madVR - high quality video renderer (GPU assisted)

[Werewolfy]

Originally Posted by omarank

I tried playing with the Custom Deband settings and settled on “angleBoost” 1.5 and “maxAngle” 0.08. With these settings, the debanding effect seems to be the same as with the old “high” preset and the image is apparently a bit sharper.

@Werewolfy: You found that “angleBoost” at 1.5 did some harm to a specific scene. Maybe you would want to test that again with the above settings. On all the samples that I tried, I didn’t see any kind of artifact or loss of detail with it set to 1.5. Can you post the original image or screenshot of the scene without debanding being applied?

I tested it and even if I set the maxAngle to 0.01, it harms the picture with Angleboost to 1.5.
Here's the video sample : https://dl.dropboxusercontent.com/u/...nshin_-_01.mkv
Look at the left corner of the picture.

[Siso]

Originally Posted by madshi

That's ok.

In max stats/present stats it shows more than 1.70 ms, is this ok?

[Qaq]

Originally Posted by Qaq

Even if *luma* is being "upscaled" from 1076 to 1080 (WEB-DLs)?

Originally Posted by nevcairiel

Even if Luma isn't being upscaled at all but just kept 1:1, you need to upscale chroma by 2x.

I understand L and C should be matched 1:1 in "size". I thought with 1-2% of "upscaling" chroma is being changed a very weird way.

[FireFreak111]

I am sort of eating my words on NNEDI3 being completely stable on Win 10. It works perfectly for Image Doubling and Chroma doubling, (if Chroma doubling the second Chroma > Jinc3 AR disappears), but for Chroma upscaling, it instantly crashes my display driver, and no image is shown (black screen). I can't pull up an OSD, but I can switch between FSE and windowed. Restarting MPC-HC with Jinc3 AR brings back the image. Same occurs if I reenable NNEDI3.

Shame, since its almost completely working. Any idea why only Chroma upscaling would be broken? Win 10 likely won't be released until mid July, so there's some time before it needs to be directly addressed, its just puzzling that Chroma doubling works but Chroma upscaling doesn't.

Originally Posted by nevcairiel

When you use SM, Dithering should already be performed at the refresh rate, since its the very last step in the entire process.

That's good to hear then, no other features I can think of besides a final choice on the debanding algo's.

[vivan]

Originally Posted by nevcairiel

When you use SM, Dithering should already be performed at the refresh rate, since its the very last step in the entire process.

SM doesn't blend frames at refresh rate, it only blends frames that overlap. E.g. for 24 fps over 60 hz: A A AB B B. That's 3 unique frames for 5 refreshes, not 5.

[mogli]

Some more with 'present a frame for every VSync' turned off:
Found a NTSC file which reliably causes constant glitches only in FSE mode with deinterlacing off then. Using deinterlacing (video or film) in FSE mode or windowed mode (with or without deinterlacing) or turning 'every VSync' on will not glitch.

[x7007]

FSE mode 10bit just doesn't run well for me in windows 7.

If I use FSE 10 bit + Exclusive mode in Potplayer I just get high rendering up to 40ms , in windows mode it's 22 ms. I can't understand what is the issue.

Windows 7 x64
Potplayer x32 Exclusive Mode Enabled - fastest

How much rendering do you get in FSE 10 bit ?

It happens in all sources , MKV , 1080P , Blu-ray AVC . 720p.

EDIT : If I enable " Present a frame for every sync " The rendering goes to 30ms but Render Queue and Present Queue doesn't max out. Render 2-8 /8 - Present 1-3 /3 . it should be Render 7-8 /8 , Present 2-3 /3. How to fix this bug or settings or what causing this.

[nevcairiel]

Originally Posted by vivan

SM doesn't blend frames at refresh rate, it only blends frames that overlap. E.g. for 24 fps over 60 hz: A A AB B B. That's 3 unique frames for 5 refreshes, not 5.

Actually thats incorrect. It may appear a simple choice to do it like this, but it is not the way SM works. I assumed that at first as well, but madshi corrected me.

Its more like this:

24p: 0ms=A, 42ms=B

60p: 0ms=A, 16ms=0.6A/0.4B 32ms=0.2A/0.8B, 48ms=0.9B/0.1C

(Factors and times rounded for simplicity)

A clean "A" is only output when the timestamps match perfectly, otherwise there is always blending performed. In the chain above, C would be output clean again at 83/84ms.
As a result of this, you can pretty much assume you get a new frame for every refresh.

If you think about it, it also makes sense. Only this way you can actually get rid of the 3:2 judder, the more simple approach wouldn't eliminate it.

[vivan]

Originally Posted by nevcairiel

If you think about it, it also makes sense. Only this way you can actually get rid of the 3:2 judder, the more simple approach wouldn't eliminate it.

Nope, it doesn't :P Try avisynth's ConvertFPS (60) - you'll see crazy amount of blending.

When rates don't perfectly match (which indeed always happens) - instead of last clean B you get BC. Basically two 3:2 patterns are blended together. That gives 2 same clean frames in succession for every pattern (5 refresh rates) or 48 fps.
SM capture https://dl.dropboxusercontent.com/u/...2014-13-16.mp4
1-bit dithering without SM https://dl.dropboxusercontent.com/u/...2014-27-30.mp4
1-bit dithering with SM https://dl.dropboxusercontent.com/u/...2014-27-50.mp4

(but it's possible to make them match by messing with audio - AFAIK this is what mpv does, they have fixed pattern and built-in "reclock")

[Anime Viewer]

Originally Posted by madshi

How high are your present numbers exactly? Anything lower than 1ms should be ok.

I didn't realized present ms is normally supposed to be so low. Mine only seem to run as low as 3.10ms-5ms (720p expanded to 1920x1080), 1.96ms (in 720p native window), 7.51ms-9ms (480 video expanded to 1920x1080) 2.26ms 480 native window. (When I say native window size I mean with settings of Zoom 100%, and normal size for video frame in view settings in MPC-HC). I've experimented with present frames in advance settings in every setting from 14 to 1 with nothing better than those stats (and with settings as low as 1 my present queue tends to drop to 0 and report presentation glitches and dropped frames).
What other settings besides present frames in advance may be effecting present ms?
(My Nivida settings for mpc-hc are set to set to maximum pre-rendered frames use global settings (use the 3d application setting), and virtual reality pre-rendered frames (guessing this had no baring since its only supposed to effect virtual reality head sets) also at use global settings (use the 3d application setting)).

[Sunset1982]

Madshi is there a roadmap for future features or do you habe any feautre plans?

[ryrynz]

Anyone here can answer that :P No roadmap and he doesn't divulge future plans, sometimes he lets the cat out of the bag early though.

[huhn]

Originally Posted by vivan

Nope, it doesn't :P Try avisynth's ConvertFPS (60) - you'll see crazy amount of blending.

When rates don't perfectly match (which indeed always happens) - instead of last clean B you get BC. Basically two 3:2 patterns are blended together. That gives 2 same clean frames in succession for every pattern (5 refresh rates) or 48 fps.
SM capture https://dl.dropboxusercontent.com/u/...2014-13-16.mp4
1-bit dithering without SM https://dl.dropboxusercontent.com/u/...2014-27-30.mp4
1-bit dithering with SM https://dl.dropboxusercontent.com/u/...2014-27-50.mp4

(but it's possible to make them match by messing with audio - AFAIK this is what mpv does, they have fixed pattern and built-in "reclock")

i see the same and there is a easy way to test this.

disable present frames in advance
take a 23p source and play at 60 hz with smooth motion.

now look at the MPC-HC statistics you will see about ~48 FPS

[pirlouy]

I think Nevcairiel is right for Smooth Motion:
http://forum.doom9.org/showthread.ph...33#post1616533

The exact blending factors (and whether blending is needed at all) are calculated for every vsync interval. The blending factors usually change all the time.

[madshi]

Originally Posted by TheShadowRunner

Yes, I got a crash box and a full report. I though the information would be more complete with a debug log, sorry ^^;
It's madVR that seems to crash, not the player (ZP).
Using DX9, on WinXP SP3 still.
Again this crash showed up with 0.88.9. I don't believe you can reproduce except if you're on XP (?).
In this case, just open the player, load a video, go in fullscreen mode (it plays fine) and then just go back to windowed mode (crash exactly happens at this point).

Ok, will look at that later.

Originally Posted by QBhd

So I decided to try the AMD registry hack to disable dithering on my R9 270X, and I am rather impressed with the way things turned out. I have a Plasma (circa 2007) with a 1024x768 rectangular pixel resolution... And I must say, it likes Ordered-Colored-Static dithering much better now. Without the reg hack, I always settled on ED1-Colored-Static, but Ordered seems to be a lot less noisy now... and I had done extensive testing when the dithering options were first implemented, and Ordered used to be very noisy. It seems that eliminating that intermediate dithering (GPU) outputs a cleaner picture. So, in this case (AMD, old Plasma and madVR), is this a placebo, or is there just too much dithering going without the registry hack. Thoughts? Anyone else with AMD and Plasma tried this?

You see a difference even with 8bit output? That's surprising. I would have expected AMD dithering to only be active when using 10bit output (and only when using a display which doesn't support 10bit).

Originally Posted by omarank

I tried playing with the Custom Deband settings and settled on “angleBoost” 1.5 and “maxAngle” 0.08. With these settings, the debanding effect seems to be the same as with the old “high” preset and the image is apparently a bit sharper.

@Werewolfy: You found that “angleBoost” at 1.5 did some harm to a specific scene. Maybe you would want to test that again with the above settings. On all the samples that I tried, I didn’t see any kind of artifact or loss of detail with it set to 1.5. Can you post the original image or screenshot of the scene without debanding being applied?

Originally Posted by Werewolfy

I tested it and even if I set the maxAngle to 0.01, it harms the picture with Angleboost to 1.5.
Here's the video sample : https://dl.dropboxusercontent.com/u/...nshin_-_01.mkv
Look at the left corner of the picture.

So 1.4 and 0.08 it is?

Originally Posted by Siso

In max stats/present stats it shows more than 1.70 ms, is this ok?

Yes.

Originally Posted by FireFreak111

I am sort of eating my words on NNEDI3 being completely stable on Win 10. It works perfectly for Image Doubling and Chroma doubling, (if Chroma doubling the second Chroma > Jinc3 AR disappears), but for Chroma upscaling, it instantly crashes my display driver, and no image is shown (black screen). I can't pull up an OSD, but I can switch between FSE and windowed. Restarting MPC-HC with Jinc3 AR brings back the image. Same occurs if I reenable NNEDI3.

Shame, since its almost completely working. Any idea why only Chroma upscaling would be broken? Win 10 likely won't be released until mid July, so there's some time before it needs to be directly addressed, its just puzzling that Chroma doubling works but Chroma upscaling doesn't.

No idea. Can anybody else reproduce this?

Originally Posted by x7007

FSE mode 10bit just doesn't run well for me in windows 7.

If I use FSE 10 bit + Exclusive mode in Potplayer I just get high rendering up to 40ms , in windows mode it's 22 ms. I can't understand what is the issue.

Windows 7 x64
Potplayer x32 Exclusive Mode Enabled - fastest

How much rendering do you get in FSE 10 bit ?

It happens in all sources , MKV , 1080P , Blu-ray AVC . 720p.

EDIT : If I enable " Present a frame for every sync " The rendering goes to 30ms but Render Queue and Present Queue doesn't max out. Render 2-8 /8 - Present 1-3 /3 . it should be Render 7-8 /8 , Present 2-3 /3. How to fix this bug or settings or what causing this.

Do you have an NVidia GPU? Then double check that the setting "max prerendered frames" (or something like that) is set to application control.

Originally Posted by Anime Viewer

I didn't realized present ms is normally supposed to be so low. Mine only seem to run as low as 3.10ms-5ms (720p expanded to 1920x1080), 1.96ms (in 720p native window), 7.51ms-9ms (480 video expanded to 1920x1080) 2.26ms 480 native window. (When I say native window size I mean with settings of Zoom 100%, and normal size for video frame in view settings in MPC-HC). I've experimented with present frames in advance settings in every setting from 14 to 1 with nothing better than those stats (and with settings as low as 1 my present queue tends to drop to 0 and report presentation glitches and dropped frames).
What other settings besides present frames in advance may be effecting present ms?
(My Nivida settings for mpc-hc are set to set to maximum pre-rendered frames use global settings (use the 3d application setting), and virtual reality pre-rendered frames (guessing this had no baring since its only supposed to effect virtual reality head sets) also at use global settings (use the 3d application setting)).

Not sure why the numbers are so bad for you. Is it the same for D3D9 with "use a separate device for presentation" either enabled or disabled?

-------

About smooth motion FRC. You can think of it like this: madVR tries to simulate a digicam which is filming a perfect monitor, with the digicam framerate being identical to the refresh rate of the GPU output.

E.g. the perfect monitor for 24fps would display the "video frames" ("vf") like this:

[-vf.1-][-vf.2-][-vf.3-][-vf.4-]

Now imagine a digicam filming this at 60fps:

[  vf1   ][  vf2   ][  vf3   ][  vf4   ]
 [d1][d2][d3][d4][d5][d6][d7][d8][d9]

You can easily see that the "d1" and "d2" digicam frames only cover the "vf1" video frame. So madVR doesn't blend either "d1" or "d2". madVR actually only renders "d1" and not "d2" to save performance. "d3" is blended, using 0.25 * vf1 + 0.75 * vf2. "d4" is not blended etc.

If you think about it, what madVR does is pretty much the same as what a digicam would capture. And that's good enough to achieve smooth motion. If I blended every frame, the result would not be smoother, just more blurry. The exact blending factors are not fixed, but are calculated dynamically, based on the exactly measured vsync position.

[huhn]

Originally Posted by madshi

Same here.

I've tried to find any API that would tell me whether the GPU actually outputs 10bit or not, but I've not found anything, at least not nothing official GPU independent from Microsoft. I can see that Direct3D11 actually tells the GPU driver to output 10bit, and the driver accepts it, although I know for a fact that my display doesn't support 10bit. So basically it's a black box and I don't know what the GPU is doing with my 10bit output...

my new TV shows the input bit deep. i test nvidia later to see what the GPU is outputting under different situations.

the screen needs to be replaced. so not much time to test with it for now.

[aufkrawall]

I still think that Shiandow with low power is a nice addition to the madVR high deband setting of the current build (judged by the Tagesschau sample).

[kalston]

Originally Posted by madshi

You see a difference even with 8bit output? That's surprising. I would have expected AMD dithering to only be active when using 10bit output (and only when using a display which doesn't support 10bit).

AFAIK it applies for 8bit too, people have reported a significantly better PQ on AMD cards when using icc profiles because of that (with nvidia you get noticeable banding with icc profiles).
- I don't use AMD hardware, I don't know anything more than that but I have read it many times, especially when those cheap (and overclockable) Korean 1440p panels were the craze and required calibration to look decent.

[QBhd]

Originally Posted by madshi

You see a difference even with 8bit output? That's surprising. I would have expected AMD dithering to only be active when using 10bit output (and only when using a display which doesn't support 10bit).

I will double check and triple check... like I said, I was wondering if I was falling victim to placebo effect. But yes, after the reg hack I thought I saw a difference. I will remove them and test the dithering options again to be certain.

QB

[Shiandow]

Originally Posted by madshi

If you think about it, what madVR does is pretty much the same as what a digicam would capture. And that's good enough to achieve smooth motion. If I blended every frame, the result would not be smoother, just more blurry. The exact blending factors are not fixed, but are calculated dynamically, based on the exactly measured vsync position.

By the way, I've been wondering, do you also change the blending factors depending on the horizontal position of the image? It seems like that would matter as well, at least for most screens.