madVR - high quality video renderer (GPU assisted)

[ryrynz]

Simple MadVR options..

"Should I enable X?"
"Can you see a difference enabling X?"

Repeat ad nauseam.

Think we've covered this about half a dozen times already.

[Manni]

Originally Posted by Asmodian

Starting with madVR's dithered 10-bit output, what would be the difference between 10 bit with trailing zero and "true" 12 bit? The way you perfectly convert 10 bit to 12 bit is to add two zeros. This is a 2D image, we do not have new samples or anything to interpolate between. The best thing to do is simply add zeros, any further processing would generate some non-zero bits in the least two significant positions, reducing the effective bit depth loss from that processing step, but the source would simply have nothing but zero in those positions.

I didn't say one would be better than the other, I just said that I have no idea what the driver does. So I don't know if it pads the 10bits handed by MadVR with 00 (which I indicated is more likely and I agree would be preferred) or if the driver does some kind of interpolation resulting in "true" 12bits. This is why I put quotes around the true, to make it clear that it would be "true" in name but not necessarily better. I clearly failed.

Now my fingers are tired too, so I'm off

[Asmodian]

There is nothing to interpolate through, I just don't understand what you imagine it could be doing. Making up non-zero values? There is only one value, nothing to interpolate between.

[huhn]

there is a way to fill it up with something else than 00. but we don't know and we can't change it anyway.

[mclingo]

Originally Posted by ryrynz

Simple MadVR options..

"Should I enable X?"
"Can you see a difference enabling X?"

Repeat ad nauseam.

Think we've covered this about half a dozen times already.

I think what some people want is objective answers and with MADVR there should be some objective advice that can be given. However with everyone having different eyesight, TV's, lighting conditions and setups in general most of the processing done in MADVR can now really on be subjectively different / better / worse in most cases.

I am aware there is a clear objective difference between 444 rgb and 420 ybcr but I personally choose to run everything in 10 bit 4:2:0, I get hammered for this everywhere but I genuinely see no difference at all between 8 bit 444 full RGB dithered and 10 bit 420, or maybe I should say I see no difference in real world viewing. I do however see a difference between madvr and standard EVR renderer.

set this way you dont have to worry about bit depth switching, bandwidth or anything, everything just works as it should.

I did try using 4:2:2 but I get small picture dropouts, a couple every evening, I see my receivers DIGI icon disappear when it happens so its losing connection, something in my setup cant handle that extra bandwidth, i've given up trying to find out what after 3 sets of HDMI cables. I'm sure its bandwidth as it doesnt do it at 422 4k 24hz, only 422 4k 60hz.

[Manni]

Originally Posted by Asmodian

There is nothing to interpolate through, I just don't understand what you imagine it could be doing. Making up non-zero values? There is only one value, nothing to interpolate between.

I'm not saying there is something to interpolate. I'm only saying that I didn't look with an analyzer at the bits coming out of the GPU to be able to say for sure what the driver does. I've seen Sony bluray players "creating" data instead of padding with zeros when enabling "deep color", so I'm not going to say that this is what it is when I have no evidence. What makes you so sure that a driver might not do the wrong thing and "interpolate"?

If you are 100% sure that the driver pads with zeros because you've looked at the stream coming out of the nVidia GPU set to 12bits when MadVR sends 10bits dithered, by all means say so, otherwise please allow me to not state something that I am not sure of.

The only thing I can guarantee is that the driver does output 12bits when set to 12bits, because I can see that with the Vertex. What's in the last two bits, I can't say for sure, even if chances are that it's padded zeros (fingers crossed).

[huhn]

because 8 bit 11111111 is not the same as 10 bit 1111111100 in term of top brightness 1111111111 has the same brightness.

[Asmodian]

I simply don't understand what you imagine it might be doing? Making up non-zero values? There is only one data point, you cannot interpolate with one point. I suppose the driver could add noise?

Edit: Ah, true huhn. So maybe they strech the range 000000000000-111111111100 to 000000000000-111111111111, with dithering, which would be odd but possible.

[nevcairiel]

There is actually a bit of a trick when extending the bitdepth of a full range signal, you fill the new bits with the leading bits of the pixel, ie. like this for 8 to 10-bit:

- You shift to 10-bit first, which adds two empty bits.
- Then you fill those empty bits with the top bits from the original signal

  1111111100
+         11111111
= 1111111111

This has several good properties, namely:
- All 1 also remains all 1, ie. maximum 8-bit (255) remains maximum 10-bit (1023)
- Zero also remains zero
- Its easy and fast

I can't say that this is what its doing, but it is generally regarded as producing a more faithful signal when increasing bitdepth then plain zero padding, and a full stretch from 0-255 to 0-1023 is computationally rather expensive.

Note that this does not apply when you are dealing with limited range (ie. 16-235), because 16 and 235 map to the limited-range 10-bit values exactly when you simply shift them up (ie. 64 to 940)
But when handling full-range signals like RGB, plain zero padding is not entirely accurate.

[e-t172]

Originally Posted by ryrynz

Simple MadVR options..

"Should I enable X?"
"Can you see a difference enabling X?"

Repeat ad nauseam.

Think we've covered this about half a dozen times already.

That's not a good answer though. Checking if an option makes a visible difference can take a lot of time, because ideally you want to check a variety of different scenes, maybe even different formats of various bitrates. This gets very time-consuming, especially if you want to do that for several options or for several possible settings of a given option. It would be way more useful to do something more universally applicable, like conducting surveys to see if a given option makes a difference for most people, and if it does, how big the difference is. This could provide reasonable "guidelines" for people to follow.

Honestly I think the "what don't you try it" answer is a bit rude IMHO. It provides no information, and presumably if the user had the time and motivation to do the comparison themselves they would already have done so. If you don't know the answer, then don't post.

[Manni]

Originally Posted by nevcairiel

There is actually a bit of a trick when extending the bitdepth of a full range signal, you fill the new bits with the leading bits of the pixel, ie. like this for 8 to 10-bit:
/snip

Thank Nev, glad that I was right to err on the side of caution (when you don't know, don't assume is my motto). I've learnt my lesson

[huhn]

there are no general answer so we simply don't say anything anymore or we say there are defaults keep them?


and manni if you really want to get an headache think what this does to the dithering and how do you turn this back to a lower bit deep.

[ryrynz]

Originally Posted by e-t172

That's not a good answer though.

My point is... that it's already been answered many times.

Originally Posted by e-t172

Honestly I think the "why don't you try it" answer is a bit rude IMHO. It provides no information, and presumably if the user had the time and motivation to do the comparison themselves they would already have done so. If you don't know the answer, then don't post.

It's really not, it's the most useful information one could provide, and in this case it's very much the answer. Not having "time or motivation" are nothing but excuses.

Asmodian, perhaps you could add a bit more info on this in your guide?

[Asmodian]

Originally Posted by Manni

Thank Nev, glad that I was right to err on the side of caution (when you don't know, don't assume is my motto). I've learnt my lesson

I don't know, stating your assumption can also be very helpful. I would never have guessed Nev's trick but it is probably what really happens. I would never have known if Nev didn't need to correct me.

[huhn]

or they do 16 bit dither to output.
again we don't know and we can't change it.

[nevcairiel]

Originally Posted by Asmodian

I would never have guessed Nev's trick but it is probably what really happens.

Its definitely in use in some things, but I can't exactly dump the HDMI data into a file to look at it, so... (now that would be a handy thing to have!)

[huhn]

there are more than enough lossless HDMI capture cards. they just have there price...

[Asmodian]

Hmm.. I actually have one... Maybe I should do some testing.

[BatKnight]

Originally Posted by brazen1

(...) Fwiw, my 'go to' movie for checking this discussion is Allied 2016. Scene 2:15 through 3:00 shows a dessert slow pan with a cloudy sky. That sky shows banding using 12bit settings in NCP. Using the 8bit settings (that it's going to revert to after a reboot anyway), there is no banding. (...)

I was puzzled by this and had to try it myself. And brazen1 is right, to a certain extent.
I've come to a conclusion, but I don't know how to explain why it happens. I tested wit the movie Allied, too and these are my findings:

Whether I chose "Use default color settings" or "Use NVIDIA color settings with RGB 12 bpc" didn't matter for the results I got. Either settings ended up with the same result, but to be consistent I used these settings (https://imgur.com/bx6QlS8) during the following tests:

1st test: enable "passthrough HDR content to the display" and enable "send HDR metadata to the display". This resulted in an automatic switch to HDR mode when I played the video and the madVR OSD shows (NV HDR): https://imgur.com/OZnoM6A

I can see banding on the blue sky when on D3D11 fullscreen windowed (10 bit) and the banding disappears with I right click the video and it shows D3D11 fullscreen windowed (8 bit)

2nd test: enable "passthrough HDR content to the display" and disable "send HDR metadata to the display". This doesn't automatically change to HDR mode and I have to manually enable "HDR and advanced color" on the Windows 10 Display Settings and the madVR OSD shows (OS HDR): https://imgur.com/i1hLNmC

I can't see any banding on the blue sky when on D3D11 fullscreen windowed (10 bit) or (8 bit).


I believe that when playing a 2160p 10bit HDR 23.976fps video and having madVR set up to output 10 bit, that all is going the best way possible.
What I don't understand is why when sending the HDR metadata, that results on NV HDR, I get banding and when not sending the HDR metadata, and manually enabling the OS HDR, I get perfect image and no banding... Why is OS HDR behaving differently than NV HDR for the same settings?

[seiyafan]

Is 3GB card enough for upscaling to 1440p? I am thinking of getting a 1060 instead of 1050TI but the 6GB version is too expensive right now. Not sure if 1050TI is enough for upscaling to 1440 though.