madVR - high quality video renderer (GPU assisted)

[anta777]

How to test the speed of a graphics card for an algorithm NGU sharp ?
Is interested in NVIDIA 1070-1080 vs AMD RX 570-580.

[Asmodian]

The 1080 is faster (and 2x the price). Polaris doesn't like NGU very much.

[anta777]

How many times faster?
2x as price or less ?

[Asmodian]

Much less than twice... though I am not sure as to the exact ratio. I haven't seen any good benchmarks, only anecdotes from this thread.

Edit: Oops, it is probably around twice as fast.

[huhn]

it's not easy to do. boost clocks are unstable other algorithm are using the GPU too.

a 1060 is about 3-5 times as fast using NGU as a 480.
i wouldn't be shocked if a 480 is generally fast when NGU isn't used.

[pose]

Originally Posted by huhn

a 1060 is about 3-5 times as fast using NGU as a 480.

I doubt it.

[huhn]

i had both cards what's your reason for the doubt?

just as a reference i can do NGU sharp very high normal chroma with my 1060 when scaling 1080p23 -> 2160p23.
compare that. high is already a problem for the RX 480.

there goes something terrible wrong when using NGU on polaris cards. my r9 270 was totally fine using NGU.

[pose]

Originally Posted by huhn

i can do NGU sharp very high normal chroma with my 1060 when scaling 1080p23 -> 2160p23.

Really? What are your render times? Can you post OSD screen?

Originally Posted by huhn

high is already a problem for the RX 480.

My RX 470 can do NGU Sharp High and has a lot of headroom left. Test results below.

1080p23 -> 2160p23. All settings identical except for Luma:

NGU Sharp Very High - 69.5ms (SM ON); 51ms (SM OFF)
NGU Sharp High - 24.7ms (SM ON); 22.8ms (SM OFF)
NGU Sharp Medium - 15.5ms (SM ON); 13.5ms (SM OFF)

NNEDI3 256 - 89ms (SM ON); 83.2ms (SM OFF)
NNEDI3 128 - 50.5ms (SM ON); 47.3ms (SM OFF)
NNEDI3 64 - 31.6ms (SM ON); 29.6ms (SM OFF)

I included smooth motion results because with NGU Very High performance hit was very hard for some reason, thought it was interesting.

[huhn]

sure https://abload.de/img/nguveryhigh2zcjho.png

way better result i'm used to.
make me wonder if a driver update is the reason because my finding back in 2016 early 2017 where confirmed by other user and yours are not even close to it.
i wonder why noone else is noticing such a giant permanence increase and reportet it.

that's only about 1.5-2x.

BTW. i was using SM back in the days.

i saw stuff like 150 ms with my 480 back in the days (ngu was used on chroma doubling too).

[pose]

huhn, what kinda rendering times youre getting with NGU Sharp High 1080p23 -> 2160p23?

[huhn]

NGU very high
in FSE ~32 AVG and ~35 ms max.
but i'm not a huge friend of rendertimes i get rendertiems of 39 ms with super xbr because the cards goes for a nap.
with advanced rendertimes NGU takes ~29 ms. it is hard to read because it is called finalsteps there right before smoothmotion.

GPU usages is abut 90% and the AVG boost clock is 1700
i use other filter at the same time.

in short
NGU very high
32 ms
NGU takes about 29 ms
GPU load 90 %
boost clock 1700

NGU high
16-17 ms
NGU take about 11 ms (a ~3rd like it should be)
GPU load is 45 %
boost clock is 1900 (don't think about this)

[pose]

huhn, impressive. What about NNEDI3? I wonder if its also faster on Pascal.

[Asmodian]

Interesting, it does seem like NGU has improved on Polaris.

My Pascal Titan X doesn't do that much better at NGU Sharp very high 1080p -> 2160p.

These are the rendering times for the NGU step by itself (ShowRenderSteps). If you disable subtitles and smooth motion you can see it reported by itself, the times weren't quite the same when it was mixed into another line. I used chroma doubling set to normal (Bicubic 60). madVR v0.91.9 of course, DX11 fullscreen windowed, Win10 w/ creators update, Nvidia driver version 382.05. I added image enhancements and changed chroma upscaling to keep the GPU clocks up. The GPU was at 1860 MHz unless otherwise noted.

NGU Sharp (low): 1.7 ms
NGU Sharp (med): 2.6 ms
NGU Sharp (high): 5.2 ms
NGU Sharp (very high): 16.8 ms

edit:
I remember NNEDI3 differences being interesting too:
NNEDI3 32: 4.18 ms
NNEDI3 64: 7.54 ms
NNEDI3 128: 18.83 ms
NNEDI3 256: 32.93 ms (1835 MHz)

[madshi]

Originally Posted by ryrynz

Madshi, the .ax files are quite a bit larger than in previous releases, zip is now 10MB larger.
Not that it's a big deal but, what gives?

Not sure why that happened. In any case, it will be fixed in the next build.

Originally Posted by Flokke

Does anyone know why no matter what I do my MadVR Control Panel will not show up.

Seems weird. You sure it's not in the list of "hidden" tray icons? (The OS sometimes hides some tray icons.) The tray icon is handled by madHcCtrl.exe, that process is running?

Originally Posted by akim

> Why are you worried about memory consumption or short term chip/memory frequency?
> These things are not really very important.

Looking through movies and different clips, I can constantly switch to full screen mode and back and then it turns out that the graphics card works for a long time in maximum mode
Forgot to say that the frequency of the chip and memory is growing and when rewinding the video
This is not what I would like to get from soft decoding
I have a gluttonous 130 watt graphics card
And for the processor, the video formats sd and full hd are not even noticeable if the program is competently optimized

SD format(default setting):
Mplayer(frontend Smplayer) Cpu 1-2% memory 35mb (video card is resting)
Mpc-hc lav-filters madvr Cpu 2-7 memory 380 mb (increase in the frequency of the video card under certain conditions)

Your post doesn't really answer my question. I asked why you're worried about memory consumption and chip/memory frequency. I see zero reason to worry about memory consumption, as long as you don't have some other application running at the same time which needs the memory. Consuming more memory doesn't cost any electricity.

I can understand if you want to save power, and have your GPU fan be silent. So I can understand that you want low chip/memory frequency *during normal video playback*. However, frequencies being temporarily higher after changing the window size or after seeking is totally irrelevant, IMHO. I don't see why that would matter in any way.

Originally Posted by dipje

I absolutely understand how that pro pricing works, I'm just wondering what's going to be missing from the pro version then. But I'm guessing you don't know for sure yourself yet .

I see a lot of tools where everything is available, but then only in 8 bit for instance. And it's frustrating for me that I would spend close to a $100 no problem to get some very good algo in 10bit or 16bit support, but if the pro license then jumps up to $500 and more it's a shame (for me). So close yet so far :P. Like I said, understandable though.

Having access to things like special chroma upscaling and NGU stuff (gpu accelerated all?) would be awesome. Upscaling stabilised footage back up to it's original res with NGU or being able to process and mix some of my 1080p footage with my 4k footage (ngu upscaling + debanding + knlmeanscl) sounds like a lot of good video toys to be testing.

Of course if someone asked me if I really needed 10bit+ they probably have a point. Really needing something really wanting something are not the same but related :P.

Actually limiting the input to 8bit is currently the most probable option, because for most "hobby" users 10bit input support is not crucial, while for pro users 10bit support is probably quite important. If you have a better idea, I'm all ears. There does need to be some sort of limitation, or else I'd lose many sales. Ideally the limitation should hurt only pro people but not hobby users.

Originally Posted by qjb2016

Hi madshi,I also,the HDR, color space can be identified, the transfer function can not be identified, windows 10 1703, epson tw8300

I'm not sure what you mean. Unfortunately your attached image is not visible to me. Attaching stuff to this forum rarely works, so it's better you upload it to some external image host and then link to it here.

Originally Posted by JarrettH

madshi, since I can't get the ytp test images to load anymore, can I just as well drag one of the png test images from http://www.lagom.nl/lcd-test/ into MPC-HC? The gradient one specifically. Thanks

I'm not familiar with those png test images, so I can't comment on that. The ytp test patterns currently only work with 32bit media players.

Originally Posted by ryrynz

Ran NGU sharp for the first time on some a 720 movie and was quite impressed with how it looked. Couldn't run Very High on my 960 but High was a go with about 20ms render time.

Glad to hear that!

Originally Posted by igvk

I have created .3dlut in displayCal for my display (NEC PA302W), using i1 Display 2 calibrator.
And recently I noticed some bad artifacts in video when this 3dlut is used in madVR.
Here is a screenshot (the saturated color patches can be seen, especially in the left part of the image).

The video is ok when i disable calibration LUT.
It seems that something is wrong with calibration or its usage.
What can be done to fix this?

Can't open your image, unfortunately. Please ask DisplayCAL related questions here:

http://www.avsforum.com/forum/139-di...argyllcms.html

Originally Posted by IanD

I'm attempting to play 3D @ 60fps to an LG OLED in PC mode using LAV, MPC-BE and MadVR in order to bypass as much processing in the LG as possible and have greater control.

Currently 3D rendering is optimized for 1080p displays. I'll add to my to do list to look into offering a line alternative 4K mode. But I don't know when I'll get to that, could take a while.

Originally Posted by IanD

As a preliminary question, I have come across comments that MadVR only works with MPC-BE x86 and not x64: is this still true, or doesn't it have an impact on speed anyway?

madVR has supported x64 for a long time now.

Originally Posted by IanD

I'm also confused over having 3D output format settings in both MPC-BE and MadVR: what should I set each one to in order to get unmolested line interleaved 3D with the least processing overhead?

I'm not sure what those settings in MPC-BE do, haven't tested that yet. Maybe they only affect MPC-BE's internal renderers? Or maybe they overwrite the madVR settings? I don't know, that's a question you should ask the MPC-BE devs. Generally, the media player doesn't even have to know you're playing 3D. LAV and madVR do all the dirty work. So IMHO you can probably leave all MPC-BE options at default settings (or turned "off" for 3D, if there is such a setting).

Originally Posted by chileverde

I've been using MadVR (currently v0.90.13) with Kodi DSPlayer 16.1 on my Windows 10 HTPC for a long time with no problems, but just recently started experiencing video/audio stutters two or three times per hour of playback. The best way I can describe it is like a skipping vinyl record, where the same millisecond of audio/video repeats in a rapid-fire machine gun fashion for 1-2 seconds. Sometimes the video/audio are out of sync afterwards and I have to stop and restart playback.

Interestingly, the MadVR playback stats don't recognize these stutters as frame repeats or presentation glitches - it seems like as far as MadVR is concerned, nothing is awry, which makes me wonder if something else in my system is causing the problem.

Any advice on how to troubleshoot this?

"recently started"? What has changed? Did you change drivers? Or madVR version? Or did you install Windows 10 to Creator's Update?

Are you using fullscreen exclusive mode?

Originally Posted by Frexxia

I'm encountering some very strange behavior if I disable exclusive mode:

If I seek in the video and keep the seek bar active I'm getting comparable render times to exclusive mode
http://i.imgur.com/64UvGhM.jpg

As soon as the seek bar becomes inactive and madvr changes from D3D9 windowed to D3D9 windowed fullscreen I'm getting very high render times
http://i.imgur.com/MpJU5u0.jpg

But it isn't only related to D3D9 windowed fullscreen, because if I activate the seek bar without actually seeking I'm still getting very high render times
http://i.imgur.com/3nc7OAN.jpg

Edit:
Immediately after posting this I discovered that enabling the option "use Direct3D 11 for presentation" seems to be a solution for my problem.

I've also received some crash reports from various users (mainly NVidia) which suggest that D3D9 is very unstable on NVidia atm in Windows 10. For most of those users it seems that activating the "use D3D11 for presentation" option helps.

Originally Posted by oddball

Since [...] MadVR takes ages to load a video now

It shouldn't. It certainly doesn't for me.

Originally Posted by ipanema

I'm just starting to try madVR in graphstudionext. I'm connecting the LAV splitter and decoder to it. I'm using Windows 7.

When the input pin of madVR is connected, a dialog box appears that says:

File Open - Security Warning
Do you want to run this file ?
...\madHcCtrl.exe

with Run and Cancel buttons.

The madHcCtrl.exe file seems to be the configuration gui for madVR. If I click "Run" to allow it to run, nothing seems to happen - the gui window does not appear.

It seems a bit strange that madVR tries to run the configuration gui just because the input pin is being connected - is this normal behaviour ?

It seems a bit pointless, especially when the gui doesn't even appear.

madHcCtrl's main purpose is to handle the madVR tray icon and display the madVR settings dialog. In certain situations madVR also starts additional madHcCtrl instances which then have a special job, like watching over the media player and restoring the original display mode in case the media player crashes, or things like that.

[madshi]

madVR v0.91.10 released

http://madshi.net/madVR.zip

* improved Windows 10 HDR metadata support
* added new option "hack Direct3D to make 24.000Hz and 60.000Hz work"
* performance improvement with "trust DXVA" + "use D3D11 for presentation"
* fixed: Intel GPU driver flickering bug with "use D3D11 for presentation"
* fixed: madVR(64).ax files were larger than necessary
* Intel: 10bit DXVA surfaces are now converted to 10bit RGB (instead of 8bit)

1) I hope switching your display into HDR mode works now? Currently only NVidia and AMD drivers support this, Intel not yet.

2) There was a long standing bug with Intel GPUs where activating the "use D3D11 for presentation" option could result in video frames jumping back and forth. I was finally able to reproduce the bug on my new Kaby Lake test PC. It turned out to be a GPU driver bug (as usual). I've found a workaround, so the problem should hopefully be fixed now.

3) I've had a hack in madVR which hooked deep into D3D9 and D3D11 to make sure that 24.000Hz and 60.000Hz display modes were correctly used in Windows 8.1 and Windows 10. However, I've found that this hack can sometimes lead to presentation queues not filling properly. So this hack is now optional (and disabled by default). Without the hack, you will probably always get 23.976Hz and 59.940Hz, even if you want 24.000Hz and 60.000Hz. With the hack, you will probably get proper refresh rates, but may run into issues with the presentation queue not filling. Pick your poison...

[ryrynz]

Solid maintenance release. Glad to see that Intel bug finally "fixed".
Please, anyone posting after this, don't quote madshi's entire post..

[imhh11]

Originally Posted by madshi

madVR v0.91.10 released


1) I hope switching your display into HDR mode works now? Currently only NVidia and AMD drivers support this, Intel not yet.

Thanks for the new release.
Does this mean we dont have to manually switch the windows hdr slider anymore ?
if so, it didnt work on my system.

[igvk]

Originally Posted by madshi

Can't open your image, unfortunately. Please ask DisplayCAL related questions here:

http://www.avsforum.com/forum/139-di...argyllcms.html

Here is another example:


I think, it could be due to whiter-than-white values or something like that in the original video. But all that leads to unpleasant artifacts in the video.
Is it possible to still convert these values for them to not be out of sRGB gamut?

[arrgh]

Originally Posted by madshi

madVR v0.91.10 released

...
* fixed: Intel GPU driver flickering bug with "use D3D11 for presentation"
...

2) There was a long standing bug with Intel GPUs where activating the "use D3D11 for presentation" option could result in video frames jumping back and forth. I was finally able to reproduce the bug on my new Kaby Lake test PC. It turned out to be a GPU driver bug (as usual). I've found a workaround, so the problem should hopefully be fixed now.

excellent... works...

[Asmodian]

Originally Posted by igvk

Is it possible to still convert these values for them to not be out of sRGB gamut?

If you use the rendering -> general settings -> "enable windowed overlay (Windows 7 and newer)" option, and don't have exclusive mode or DX11 enabled, WtW values are clipped to in gamut.