madVR - high quality video renderer (GPU assisted)

[huhn]

Originally Posted by XMonarchY

But I do get the right colors. I verified with HCFR Calibration when I placed Rec.601 3DLUT into Rec.709 line and selected Rec.709 in madTPG for verification. 3DLUT already sets the colors for you. I don't think madVR messes with them...

you verified a smpte C calibration treated as bt 709 with bt 709.

if you got correct results fine but... that's not how it should work.

Madshi, sorry, I did not see your answer, but I previous posted my exact settings. Here they are again:
3DLUT
Artifact Removal = Low for default, Medium for fade in/out + Shiandow's Deband with default settings
Image Enhancement = none
Chroma Upscaling = NNEDI3 32n + SuperRes Filter with default settings
Image Downscaling = Catmull-Rom + AA + Scale in Linear Light
Image Doubling = NNEDI3 32n (set to Always) if the content is sub-1080p on my 1080p monitor/TV
Image Upscaling = Jinc 3tap + AA
Image Refinement = SuperRes with High setting for 1080p content only
D3D11 & Exclusive mode = disabled

Problems:
Image Quadrupling = severe frame drops for sub-1080p content, even when SuperRes is disabled for Image Refinement and Chroma Upscaling. Used to use-able (both Chroma and Luma) in madVR 0.87.21 without frame drops.
Image Refinement = SuperRes can only be enabled IF Chroma Upscaling SuperRes filter is NOT activated for sub-1080p content, regardless of High/Medium setting selected for Image Refinement. For 1080p content, both Filter and Refinement can use SuperRes without problems.
D3D11 = severe presentation glitches, even without frame drops and/or delays.

Some questions I had:
1. Why is Quadrupling creating frame drops in the latest version of madVR, but not in the older version, where I had both Luma and Chroma Doubling and Quadrupling, all set to NNEDI3 32n? This occurs regardless of whether I use SuperRes Chroma Upscaling filter and/or SuperRes in Image Refinement. I tried with no SuperRes enabled anywhere, but Quadrupling still produced severe frame drops.
2. Why can I not enable SuperRes in both places - Image Refinement and Chroma Upscaling for sub-1080p content? Is it just too heavy on my system?
3. What is the benefit of using D3D11?
4. What are D3D11 presentation glitches? They often occur without dropped and/or delayed frames. Should they be ignored?
5. Is there any point to use D3D11 "present a frame for every VSync" option if one is running ReClock?
6. Are there any comparisons of how SuperRes Chroma Upscaling filter and SuperRes Image Refinement compare against each other?

[aufkrawall]

Originally Posted by madshi

In the NVidia control panel, have you set "Max Pre-Rendered Frames" to "application controlled"?

No, it was set to 1. With app controlled, it says 7-8 / 8 now.
I was about to write that it's working well, but when switching from FSE to windowed via hotkey, MPC HC x64 crashed without error description.

Btw: When using e.g. FineSharp as an upscaling refinement together with NNEDI3 image doubling, an issue with the Nvidia driver is triggered: The image gets corrupted as shown in the following (dark & red) and the NV driver fails to enter fast clock states, it can't go any higher than P5 state (faster P2 state would be correct for OpenCL/CUDA):

Often MPC HC also crashes when this issue occurs.

[tobindac]

Demonstrating what big leap is one extra bit. It's an 8-fold increase every step. But that also means if you can already not see differences after 7, 8 or 9, the diminishing returns would be enormous, unless of course you're only getting rounding error improvements.

[JarrettH]

Originally Posted by tFWo

0.88.2 Win8.1x64

MPC-HC (32bit) fails to exit properly. I have to end task manualy.

I noticed MPC fails to exit as well and posted about it:

http://forum.doom9.org/showthread.ph...41#post1721141

[tFWo]

Originally Posted by JarrettH

I noticed MPC fails to exit as well and posted about it:

http://forum.doom9.org/showthread.ph...41#post1721141

Download the latest version 1.7.8.162 and it should be fixed.

[e-t172]

Originally Posted by dbcooper

When the DCI standards were developed, tests were made on audience ability to see differences in luminance gradient levels. IIRC, many could see differences equivalent to 10bit gradients, and some could see 11bit gradient differences.

Without dithering, maybe. With proper dithering, I would be very, very surprised if this is actually true.

Originally Posted by dbcooper

Another benefit of 10bit output from MadVR is with the use of 3D LUTs. With RGB gamma ramps or a 3DLUT, you need to change the spacing between RGB values from the default unity, to values that minimize display error. In doing so, banding can be introduced (you only have 0 to 255 to represent all 256 levels in 8bit). Using 10bit output to a 10bit display for 8bit material can reduce this banding, and give a better calibration.

AFAICT, when using a 3DLUT within madVR, dithering is done after the 3DLUT and before conversion to integer, effectively removing the banding introduced by the coarse steps.

[dbcooper]

Originally Posted by tobindac

Demonstrating what big leap is one extra bit. It's an 8-fold increase every step. But that also means if you can already not see differences after 7, 8 or 9, the diminishing returns would be enormous, unless of course you're only getting rounding error improvements.

That is for all colours, not for gradients. Gradients go in one bit steps. i.e. 256, 512, 1024.

[Qaq]

Originally Posted by tobindac

are you a co-programmer of this by the way, because it's closed source.

I'm not involved in madVR development, but I do programm some stuff at my work (measurements & automation).
madVR code is closed, but its technology was explained and discussed in details many times.

Originally Posted by tobindac

To what extend that can be ever noticed is another question.

I wouldn't be surprised if it's humanly impossible to see.

Well, we can spend tons of precious time on tests (then upgrade our software/hardware and start it again) or just use insine precision and call it a day. 10-12 bits don't cost anything to us.
I think I've made myself clear now, so I stop right here to not mess with the actual discussion.

Originally Posted by e-t172

AFAICT, when using a 3DLUT within madVR, dithering is done after the 3DLUT and before conversion to integer, effectively removing the banding introduced by the coarse steps.

Correct! Even I know that

[Hprd]

Not sure if this has been reported, but I get a total freeze/crash (have to kill MPC-BE with task manager) when entering fullscreen with super res enabled for upscaling (chroma is fine) with image quadrupling (nnedi3, any setting) regardless of FSE mode (d3d11/9, 8-10bit). With no super res, video plays fine with quadrupling. On this screen (1920x1200) only affects DVD resolution videos obviously.

[Zachs]

Hi madshi,

MadVR is currently in violation of MPDN_Extension's license which is LGPLv3. Can you please correct this violation?
For more information, see MPDN's thread here - http://forum.doom9.org/showthread.ph...94#post1721194

Thanks.

[har3inger]

Has anyone with switchable graphics managed to get the dx11 working? I get a black screen with audio, and no OSD.

(Same thing has always happened if I try to use a separate rendering device)

[ryrynz]

Originally Posted by Qaq

Well, we can spend tons of precious time on tests (then upgrade our software/hardware and start it again) or just use insine precision and call it a day. 10-12 bits don't cost anything to us.

Exactly. Set to 10 bit, avoid dithering to gain some performance if u want as chances are u won't see any benefit enabling it (tests required) and call it a day.
I think dithering looks a bit messy on a plasma less noise = cleaner/clearer picture.

[James Freeman]

Adding to the discussion about 8bit vs 10bit;

You are forgetting that Contrast Ratio has a huge influence on whether you'll see the difference between steps or not.
8bit on a typical 1000:1 display is almost enough to not see the gradients, but on a 20,000:1 plasma the gradients are clear as day, 10bit on a 1000:1 is an overkill.
Now if HDR display will have a contrast ratio or 100,000:1 it most definitely will benefit from 10bit+ we'll see.
This on a smooth gradient pattern which is far from what real video is;
Quality (original -> not re-compressed) bluray video has its own dithering done when it was converted from 12bit DCI to bluray, so the chances anyone will see banding is slim.
On a super compressed anime where smooth color changes are almost like a test pattern, is a different story.

I wonder what is the visible bit depth of Dithered 10bit... maybe like 14bit?
And to think that madVR is capable of that right now...

madshi,
I'm totally satisfied with the current 3 Debanding options.
They are doing to job just fine, from 1GB super compressed flicks to full 50GB remuxes.
I don't see any detail loss or banding when it is doing its job.

[ladersu]

madVR 0.88.2 32bit
ati 5770 13.12
Win7 64bit

d3d11 render queue doesn't fill up, and present time is 0.26 vs 0.07 for d3d9.

[harshal_rio91]

Originally Posted by har3inger

Has anyone with switchable graphics managed to get the dx11 working? I get a black screen with audio, and no OSD.

(Same thing has always happened if I try to use a separate rendering device)

Works fine on my laptop with Intel HD 4600 + GTX 860M

[baii]

0.88.2 seem to broke next/skip file for me, tried default setting, both mpc-hc and pot player does it. revert back to 0.88.1 and it is fine.

64bit nnedi3 will make image darken, it that known issue?

[ryrynz]

Originally Posted by James Freeman

8bit on a typical 1000:1 display is almost enough to not see the gradients, but on a 20,000:1 plasma the gradients are clear as day, 10bit on a 1000:1 is an overkill.

Overkill that costs you nothing. Disable dithering and gain performance.

[Schwartz]

It appears that homemade VfW H.264 put into .avi files cannot be played in 0.88.2 anymore without an ugly green bar at the top of the video. It used to be good with 0.87.X. And yeah, I realize you're not supposed to record H.264 in .avi, but it used to work fine. (and I put them in .mkv afterwards, but still no dice)

Edit: Looks as if the chroma part of the image is 'pushed downwards' by X amount of pixels. Smaller video, bigger bar.

Link to a small test video.