madVR - high quality video renderer (GPU assisted)

[Xaurus]

Originally Posted by pirlouy

Because if you use 24 or 50Hz, you're not in direct RGB 4:4:4 anymore. There will be a conversion (RGB 4:4:4 RGB -> (by TV) YUV 4:2:2 -> (by TV) RGB 4:4:4)
PS: Samsung limitation, but maybe it changed recently, even if I doubt it.

Okay, is there a way for me to check this? I mean, your TV from 2009 is pretty old by today's standards (no offence!).

[pirlouy]

From first post of this thread: http://madshi.net/madVR/ChromaRes.png in full screen

[MSL_DK]

Originally Posted by pirlouy

but maybe it changed recently, even if I doubt it.

They have not changed anything.

[Xaurus]

Originally Posted by pirlouy

From first post of this thread: http://madshi.net/madVR/ChromaRes.png in full screen

Thanks.

Ok, so I'm using 23,976 refresh rate and I can see the image saying 4:4:4. Is there something more I need to do?

[MSL_DK]

Originally Posted by Xaurus

Thanks.

Ok, so I'm using 23,976 refresh rate and I can see the image saying 4:4:4. Is there something more I need to do?

And 4:2:2?

[Xaurus]

Originally Posted by MSL_DK

And 4:2:2?

I see only 4:4:4.

Let me upload a photo:
http://home.halden.net/mordor/img.jpg

[MSL_DK]

Then your display supports RGB 4:4:4, nothing more to think about I envy you your TV

[Xaurus]

Originally Posted by MSL_DK

Then your display supports RGB 4:4:4, nothing more to think about I envy you your TV

Ok cool, so what was mentioned by pirlouy for other refresh rates than 60 doesn't apply to this Samsung set then?

[MSL_DK]

Originally Posted by Xaurus

Ok cool, so what was mentioned by pirlouy for other refresh rates than 60 doesn't apply to this Samsung set then?

No ... and with your picture, I regret that I have not bought a samsung 4k What model is it?

[Xaurus]

Originally Posted by MSL_DK

No ... and with your picture, I regret that I have not bought a samsung 4k What model is it?

It's this one.

A couple of questions:

• Should I still select NV12 in LAV filters?
  
• What is preferred when choosing between NNEDI3 for chroma upscaling, or "use NNEDI3 to double Luma resolution"? Is only one of them needed? For example use Jinc in chroma upscaling then NNEDI3 to double Luma / chroma resolution.

[pirlouy]

By curiosity, when you're in PC mode, are you able to enable "LED Clear Motion" in your Samsung options ? Again I doubt it, but it would be nice if you surprised me again.

[madshi]

Originally Posted by aufkrawall

It will be interesting to see how much ALU power will be left for the renderer with HEVC 10 bit decoding (not supported by Maxwell 2.0 VPU). Luckily, this will most likely apply only to 4k which won't require something like NNEDI3 scaling.

Not sure if the GPU will be used at all for HEVC decoding, if the VPU doesn't support it. I don't really know.

Originally Posted by ThurstonX

Quick question: how can I tell if my Toshiba laptop running Intel HD Graphics 3000 on a Sandy Bridge i5 + nVidia GTX 560M (Optimus) is using NNEDI3?

IIRC, you need at least an HD4000 to get hardware OpenCL support. But even the HD4000 is by far too slow for NNEDI3. You can force madVR to use the NVidia GPU for OpenCL, though, by using the following registry value:

HKCU\Software\madshi\madVR\OpenCL\forceVendor

A search for "forceVendor" should give you more details on how to use that.

Originally Posted by leeperry

Shiandow posted impressive NEDI/NNEDI3 comparisons in this thread, you said that some of your test patterns didn't pass the test but does vegetation look equally artificial?

From what I've seen, although Shiandow did a really good job to reduce these artifacts as much as possible, NEDI has noticeably stronger directional artifacts compared to NNEDI3. If you are bothered by the fractal like NNEDI3 artifacts, you will be even less happy with NEDI, I believe.

Well, at least that's without the SuperRes post processing. Not sure how that changes things. I haven't really had time yet to study Shiandow's algorithms in detail. Hopefully I'll get around to do that rather sooner than later. But I don't know yet...

Originally Posted by Xaurus

Should I still select NV12 in LAV filters?

Leave LAV filters to default. The defaults are usually optimal for madVR.

Originally Posted by Xaurus

What is preferred when choosing between NNEDI3 for chroma upscaling, or "use NNEDI3 to double Luma resolution"? Is only one of them needed?

Best bang for the buck is luma doubling, followed by chroma upscaling. Usually doubling chroma has the lowest benefit. If you have performance to spare, up the neurons for luma doubling.

[Shiandow]

Originally Posted by madshi

From what I've seen, although Shiandow did a really good job to reduce these artifacts as much as possible, NEDI has noticeably stronger directional artifacts compared to NNEDI3. If you are bothered by the fractal like NNEDI3 artifacts, you will be even less happy with NEDI, I believe.

Well, at least that's without the SuperRes post processing. Not sure how that changes things. I haven't really had time yet to study Shiandow's algorithms in detail. Hopefully I'll get around to do that rather sooner than later. But I don't know yet...

I agree that NEDI will likely look more 'artificial' than NNEDI3.

FWIW there'll soon be a working (open source) implementation of SuperRes (with or without NEDI) in MPDN. It's still a bit experimental though. I still need to try out all the things that were impossible to achieve using only shaders, but have become possible with MPDN's more flexible system.

[madshi]

Which limitations of the shaders were the main problem?

[Xaurus]

Originally Posted by pirlouy

By curiosity, when you're in PC mode, are you able to enable "LED Clear Motion" in your Samsung options ? Again I doubt it, but it would be nice if you surprised me again.

No sorry, almost everything in the menu is greyed out, including those kind of options.

[Xaurus]

Originally Posted by madshi

Best bang for the buck is luma doubling, followed by chroma upscaling. Usually doubling chroma has the lowest benefit. If you have performance to spare, up the neurons for luma doubling.

Thanks madshi, this is useful info.

So if I understand you correctly, just push chroma doubling as far as it goes, then put "the rest" into chroma upscaling (something like Jinc)?

[Shiandow]

Originally Posted by madshi

Which limitations of the shaders were the main problem?

Mainly the fact that I couldn't save any intermediate results. I had to somehow store the current image, the original image, and the difference between the current and original image, simultaneously. I could just barely manage this by abusing the alpha channel, but I could save at most 2 channels of the original, it didn't improve the legibility of the code and also limited the kind of processing I could do. For the chroma version of SuperRes it also didn't help that I couldn't make use of MadVR's chroma conversion. I could either disable it entirely or try to reverse it. In both cases I essentially needed to reverse engineer MadVR's chroma processing.

That said not even I am entirely sure why but for some reason the adaptation I made for MPDN seems to beat the (unpublished) experimental version of SuperRes that I had developed. Maybe because I can save more information of the original image, or because it sped up enough that I could use 16 bit processing.

Oh and the fact that I could reuse MPDN's scaling algorithms also helped. Although in the end the shader version couldn't handle anything more complicated than bilinear anyway because I had to do everything in 1 pass (nowhere to store the intermediate result).

[ryrynz]

Originally Posted by Xaurus

So if I understand you correctly, just push chroma doubling as far as it goes, then put "the rest" into chroma upscaling (something like Jinc)?

You mean luma and yes. I'd recommend Jinc AR on chroma for bang for buck also.

[madshi]

Originally Posted by Shiandow

Mainly the fact that I couldn't save any intermediate results. I had to somehow store the current image, the original image, and the difference between the current and original image, simultaneously. I could just barely manage this by abusing the alpha channel, but I could save at most 2 channels of the original, it didn't improve the legibility of the code and also limited the kind of processing I could do. For the chroma version of SuperRes it also didn't help that I couldn't make use of MadVR's chroma conversion. I could either disable it entirely or try to reverse it. In both cases I essentially needed to reverse engineer MadVR's chroma processing.

That said not even I am entirely sure why but for some reason the adaptation I made for MPDN seems to beat the (unpublished) experimental version of SuperRes that I had developed. Maybe because I can save more information of the original image, or because it sped up enough that I could use 16 bit processing.

Oh and the fact that I could reuse MPDN's scaling algorithms also helped. Although in the end the shader version couldn't handle anything more complicated than bilinear anyway because I had to do everything in 1 pass (nowhere to store the intermediate result).

Ok, thanks. Sounds like the plans I have should take care of most of those problems.

[Xaurus]

Originally Posted by ryrynz

You mean luma and yes. I'd recommend Jinc AR on chroma for bang for buck also.

Yes luma doubling, of course. Thanks guys.