madVR - high quality video renderer (GPU assisted)

[flossy_cake]

Originally Posted by scollaco

Are there any issues with this card playing 23.976 at 3840x2160p? Do I still need to come up with custom refresh rate for 23.976?

Your JVC projector almost certainly supports 23.976hz and probably defines that mode in its EDID so you will probably get a 23hz mode appearing in NVCP by default. Although NVCP may or may not incorrectly label it 24hz. If in doubt see Windows display settings as that seems to correctly differentiate between 23 and 24.

The other issue is that on ctrl+J you may get 1 dropped frame every x minutes and if that bothers you then one solution is to nudge the refresh rate slightly by like 0.0005hz via a custom res until the drops are minimised to your satisfaction.

Originally Posted by QBhd

One reason why I long ago stopped chasing that dragon. I now use WASAPI Audio Renderer and drop/repeats are always minimum 1-2 hours for all refresh rates. Good enough!

QB

Yeah I would be ok with 1-2 hours. Actually that's what my R9 380 gets without having to touch anything and I was using that without complaint. But then after swapping GPU for 1070 I got a drop every 12 minutes which I tried to ignore but it was too noticeable so I had to do something and here we are.

[x7007]

Originally Posted by huhn

the amd decoder are pathetic use software to even have a chance for proper seeking.
i'm not sure if it is even possible anymore to see the hardware decoder usage on AMD anymore.

the "decoder block" is called UVD you can still see it in GPU-Z but just the frequency.

Do you use the Windows 11 New right click Context or the Classic Full context menus using the Registry? Because I have 2 windows installation which are almost identical, one I use the classic and the other one I use the new and on the new it works 100% of the times even with 0.92.7..

Could this bloody be related? they are almost 100% identical os, I used the same windows backup image to install windows 11 on the new windows with the classic registry. 100% of the other registries and apps are exactly the same.

Fuck sake, it stopped working even with the newest, why the hell it work on my other OS and not on this one.

Do you have PowerShell 7 installed?? I noticed that's the only difference I have on the two OS, one has installed 7.1.5 and the other have 7.2.2, so I uninstalled 7.1.5 and installed 7.2.2 and I see if that's fixed it.

pfff still not..... it stopped working again. what the hell..

EDIT: Make sure you set 10bit or higher on Bit Depth!!!! DON'T USE AUTO!!!! I sat like 1 hour just to understand it works fine with 10bit instead on Auto!!

I also disable everything on Display Modes, but it works the same with it enabled for me at least.

I can't understand why it works perfectly on my other windows 11 installation with same everything !! 100% of the times and it's set to AUTO! with 0.92.7!!!

I use Let Madvr decide. can't understand what the bloody different, same GPU Drivers and everything else.

Update me if it fixed your issue!! Works fine now on my Main OS 100% of the times. I'm finally DONE ( I think it's also related to GPU OVERCLOCK!) with my 6900xt if I go overboard with GPU Clock Frequency even if the game works and I try to play a video it would reset it back to 8bit after I exit the game! even when it worked fine right before it! or if you go too far it would crash and it would work 8bit no until you close explorer or sign off etc. Because that's what I didn't do on my other OS!!! Funny yes!

I also updated Anydesk from 7.0.4 to 7.0.7 and Teamviewer from 15.26 to 15.28.5. I even tried lower LavFilters 75 didn't change anything at the time when Bit Depth was Auto.

[tij]

I have problem with subtitles in 3D movies (frame pack). Specifically 3D movies that have "with right eye first" ... subtitle depth seems totally wrong ... the feeling i get as if video is playing correctly "right eye first" but subtitles still get played "with left eye first" ... looking at subtitles is painful in these cases

Subtitles for movies that have "with left eye first" works perfectly

Can anyone confirm this

PS.

[Siso]

@huhn

Which SM option is the right choice for 120hz - "only if there would be motion judder without it..." it doesn't activate SM, "...or if the display refresh rate is an exact multiple of the movie frame rate" this activate SM, and the last option is "always" which also activates SM. Should I choose "always" or "...or if the display refresh rate is an exact multiple of the movie frame rate"?

[huhn]

do you watch 120 hz video if not just use always.

[Siso]

Originally Posted by huhn

do you watch 120 hz video if not just use always.

I watch mostly 23,976 and 24 fps movies.

[tij]

Originally Posted by tij

I have problem with subtitles in 3D movies (frame pack). Specifically 3D movies that have "with right eye first" ... subtitle depth seems totally wrong ... the feeling i get as if video is playing correctly "right eye first" but subtitles still get played "with left eye first" ... looking at subtitles is painful in these cases

Subtitles for movies that have "with left eye first" works perfectly

Can anyone confirm this

PS.

I was able to consistently reproduce this in Jriver MediaCenter and on MPC-HC running MadVR v0.92.17 and 113.

Upon closer inspection ... its not just subtitles but OSD too

My guess ... for 3D movies that have "with right eye first" ... MadVR swaps not just video frames for eyes ... but rendered overlays too (subtitles and OSD), which it should not as overlays will be brought "inside" the screen causing perception problems

[flossy_cake]

I reckon smooth motion would be an ideal solution to the "1 frame drop every..." problem, if it could be modified to only blend for a few frames either side of the frame drop, and leave the rest of the frames untouched.

eg. if you're playing 23.976 video & have a frame drop every 10 minutes (14385 frames) then smooth motion could blend across, say, frames 14382 to 14388 and leave the rest alone.

If I'm imagining how SM algorithm works, perhaps this could be trivial to implement by something like complementing and rounding its internal blend coefficient.

[huhn]

that would be horrendous

[flossy_cake]

Well the idea is to minimise the amount of frame blending and therefore minimise the amount of blur.

From what I can tell by looking at it, SM is just a simple weighted frame blending based on how "far" the current video frame is between refresh cycles.

Maybe SVP can do it better.

[huhn]

so you fix a dropped repeated frame by still seeing the drop repeated and a blurred frame? to minimise blur you use the highest available refresh rate. 1 frame per input frame is blended that means with 60 HZ you have a blended frame on for ~16 ms but for 120 hz for ~8 ms with other words the higher the refreshrate the lower the blurred frames are on the screen.

sorry but you tried to use it a 24 HZ which should be utterly illegal or with other words only be used to prove that it should not be used like that.

SVP is a frame interpolation algorithm not only has it little to nothing todo with SM it's also in capable of fixing frame repeates/drops because it is just integar based not using the audio clock as reverence and it'S also pretty impossible to motion interpolate a fraction of a ms.

it can't even do 24 to 60 it doing 120 and just throws half the frames away including the originals that's not part of a bad implementation it's just as good as it get's

[flossy_cake]

Well the best solution is to slightly nudge the monitor refresh to sync with audio clock and minimise "1 drop every..." and this is what I'm using, but sadly not reliable across all systems.

120hz & 144hz are divisible by 24 so no gain for SM at removing judder.

For hiding a dropped frame every 10 minutes, it seems that SM would still blend for 9.999 minutes leading up to the drop @ 10 mins, so I'm not sure how higher refresh rates would help minimise SM blur.

I do agree that higher refresh rates should help hide the dropped frame as in the case of 120hz the drop should be a 8ms stutter instead of 16ms stutter for 60hz. But 8ms is still going to be noticeable, and most HDTVs in the wild are 60hz max.

SVP's motion vectors would be better at interpolating between the skipped frame, but I'm not sure it could be configured that way.

[huhn]

wow that's disappointing.

i never talked about how 120 hz shortes dropped frames which are less notable (but need far more because they are only correcting 8 ms if madVR would even work like that which i never tested because why would i)

For hiding a dropped frame every 10 minutes, it seems that SM would still blend for 9.999 minutes leading up to the drop @ 10 mins, so I'm not sure how higher refresh rates would help minimise blur.

it will never reach the dropped or repeated frame because the audio and video are always in perfect sync at all times it does not drift madVR only calculates the numbers for you that's all.
the deviations is correct all the time with no pause.

For hiding a dropped frame every 10 minutes, it seems that SM would still blend for 9.999 minutes leading up to the drop @ 10 mins, so I'm not sure how higher refresh rates would help minimise SM blur.

that's how

1 frame per input frame is blended that means with 60 HZ you have a blended frame on for ~16 ms but for 120 hz for ~8 ms with other words the higher the refreshrate the lower the blurred frames are on the screen.

let me do some math here even through you are most likely going to ignore it like other the other things i have listed:
if you use SM at 24 with a 24p source you will get blended frames with the length of 42 ms every 42 ms
or in short you get garbage on the screen.
makes total sense very simple math.
if you play 24p at 60p you will get 16 ms of blended frames every ~3 frames with means you don't get garbage on the screen. ~1/3 of all displayed frames are blended the rest is original
and now we comes to 120 with 120 hz you get 8 ms of blended frames every 6 frames about 1/6 of all displayed frames are blended for barely any time at all.

I do agree that higher refresh rates should help hide the dropped frame as in the case of 120hz the drop should be a 8ms stutter instead of 16ms stutter for 60hz. But 8ms is still going to be noticeable, and most HDTVs are 60hz max.

the blended frame duration is 8 ms not the dropped repeate frame time...

SVP's motion vectors would be better at interpolating between the skipped frame, but I'm not sure it could be configured that way.

no why would it be it makes zero sense the interval is still interrupted by 42 ms with a drop there are now 42 ms missing and with a repeat 42 are repeated interpolation or not doesn't change that.
i talked about that before.

SVP is a frame interpolation algorithm not only has it little to nothing todo with SM it's also in capable of fixing frame repeates/drops because it is just integar based not using the audio clock as reverence and it'S also pretty impossible to motion interpolate a fraction of a ms.

it can't even do 24 to 60 it doing 120 and just throws half the frames away including the originals that's not part of a bad implementation it's just as good as it get's

[e-t172]

If you want to experiment with various ways of doing frame blending, try out the time interpolation options in mpv (--interpolation, --tscale, etc.), which has lots and lots of various algorithms to try (including one that is similar to madVR Smooth Motion). For me it was a great way to convince myself that the approach madVR uses is probably the best one - other approaches tend to come with very obvious, very obnoxious motion blur.

[flossy_cake]

Originally Posted by huhn

120 with 120 hz you get 8 ms of blended frames every 6 frames about 1/6 of all displayed frames are blended for barely any time at all.

Not if SM weights the blending by how close it is to the next/previous video frame, which is the premise I was writing under.

Going by the link in e-t172's post, it seems that is NOT how SM works -- it only blends 50% with next/prev frame, and no other amount of blending is used other than 50%.

Why do I believe this? I took a dozen screenshots of SM frames (24@60hz) and they all have the same amount of 50% blending like this:



I am not sure why there would be any blended frames with 24@120hz as 120 is evenly divisible by 24.

However my 120hz mode is tuned to 59.94x2 so 23.976 goes evenly, maybe that's why.

[huhn]

i recommend you to read up some math stuff.

it needs 1000 of frames to move the blending around it is not 50/50 a dozen doesn't mean anything and if you don't have a clock deviation it will not move at all.

if you get a frame drop every 40 min that with 24p that's 54600 frames to get it deviate by ~42 ms

seriously just throw 25 at 24 no no no throw 60 at 24 and tell me that it is only blending 50 %
i can't wait for your explanation with "dozen" of screenshoots.

[flossy_cake]

Originally Posted by flossy_cake

Going by the link in e-t172's post, it seems that is NOT how SM works -- it only blends 50% with next/prev frame, and no other amount of blending is used other than 50%.

This screenshot of 30.000fps video displayed at 59.94hz appears to indicate otherwise -- note the grey is darker on the left side than right side -- not a 50% blend.


https://www.youtube.com/watch?v=-Ht8EIAOrj8

Originally Posted by huhn

if you get a frame drop every 40 min that with 24p that's 54600 frames to get it deviate by ~42 ms

That's what I thought too, but then I did a test with 23.976@23.8hz to induce a frame drop every 5 seconds and it was blending 4/5ths of the time. I was expecting it to blend only for a short amount of time.

Originally Posted by huhn

seriously just throw 25 at 24

I tried 24@50hz and it spends majority of time blending too. So SM is definitely weighting it, I just don't understand how the weights work cause with 24@60hz we DON'T get weighting like this:

Refresh#  60hz interval  24fps frame#  24fps interval  Blend weight  Blend formula
1         0ms            1             0ms             0%            (0-0)/41.6
2         16.7ms         1             0ms             40%           (16.7-0)/41.6
3         33.3ms         1             0ms             80%           (33.3-0)/41.6
4         50.0ms         2             41.6ms          20%           (50.0-41.6)/41.6 
5         66.6ms         2             41.6ms          60%           (66.6-41.6)/41.6
6         83.3ms         3             83.3ms          0%            (83.3-83.3)/41.6	
7         100.0ms        3             83.3ms          40%           (100.0-83.3)/41.6	
8         116.6ms        3             83.3ms          80%           (116.6-83.3)/41.6	
9         133.3ms        4             125.0ms         20%           (133.3-125.0)/41.6	
10        150.0ms        4             125.0ms         60%           (150.0-125.0)/41.6	
11        166.6ms        5             166.6ms         0%            (166.6-166.6)/41.6

Maybe Madshi wrote an exception to the rule for 24@60 since limiting it to 50% weighting in this mode reduces the total amount of blending by quite a lot. Or maybe his algorithm just naturally determines the best way to minimise total blending.

[huhn]

yes he wrote that it fixes deviation by simulating a unlimited frame rate screen.

but more important do these files even have audio?

it's "simply" (not easy to programm) vsync time changed by audio clock (42 ms treated as 43 ms)
now you take the 43 ms and look what is displayed at it for example frame 1 for 8 ms and frame 2 for 35 so you blend in linear light 8/43 frame one and 35/43 and present it(not in linear light).

[flossy_cake]

Originally Posted by huhn

but more important do these files even have audio?

The pendulum one I didn't bother remuxing an audio track because I'm not using it to minimise drops, only to induce drops so I can try to understand how SM algorithm works in different scenarios.

Originally Posted by huhn

now you take the 43 ms and look what is displayed at it for example frame 1 for 8 ms and frame 2 for 35 so you blend in linear light 8/43 frame one and 35/43 and present it(not in linear light).

Yes but I'm not sure that is how SM works @120hz because 24 goes evenly 5:5 into 120hz & I took screenshots and couldn't find any blended frames.

[flossy_cake]

Originally Posted by Asmodian

You don't want to use smooth motion when the source frame rate almost equals the display's refresh rate

This is consistent with what I'm observing, so SM is not good at minimising "1 frame drop every n minutes" but in theory I think it could be if its algorithm was tweaked a little.

Originally Posted by Asmodian

but it is great when the display's refresh rate is much higher

This is consistent with huhns comments, but it's not quite correct because 24@120 doesn't appear to have any blended frames given the 5:5 cadence. So cadence needs to be unsmooth, then yes higher refresh will give better results.

But higher refresh will give better results anyway even with SM off, since the rendered frame will "land" closer to a refresh cycle as there are more refreshes available. An easy way to think about this is a hypothetical 1000hz display (refresh every 1ms) in which max stutter will be 1ms regardless of the video fps since the longest wait till next refresh cycle is only 1ms, and a 1ms stutter is probably invisible.