madVR - high quality video renderer (GPU assisted) [Archive] - Page 1142

fourbigkids

9th August 2019, 06:41

Hi all.

I am running MPC-BE v.1.5.3 with madvr 127.0.0.1 on a Windows 10 HTPC with an Nvidia RTX 2060. The video is fed via HDMI to a pre-pro and then to a Sony TV XBR85X850F. It can also be fed directly to the TV. I have madvr setup to open full screen and to change the refresh rate of the TV to match the video file. This works fine about 50% of the time. For the other 50% the video opens and plays as expected, full screen and with the adjusted refresh rate, but the menu bar (WITH THE PLAYER OPERATING BUTTONS) is not available when placing the cursor at the bottom of the screen and right clicking does not open the context menu. There is no way to control playback when this occurs. However, if I double click on the video on screen to 'restore down', taking the application out of fullscreen mode, the menu bar and context menus return. Once the fault occurs it then keeps happening on each successive file until I reboot the computer. It then plays as expected for a random period of time and/or playback of a random number of files.

There does not seem to be any common characteristics in the files that this fault occurs. Could be anything, an .avi, .mp4, or mkv etc. But next time they could play faultlessly.

Has anyone encountered this sort of behaviour before and managed to resolve it?

Appreciate your advice.

Cheers,

Allan.

ps. Could you please let me know if this is better dealt with in another thread. a

Hi all. This problem is still plaguing me. It only seems to occur when madvr is selected as the video renderer in MPC-BE, not for any of the others. I was thinking of reinstalling madvr but cannot find a way of backing up madvr's settings before doing so. madvr is in a folder inside the MPC-BE folder but there is no settings.bin file in that folder. I have show hidden files selected. Is there any other way to backup madvr's settings? Thanks for the help. Cheers. Allan

chros

9th August 2019, 09:53

Maybe the TV does overscan, hence you don't see the controls. Or controls are hidden in MPC-BE (I think in View menu).

The video is fed via HDMI to a pre-pro and then to a Sony TV XBR85X850F. It can also be fed directly to the TV

Then connect directly to the TV and try out that way.

ryrynz

9th August 2019, 10:28

but cannot find a way of backing up madvr's settings before doing so. madvr is in a folder inside the MPC-BE folder but there is no settings.bin file in that folder.

Because permissions aren't set to allow the writing of the file.
Give the madVR folder write permissions for Users, settings.bin will be written next time you make changes, alternatively back up \HKEY_CURRENT_USER\Software\madshi

okc_smoker

10th August 2019, 12:51

There was some talk a couple weeks ago about some sort of benchmarking protocol and a database of results for various GPUs. Is that still in the works?

Sent from my SM-G965U using Tapatalk

chros

10th August 2019, 19:29

There was some talk a couple weeks ago about some sort of benchmarking protocol and a database of results for various GPUs. Is that still in the works?
No one started it, you can be the One :)

okc_smoker

10th August 2019, 23:17

No one started it, you can be the One :)

Lol! As much as I'd love to, I don't really have the time to take that on at this point. I did just switch my AMD RX 560 for an RX 5700 (not XT), though. I'd be happy to run any tests or provide data points for various settings. I will say that, so far, the 5700 is blowing me away. I can run pretty much whatever MadVR settings I want and I barely break 30 fps (rendering). I'm happy to be a test subject for Navi if anyone is interested.

Asmodian

11th August 2019, 00:03

I was thinking of attempting something...

What about the idea of a 4K video that tries to run at high fps? H.264 fast-decode so it is likely it can be decoded fast enough on most newer systems. Record the number of dropped frames averaged over three runs. It is hard to come up with a good benchmark that doesn't require a lot of steps and/or is prone to user error. Any manual recording of MHz + ms numbers with a collection of standard videos will be pretty hard to do well. Even this idea of manually reporting dropped frames would be tricky to accurately reproduce from user to user but it would push the GPU to the max and only require one test video.

edit: probably better to use a 1080p video upscaling to 4K instead, to actually test image upscaling which is the most important place to spend GPU power. :o

huhn

11th August 2019, 00:29

the best i found is:
1080p23 supersampled
advanced render stats
pushing it as far as possible before dropping frames with the chroma option you have quite some options with that.
if that's not enough we can add a better tier later with 30-60 FPS

now you make a screenshoot with GPU-Z and the OSD.
take the MS for the doubling and the clock / GPU load and put it into the list sorted after settings followed by ms, clock or so.

is this perfect no without a question but it give user real world examples.

something similar could be done with HDR but that should be ignored until it is finish some day some day for sure.

okc_smoker

11th August 2019, 00:51

Asmodian

11th August 2019, 01:10

The problem with more comprehensive methods is that they do not scale well with GPU power, especially if we do not want to drop frames. A 1650 to 2080 Ti is a very wide range, and that is only Nvidia's current GPUs. With AMD and old GPUs collating and understanding a comprehensive data set seems like a lot of manual work. Without hooks into madVR and some software automation I think that kind of data set is too ambitious.

I think just the frame drop rate doing something too hard might be a pretty good judge of madVR performance in more realistic workloads, at least enough to test.

I agree that a settings.bin file is required along with a specific test video or set of videos. We might need two settings to cover every GPU and/or output resolution.

huhn

11th August 2019, 01:41

drop frames is like the most unreliable measurement you can do. as soon as madVR drops more then a couple it start repeating and with that it may start dropping then drops more and more totally unreliable between runs.

GPU doesn't got much faster in the past years a 2080 ti is not even close 5x as fast as a 1650.

if someone is unable to figure what is the maximum setting this card can read why would you want a his data?

okc_smoker

11th August 2019, 02:19

mpchelp10

11th August 2019, 04:13

MPC-HC 1.8.7
madVR - not sure which version but I installed via K-Lite Codec Pack
Windows 10 Laptop with 4k HDR (fake HDR, 8+2) screen
iGPU: Intel HD 630
dGPU: GTX 1050TI with Max-Q
NVIDIA settings shows that my laptop's screen is physically connected to iGPU

Files I am playing:
-Your name in 4K HDR
-Sony 4k HDR sword demo (https://4kmedia.org/sony-swordsmith-hdr-uhd-4k-demo/)

LAV settings: DXVA2 - Intel HD 630, no option (greyed out) to use NVIDIA dGPU
MadVR settings: Passthrough HDR to display, send HDR metadata to display

With these settings, I can "see the HDR" on both files. Your name more or less plays fine, but the Sony sword demo will visibly lag or be "slow".

When I try to get MPC-HC to use my GPU, I lose the HDR colors. I tried the following:
1. DXVA2 (native) + NVIDIA Control Panel Program Settings > MPC-HC > High-performance NVIDIA processor. After this, DXVA2 shows that it is using the dGPU
2. Change DXVA2 (native) to NVIDIA CUVID
Both of these options result in MPC-HC no longer showing "HDR colors".

How do I fix this?

chros

11th August 2019, 06:03

I don't claim to be an expert when it comes to video processing, but I do have 20+ years as a software engineer. Any data we can gather is going to be wildly speculative, but if we can get a broad enough sample, I would think it would still be useful. Certainly better than what we have now (i.e. nothing). The key is to eliminate as many variables as possible. Using specific videos and specific MadVR settings seems to me to be as close as we can get to a controlled study. I would think it would be worthwhile, at the very least...
I absolutely agree with you, exactly that's what I told them as well.
The problem is that something has been changed, not just in real world, but here as well, only a few is willing to do something. :) Especially when it comes to time consuming tasks.

Anyway, maybe I quickly put something together today in a new thread.
Which is the best free hosting service with direct links to files, where we can store 10GB data? (like dropbox, etc)

cork_OS

11th August 2019, 07:55

Using specific videos and specific MadVR settings seems to me to be as close as we can get to a controlled study. I would think it would be worthwhile, at the very least...
madVR upscaling speed doesn't depend on a specific video. And that's right and fine. Unfortunately, I can't say the same about HDR processing.

chros

11th August 2019, 09:26

madVR upscaling speed doesn't depend on a specific video.
It does!

chros

11th August 2019, 12:03

I would think the ideal would be to provide some sample videos and some MadVR config files and have users run the sample videos and report the results with the provided configurations. That seems like the best way to get reliable data.

There you go: madVR 4K GPU performance test (https://forum.doom9.org/showthread.php?p=1881568#post1881568)
And I already posted the first result set :)
Enjoy!

Asmodian

11th August 2019, 13:18

drop frames is like the most unreliable measurement you can do. as soon as madVR drops more then a couple it start repeating and with that it may start dropping then drops more and more totally unreliable between runs.

It isn't super reproducible but it does give an idea. I measured a 5.6% standard deviation over 10 repeats when dropping an average of 11.1% (10.5% to 12.6%) of the frames in a 240 fps 1080p encode of big buck bunny. I did have to mess with settings to get it even that reproducible.

That said, I am not sure how useful a "max settings" data set would be either. A madVR benchmark score is not a list of max settings for every GPU. Settings are too flexible and complex for that to be super useful.

okc_smoker

11th August 2019, 14:45

That said, I am not sure how useful a "max settings" data set would be either. A madVR benchmark score is not a list of max settings for every GPU. Settings are too flexible and complex for that to be super useful.

Max settings or not, if we get enough people running the same videos with the same settings, that would still be useful.

Maybe 1080p -> 4k with image scaling using NGU Low, Medium, High, etc. We could include whatever settings we want if there were standardized configs that users could just download and run.

ng4ever

11th August 2019, 14:54

Stupid noob question here sorry.

Apple TV with Infuse Pro or HTPC with madVR ?

Which one would be better, have higher picture quality, and more reliable + supported ? I know which one is cheaper.

Thanks!

ng4ever

11th August 2019, 16:27

Also which mid range video card would be best for madVR and what would be the highest settings I could use ?

I want to get a RTX video card but keep hearing about glitches, stuttering, or something with them and madVR :(

huhn

11th August 2019, 19:48

It isn't super reproducible but it does give an idea. I measured a 5.6% standard deviation over 10 repeats when dropping an average of 11.1% (10.5% to 12.6%) of the frames in a 240 fps 1080p encode of big buck bunny. I did have to mess with settings to get it even that reproducible.
12% is not enough before madVR starts "batch" dropping at some point madVR even runs async.
the next problem with 240 hz is decoding may have an impact and the overhead just for final steps is gigantic.
That said, I am not sure how useful a "max settings" data set would be either. A madVR benchmark score is not a list of max settings for every GPU. Settings are too flexible and complex for that to be super useful.

using my method gives real a world example what a GPU is capable of and using the ms to sort them gives a general list what GPU is better which is only useful for to measuring it between different GPU types everyone knows a 1660 is better then a 1650.

i didn't start a test like this because i think it is not good enough.

QBhd

11th August 2019, 22:06

I think everyone is making this "benchmark"/comparison for too complicated. We want a simple procedure to give us a simple number. If GPU "A" can perform a standardized playback with a GPU usage of 60%, and GPU "B" performs the same task with a usage of 96%, is it not safe to deduce that GPU "A" performs better with madVR? Don't most players have an OSD which shows various info, including GPU usage (in PotPlayer it's the TAB key)? Also a GPU should be able to handle at least 90% usage before frames drop (I can push mine to 96-97%), so we should be able to have a nice wide range of values to sort GPU's by. If your GPU usage is bouncing around all over the place (by more than a few % points) wouldn't that indicate an improper setup?

I'm just trying to simplify things so that questions like the one by nb4ever can easily be answered if we get enough info gathered.

QB

huhn

11th August 2019, 23:34

that the reason i only do one test video but GPU usage is irrelevant GPU have no many parts that a not used when using madVR making this metric bad the new boost behaviour doesn't make it better.

Ceremony

12th August 2019, 00:28

Hi everyone,

Development of a dark theme for MPC-HC is almost finished. Please test it and let me know if you see any visual bugs, or bugs when interacting with the player GUI.

test build (x64) (https://github.com/adipose/mpc-hc/raw/develop/testbuild64/mpc-hc64.7z)
test build (x86) (https://github.com/adipose/mpc-hc/raw/develop/testbuild/mpc-hc.7z)

Dark theme can be activated through: menu > view > dark theme

Please test both the dark theme and the normal theme.
found a bug:
entering a list with checkboxes, such as the internal filter list, and going through that list with the arrow keys, deactivating/enabling (spacebar) them, crashes the player. just toggle a few in a row (spacebar, arrow down, spacebar, arrow down, etc) and it crashes with a memory access violation. this is not the case when toggling them using the mouse.

P.S. happens with both themes

Asmodian

12th August 2019, 02:04

I think everyone is making this "benchmark"/comparison for too complicated.

I agree! However, your idea is impossible due to GPU boost so we need to come up with something easy to do reasonably accurately that provides a numeric score we can easily compare. Getting my 2080 Ti to run at full speed will definitely cause dropped frames on a 1650.

I am very familiar with madVR's options and the NGU quality levels and even I don't get a sense of how fast chros'es 1060 is compared to my 2080 Ti based on that report. Comparing GPUs with pages of those kind of reports is not easy so we really need a better test method.

ksio89

12th August 2019, 02:08

Nevermind, already found the solution to my question.

QBhd

12th August 2019, 03:13

I need a little help. I have a R9 270X that is still kicking (even after it died and I baked it in the oven to bring it back to life), It was more than enough when I had my 1024x768 rectangular pixel "720p" Plasma. When That TV died and I replaced it with a 55" LG C8, I tried playing 4k HDR though the PC, but it didn't work and I just chalked it up to a GPU with only 2GB RAM and HDMI 1.4 so since then I played the 4k HDR content using the TV's built in player.

But today I was farting around with a 4k HDR Blu-ray and it worked on the PC. I can even do HDR.... It took a little tweaking to work in super high bitrate scenes. I cranked CPU and GPU queues to the max (48/24) which makes no sense since queues that big make normal playback of 1080p or less a choppy mess, but the 4k HDR plays like silk (and turns into a choppy mess with 16/8 queue default queue depths) plus queues that big should cause serious RAM shortages shouldn't they? This card has just 2GB... it's odd, but I digress.

Now that I realize can "tone map HDR using pixel shaders", I am curious what should I set "target peak nits" to... I read the C8 can do 820 nits, but I don't have a meter to verify. I know there are a few C8 users here, just trying to pick your brains :)

QB

Asmodian

12th August 2019, 06:14

12% is not enough before madVR starts "batch" dropping at some point madVR even runs async.

I am not sure this is true with all setting combinations, I get similar results with NGU high that results in 38.7% frames dropped. Standard deviation for 10 runs was only 217 (1.43%).

I have a proposal, I wonder if it gives reproducible results? There is a lot of noise in the results on my system so I think we need to run it at least three times and average the results, 10 or more would be even better.

Set your display refresh rate to 23 Hz
Reset your madVR settings and use this settings.bin (http://www.mediafire.com/file/k9gy2np6kmch501/240HzBenchmark1.zip/file). Do not change any settings.
Use this video (http://www.mediafire.com/file/9gpknwhmzkmtpaz/Big_Buck_Bunny_Demo_1080p240_fastdecode.mkv/file)
Set LAV to DXVA2 Native for H.264, or software decoding if you cannot get DXVA2 native to work
View the video at 1920x1080, either fullscreen or windowed depending on your native resolution
On startup pause and seek to the start of the video
Use ctrl-R to reset the stats
Unpause
When it ends use ctrl-j to view the number of frame dropped
Repeat as many times as you can and average the results

Make sure your player does nothing when the video ends, just leaves the last frame up paused. I suggest MPC-HC or Zoom Player.

A check to make sure it is working correctly is that the upload queue is at least 1-8 for the entire playback. Watching with the OSD open seems to give the same results. Make sure you do not have any hidden file/folder options enabled as the settings file cannot change those.

My 2080 Ti drops an average of 4584 frames with a standard deviation of 265 (5.8%) over 10 repeats.

huhn

12th August 2019, 12:05

1. that limits the number of test system quite a lot

6. stop pause wait til the queues are full should be more reliable.

9. i had the OSD on

10. to time demanding for a benchmark and maybe not needed.

i run it 2 times at 60 hz and got very reliable results from this both runs where pretty much spot on 33000 dropped frames.
is forcing SM the trick so it doesn't go "mad"?
my upload queue survived for some reason with 2-8.

i had ShowRenderSteps active but i didn't run this for the numbers now more to see if something is odd.
SM had 0 ms so madVR is skipping it the same for one final step.

just to be sure that BBB 60 at 4x speed so around 38160 frames so
not even 14 % survived in my case so about 30 MS per rendered frame?

DXVA native may cost a lot of processing power here. at 23p these settings are a joke and that's only a difference of 10 time which is matching so this worked for me with these settings.
this is far better then i expected.

HillieSan

12th August 2019, 19:57

I have a proposal, I wonder if it gives reproducible results? There is a lot of noise in the results on my system so I think we need to run it at least three times and average the results, 10 or more would be even better.

...

My 2080 Ti drops an average of 4584 frames with a standard deviation of 265 (5.8%) over 10 repeats.

I have been reading the discussion about benchmarking madvr for the various graphics cards. I think I am not interested in such a benchmark, because it is useless in my opinion. The results tell me little. I cannot base by choice on this.

I am interested in how a graphics card (example an AMD 5700 XT or a RTC 2070 Super) performs at its BEST when scaling 2K to 4K. What are the settings for the best possible video quality? What are the experiences? Different cards end up with different settings but perhaps with almost equivalent quality. Comparing these settings (and parameters) would be more useful. The cards can be ranked by medium, high and very high quality with or without certain features. The maturity of the drivers also need to be considered, like are 3D or HDR supported. Price is also an important parameter.

My AMD RX 480 is default a medium performer, but with specific settings it becomes a high performer. Very high performance (very high NGU) is too much for this card, but the difference between high NGU and very high NGU is very small on my projector screen. Thus far, I am satisfied with high NGU and the video quality. However, many features had to be disabled and I cannot crank up the quality with this card. Doubling fps and smooth motion would further improve the video quality, but this card cannot do this for high NGU and up; e.g. in the past BlueSky improved the video quality quite a lot. This is why I am interested in a new graphics card, such as the AMD 5700 XT, to find out if the video quality can be significantly improved.

I am more interest in what a particular graphics card is up to with madvr.

huhn

12th August 2019, 20:06

that the point of his test. because there are to many combination of setting a GPU that has a better ranking will be able to do more settings in general.

Asmodian

12th August 2019, 20:27

I am simply not interested in putting together that kind of benchmark/database of user reports. There are way too many variables and what is important changes from person to person. Can we do both adaptive sharpen and thin edges while doing 720p -> 4K with NGU 2x very high and 2x high? What about with NGU direct 4x very high? How about with SSIM100 downscaling or Lanczos3?

It is simply never ending.

My benchmark is simply a "madVR performance score". Once we have enough data we could say what score you need to do 1080p24 -> 2160p with NGU high "with some headroom", or any setting we are interested in. The same with adding BlueSky or similar. If users report their experiance with a particular GPU and a particular configuration we could compare the performance score of their GPU with another no one has posted anything about to decide if it is likely to work.

Of course, my benchmark is mostly a NGU benchmark (NGU low and medium to be exact), we don't know if the score would compare well for other settings too... honestly we really don't have any good data sets now. All these reports like, "I use NGU very high for 1080p24 -> 2160p60, all artifact removal disabled, ordered dithering, smooth motion, and a 3DLUT on my 2080 Ti" aren't very useful. If the benchmark does not test this combination for a pass/fail it isn't useful for me because it doesn't test my use case and it is really hard to use a big collection of these to come up with something useful, not to mention the effort involved in collecting and presenting it all. There is a reason popular benchmarks always return a single numeric score.

ng4ever

12th August 2019, 20:40

Is it true Nvidia RTX cards have a freezing issue with madVR ?

Asmodian

12th August 2019, 21:51

I have no issues with current drivers on my 2080 Ti.

ng4ever

12th August 2019, 22:02

I have no issues with current drivers on my 2080 Ti.

Thanks. :thanks:

fourbigkids

12th August 2019, 23:13

Hi all. I have been having problems with madvr for some time. The latest issue occurred when I upgraded the driver for my video card, a Gigabyte Nvidia RTX 2060, and then had to downgrade to an earlier driver when the latest gave me a black screen when running two displays.

I use MPC-BE. When Madvr is selected as the video renderer, the colours are washed out. These are just ordinary DVD or blu-ray rips. Nothing fancy. If I change the renderer, or if I use a different player such as VLC or Media Player, the colour appears normal, just not with madvr. The TV is UHD and automatically handles the colour gamut and madvr is setup for PC 0-255. I have reinstalled MPC-BE and madvr (since the downgrade in the driver), but no luck. Still washed out compared to any other player or renderer.

Appreciate any pointers you might have.

Cheers,

Allan

Asmodian

13th August 2019, 00:02

Do you have a "this display is already calibrated" display setting configured? Set it to BT.709 with a 2.2 gamma. :)

SamuriHL

13th August 2019, 02:26

Is it true Nvidia RTX cards have a freezing issue with madVR ?

Freezing, no. There are microstutter issues that some of us have. But we've worked around it 1 of 2 ways. Either use D3D9 or set the present frames in advance to 1 if you are impacted by the issue.

ng4ever

13th August 2019, 02:27

Freezing, no. There are microstutter issues that some of us have. But we've worked around it 1 of 2 ways. Either use D3D9 or set the present frames in advance to 1 if you are impacted by the issue.

Cool. Thanks! :thanks:

fourbigkids

13th August 2019, 02:50

Do you have a "this display is already calibrated" display setting configured? Set it to BT.709 with a 2.2 gamma. :)

Thanks Asmodian. That fixed it! You're a star! :-)

Cheers,

Allan

Alexkral

13th August 2019, 10:15

Do you have a "this display is already calibrated" display setting configured? Set it to BT.709 with a 2.2 gamma. :)

Why does that need to be configured? Isn't that the default setting? Or maybe he had another configuration and set "disable calibration..." does the same?

chros

13th August 2019, 19:24

My 2080 Ti drops an average of 4584 frames with a standard deviation of 265 (5.8%) over 10 repeats.

i run it 2 times at 60 hz and got very reliable results from this both runs where pretty much spot on 33000 dropped frames.

MSI 1060 Gaming X 6GB, 9 runs, median: 36855
36898 , 36469 , 36921 , 37039 , 33106 , 36855 , 36900 , 36964 , 35671

What does it tell us? (Apart from 2080TI is faster than 1060 6GB)

huhn

13th August 2019, 19:25

i was using a 960.

Asmodian

13th August 2019, 19:38

I am not even sure this method is useful... we don't have enough data (yet?) but it is odd that the 960 seems faster than the 1060 based on your reports.

Upload queue was 1-8 or higher the entire time?

What does it tell us? (Apart from 2080TI is faster than 1060 6GB)

Did you see my previous post?

My benchmark is simply a "madVR performance score". Once we have enough data we could say what score you need to do 1080p24 -> 2160p with NGU high "with some headroom", or any setting we are interested in. The same with adding BlueSky or similar. If users report their experiance with a particular GPU and a particular configuration we could compare the performance score of their GPU with another no one has posted anything about to decide if it is likely to work.

Edit: There are 38073 frames total, maybe dropping that many is testing some interop step more than NGU speed. :(