with other words when i use NGU sharp 2 times.
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?

EDIT: for me NGU AA needs a "lot more" performance than NGU pix art 3. this is working as intended?
i can't use it in real time anymore.
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?

Is NGU Sharp sharper than 0.91.5?
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.

[leeperry]

The only reliable way I know is if your receiver or TV report the incoming bitdepth. If madVR sends 10bit to the GPU driver, the GPU driver could still ouput it as dithered or undithered 8bit, 10bit, 12bit or 16bit. I've no control over that, I can't even ask the GPU which bitdepth it actually outputs.
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.

[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]

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.

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]

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.

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]

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]

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.

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/z27qsdw82ezb3n1/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).

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]

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]

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]

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]

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]

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

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.

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.

@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.

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...

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.

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]

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]

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]

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]

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]

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.

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.:thanks:

[madshi]

Nope, it would just be nice if the edge smoothing strength for direct quadrupling NGU would be as strong as for 2x doubling.
I'm not completely sure what you mean. Basically you seem to say that edge smoothing strength for direct quadrupling is less strong as it is for 2x doubling. But is it? And what exactly do you mean with "edge smoothing"? Do you mean anti-aliasing? Or something else?

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.
NGU AA low is better in some images than NNEDI-16 and worse in other images. NGU AA med should beat NNEDI-16 in most images and still be faster.

This new version doesn't have PixArt installed within, no?
What do you mean? The NGU PixArt algo was renamed to NGU Anti-Alias, as mentioned in the v0.91.6 announcement post.

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.
There are too many drop-down menues for my taste. As you might have noticed, I had tried to reduce the number of options, but many users weren't happy with my cuts, so I've brought most settings back. However, to make things easier for noobs, most settings have a "let madVR decide" setting now, and also default to that setting.

FWIW, what you're seeing now is the "expert" version of the settings dialog. When I reach v1.0, there may be a noob view with limited options. But I've not decided that yet.

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.
Agreed.

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.
Thanks. But it's not the right time for that now. Right now I'm trying to do only what's absolutely necessary, in as short time as possible.

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

I dont know what madvr settings to use. example, super-xbr ? Using amd 470 card.
If you like sharp, for good quality sources try "NGU Sharp". For bad quality sources, try "NGU Anti-Alias". It's less sharp, but more forgiving.

NGU AA need more performance than NGU pix art 3.
It's not supposed to!! You're the 2nd user who reports this, and I don't know why you have that experience. Can you try to find out more about that? NGU AA is supposed to have exactly the same speed as NGU pixart 3. Except for the "high" quality preset, which wasn't available with NGU pixart.

[madshi]

madVR v0.91.7 released

http://madshi.net/madVR.zip

* fixed: doubling twice with "normal" chroma quality resulted in weird colors
* fixed: "downscaling after doubling" dropbox was missing an item

[jmonier]

I'm trying out 91.6 replacing 91.5 on a GTX1060 6gb and trying to have the same settings as before: NGU high for luma and NGU Sharp med for chroma (was NGU med). With the same material I'm getting 13-14 ms times vs 7-8 ms with 91.5 and the GPU utilization is 80% vs 50% (as reported by GPU-Z). Am I doing something wrong? Could "let madVR decide" have a greater effect now? Or is there something else that has changed that much in 91.6?

EDIT: I went back to it and now 91.6 is about the same as 91.5. This is something that I've been seeing sporadically before but haven't been able to reproduce when I want to.

[madshi]

Weird! Which GPU? Can you make a screenshot of the same movie showing the rendering time difference, with the OSD turned on, with both v0.91.5 and v0.91.6? Please also create an empty file "showrendersteps" in the madVR folder before creating the screenshots.

[Magik Mark]

madshi,

Which NGU variant would you recommend that would perform similarly as of neddi3 64 neurons?

[Damien147]

I think a benchmark like SVP for noobs to automatically set settings is a good idea for v1.0.Also is 1.0 gonna be a paid version?Just curious.8 years is a long time and you need something like a ''prize'' for your work.I don't need it but I'll buy it because I feel grateful.:)Maybe a donate button!That works too.


But fix the black screen with dxva2(native) 10bit hevc amd rx4x0 first.:p

[ryrynz]

Which NGU variant would you recommend that would perform similarly as of neddi3 64 neurons?
You know there's only two versions of NGU right? And you read madshi's release post a couple of pages back right?.. It's not hard to put two and two together.

[har3inger]

The anti alias NGU is quite good, and IMO is probably good enough to replace nnedi3. I'm more interested to see how the direct 4x for it will look, since that's the most relevant application for speed/performance ratio on 480p and lower sources.

Did you change something in the NGU sharp scaler? It runs a smidgeon (5-10%) slower than it used to in 91.5.

Edit: Yep. Reverted to 91.5, and lost 3ms worth of render time from 36ms when using NGU high. Unfortunately, this is the difference between perfectly playable and stuttering mess for me. (91.6 is slow. 91.7 is similar, or even a tiny bit slower than 91.6)

[dioxholster]

I'm trying out 91.6 replacing 91.5 on a GTX1060 6gb and trying to have the same settings as before: NGU high for luma and NGU Sharp med for chroma (was NGU med). With the same material I'm getting 13-14 ms times vs 7-8 ms with 91.5 and the GPU utilization is 80% vs 50% (as reported by GPU-Z). Am I doing something wrong? Could "let madVR decide" have a greater effect now? Or is there something else that has changed that much in 91.6?

EDIT: I went back to it and now 91.6 is about the same as 91.5. This is something that I've been seeing sporadically before but haven't been able to reproduce when I want to.



I dont know, but it feels like the framerate is more sped up with the new one, or it could just be me.

[Neo-XP]

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?

It would be the same as the upscaling algo after doubling, as in v0.91.1.
Something like "same as upscaling after doubling algo". Would that be possible ?

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

I can't wait to test that ;)

[AngelGraves13]

I hope NGU AA will get direct 4X soon.

[Arm3nian]

How does NGU quadrupling work with "refine the image after every 2x step / only once" options? If I have sharpen edges at 4.0, refine the image only once, and NGU on "double again" then it seems to be at 4.0. Same with direct quadruple. But if it's "refine the image after every step" then it looks like 8.0. And if NGU is on direct quadruple and I select refine after every step then it overrides it and uses NGU twice. So if you could select sharpen edges at 8.0 then you could use NGU direct quadruple with the same look as doubling twice and refining the image twice right now. But then I guess that would go up to 16 in the new version? :)

[Anima123]

NGU-AA low consumes more GPU power than previous NGU pixart, the Intel Graphics 630 can hardly handle 720p -> 1024p @ 23.976 fps fluently, the playback stuttered from time to time, maybe due to throttling?

[HillieSan]

OK, so both 10 & 12 bit 4:2:2 seem possible in HDMI 2.0b (http://www.acousticfrontiers.com/uhd-101-v2/) but do the RX480 drivers provide an option to do that? FWIR only nvidia allow this kind of kludge in RGB?

Actually that link seems to imply that it would even be possible with HDMI 1.4, it's all pretty dang confusing and I haven't seen any option to output 4:2:2 RGB in the 13.12 AMD drivers.

There is an option in the Crimson driver where you can select 4:4:4, 4:2:2 or 4:2:0 and 8, 10, or 12 bits. You can only selects 8 bits for 4:4:4. This is the same for nVidia. It's the movie that will determine if 8, 10 or 12 bits are send to the TV/Projector.

[madshi]

Which NGU variant would you recommend that would perform similarly as of neddi3 64 neurons?
On a pure speed level, I think "NGA Anti-Alias High" should be faster than NNEDI3-64 (except on Polaris, maybe), while "Very High" might be slightly slower. But why don't you simply benchmark yourself?

If you're asking about quality instead of speed, that's what I actually want to hear from you!!

Did you change something in the NGU sharp scaler? It runs a smidgeon (5-10%) slower than it used to in 91.5.

Edit: Yep. Reverted to 91.5, and lost 3ms worth of render time from 36ms when using NGU high. Unfortunately, this is the difference between perfectly playable and stuttering mess for me. (91.6 is slow. 91.7 is similar, or even a tiny bit slower than 91.6)
I've changed the downscaling algo after doubling/quadrupling. Do you see the same slowdown if you do a straight 2x upscale without any downscaling involved? If yes, please do what I suggested here:

https://forum.doom9.org/showpost.php?p=1799698&postcount=42902

It would be the same as the upscaling algo after doubling, as in v0.91.1.
Something like "same as upscaling after doubling algo". Would that be possible ?
Oh, I see. I'm not sure if anybody would understand that concept, though. I mean one dropdown box referring to another is kinda weird.

How does NGU quadrupling work with "refine the image after every 2x step / only once" options? If I have sharpen edges at 4.0, refine the image only once, and NGU on "double again" then it seems to be at 4.0. Same with direct quadruple. But if it's "refine the image after every step" then it looks like 8.0. And if NGU is on direct quadruple and I select refine after every step then it overrides it and uses NGU twice. So if you could select sharpen edges at 8.0 then you could use NGU direct quadruple with the same look as doubling twice and refining the image twice right now. But then I guess that would go up to 16 in the new version? :)
I can't really follow your thoughts. Where do you get the idea from that madVR would do 8.0 or 16.0 in your case? As you said yourself, if you select "refine after every step", quadrupling is turned off because it can't work in that case.

NGU-AA low consumes more GPU power than previous NGU pixart, the Intel Graphics 630 can hardly handle 720p -> 1024p @ 23.976 fps fluently, the playback stuttered from time to time, maybe due to throttling?
I've changed the downscaling algo after doubling/quadrupling. Do you see the same slowdown if you do a straight 2x upscale without any downscaling involved? If yes, please do what I suggested here:

https://forum.doom9.org/showpost.php?p=1799698&postcount=42902

[Anima123]

I've changed the downscaling algo after doubling/quadrupling. Do you see the same slowdown if you do a straight 2x upscale without any downscaling involved? If yes, please do what I suggested here:

https://forum.doom9.org/showpost.php?p=1799698&postcount=42902

Before I do as what you're asking, need to tell that the rendering time goes from less than 10 ms up to crazy 3x ms average and 6x ms peak, when the stutter began. And after stuttering for some time, the rendering time would go back to normal.

With that kind of case, would it help to do the screen-shot, and when should I do it?

[huhn]

There is an option in the Crimson driver where you can select 4:4:4, 4:2:2 or 4:2:0 and 8, 10, or 12 bits. You can only selects 8 bits for 4:4:4. This is the same for nVidia.
it's not that easy.
It's the movie that will determine if 8, 10 or 12 bits are send to the TV/Projector.

no.

[madshi]

Before I do as what you're asking, need to tell that the rendering time goes from less than 10 ms up to crazy 3x ms average and 6x ms peak, when the stutter began. And after stuttering for some time, the rendering time would go back to normal.

With that kind of case, would it help to do the screen-shot, and when should I do it?
Are the rendering times identical, when comparing NGU-AA vs NGU-pixart, except for those crazy 3x slowdowns? Or are the rendering times also slower when ignoring those 3x slowdowns?

[Anima123]

Are the rendering times identical, when comparing NGU-AA vs NGU-pixart, except for those crazy 3x slowdowns? Or are the rendering times also slower when ignoring those 3x slowdowns?

Under 10 ms for both of your algorithms, except for the crazy period of time.

[madshi]

So the real problem for you are just the crazy slowdowns? No idea how to debug that, unfortunately... :( And you can reproduce it 100% that you get those slowdowns with NGU-AA, but not with NGU-pixart? Does it also happen if you just do a simple 2x upscale without any downscaling involved, and without 4x?