madVR - high quality video renderer (GPU assisted)

[blaubart]

@ huhn ..so also you didn't read and understand what I'm saying - this may be the beginning of a wonderful friendship :p

[huhn]

should be a driver bug.

madVR doesn't do anything different if you send RGB, 4:2:0 or something like that.

[Thunderbolt8]

Madshi knows about this and tries to do something about it.ok, thanks.

[zapatista]

a general question for the video experten here:
is it possible with madvr to use the GPU (for ex from a gtx 1060) to play mp4 video files instead of it being processed by the cpu ?
my haswell i7 cpu seems to do most of the work , and i am hoping to offload it to the more modern GPU that might be more efficient (fine for mp4 @ 2k, but the cpu maxes out on 4k mp4 files while the GPU is only @ 10% load). with HEVC files of same resolution, the GPU does most of the work and the cpu stays @ 20 % for ex, this is much more hardware cpu/gpu efficient (i know hevc is a much more efficient file format, but am hoping the gpu can be used more for mp4 decoding now to)

Madvr doesn't do video decoding by itself, other components need to handle that. Usually, players/decoders that allow HEVC hardware decoding also allow H.264/AVC hardware decoding (H.264/AVC is the most common codec in mp4 files, but there are others). For example LAV Video/MPC-HC can do that.

thank you for the answer, i am not sure how to implement your advice however
i have installed madvr, mpc-ht and the LAV Filters 0.73.1 , and been using madvr for many years now on my older htpc setup (but have no expertise in video technology and what exactly does what part)

what settings should i look at to ensure my new GPU (gtx 1060) does the mp4 (h264/AVC) decoding so i can lower the CPU load on my older haswell processor

ps: my current newly re-installed mpc-ht and madvr settings are pretty much still set at default right now

[huhn]

the default settings from mpc-hc do use hardware decoding by default by using DXVA2.

try this guide: https://www.avsforum.com/forum/26-home-theater-computers/1357375-advanced-mpc-hc-setup-guide.html

it's old but still "accurate". the Video Decoding part is where you should look into.

i recommend DXVA2 copyback.
hardware decoder doesn't really put strain on a GPU.
because a file is "mp4" doesn't mean it can be hardware decoded just to make this clear.

[Warner306]

I can't live without LG's "active HDR". It tries to mimic HDR10+/Dolby Vision by analyzing the whole scene and adjusting brightness and tone mapping per frame. I don't know if this is something madshi is working on or is even possible with madVR, but until/if this happens I'm sticking with LG's implementation.

It really helps with shadow detail in dark scenes. Often times madVR is simply too dark compared to LG with active HDR turned on.

You can use madVR for tone mapping and still turn on LG active HDR, but then you've got kind of a double tonemapping situation where the tv is trying to dynamically adjust brightness and highlights after madVR has already tone mapped. It doesn't quite look right IMO. Often times it will just make the highlights even darker.

Both madVR and all of the current OLED TVs do dynamic tone mapping that is similar to HDR10+. The goal is to reduce brightness compression in any scene where this is possible by ignoring the single peak brightness variable and measuring each frame and each scene independently. The brightest scenes usually see no relief in compression, but dimmer scenes can be brighter.

madVR's implementation generally alters the information above reference white (100 nits), while the newest LG OLEDs (especially the 2018 models) appear to alter reference white as well as the brighter specular highlights. The LG OLEDs are known to be the most dynamic when it comes to changes in brightness compared to solutions offered by Sony, Panasonic and Phillips. This is probably a matter of taste. None of the OLEDs should be boosting the brightness of any values in an artificial way, just restoring the recorded brightness of each pixel whenever possible. PQ values are absolute in terms of display nits and each pixel is meant to be displayed at a specific brightness. Tone mapping makes the image darker and sometimes causes a visible loss of detail when done poorly as well as some loss of color saturation for bright image areas.

[Warner306]

I thought that madVR does do dynamic tone mapping to adjust brightness to the whole image, but I think this only seems to work in SDR. It should be able to do this since it does know peak brightness of each scene. However, for most of the range when output in HDR, brightness seems to be unchanged. Is this due to the static metadata that is sent for HDR? Are our TVs only relying on the HDR nits mastering in the metadata to scale brightness? Maybe if the metadata can be manipulated to send peak brightness of the scene as mastering metadata, it may change the brightness range dynamically. I don't know.

If you set pixel shader to 750 nits, only pixels above 750 nits will receive any tone mapping at all. This is how the current test builds work when "measure each frame's peak luminance" is checked. So the first 750 nits should be displayed 1:1 on your display (only if it is that bright), and anything brighter will be tone mapped. This means only very bright scenes should show a visible change.

Of course, this depends on how your display reacts to the input signal. It may take the 750 nit input and compress it instead of displaying it with clipping, which would be inaccurate and less bright than it should be.

[corporalgator]

thank you for the answer, i am not sure how to implement your advice however
i have installed madvr, mpc-ht and the LAV Filters 0.73.1 , and been using madvr for many years now on my older htpc setup (but have no expertise in video technology and what exactly does what part)

what settings should i look at to ensure my new GPU (gtx 1060) does the mp4 (h264/AVC) decoding so i can lower the CPU load on my older haswell processor

ps: my current newly re-installed mpc-ht and madvr settings are pretty much still set at default right now

How much of a load are we talking about? If you stick to letting your CPU decode, that frees up more of your GPU for madVR rendering.

[Warner306]

That's not true. A hardware decoder is a seperate part of the GPU. Hardware acceleration will actually improve performance (not in madVR, but in general) by freeing up the CPU without harming GPU performance.

[Betroz]

That's not true. A hardware decoder is a seperate part of the GPU. Hardware acceleration will actually improve performance (not in madVR, but in general) by freeing up the CPU without harming GPU performance.

For people who have a GTX 1080 card, what hardware decoder do you then recommend? CUVID or...?

[Warner306]

DXVA2 copy-back if you want hardware deinterlacing and black bar detection. D3D11 Automatic (Native) for everyone else.

[Manni]

DXVA2 copy-back if you want hardware deinterlacing and black bar detection. D3D11 Automatic (Native) for everyone else.

You can get black bar detection with D3D11 copyback (by selecting the GPU instead of leaving auto/native in LAV). That’s what I use and it works fine, although there is a performance loss with copyback compared to native.

[nevcairiel]

You can get black bar detection with D3D11 copyback (by selecting the GPU instead of leaving auto/native in LAV). That’s what I use and it works fine, although there is a performance loss with copyback compared to native.

There is also a performance loss with D3D11 copyback compared to DXVA2 copyback.
D3D11 is really only recommended in native mode, its copy-back performance is lower and can impact the GPU more due to some constraints in the API (ie. you cannot access D3D11 textures directly from software, probably so that GPUs can handle them more efficiently if they don't need to worry about that).

[Manni]

There is also a performance loss with D3D11 copyback compared to DXVA2 copyback.
D3D11 is really only recommended in native mode, its copy-back performance is lower and can impact the GPU more due to some constraints in the API (ie. you cannot access D3D11 textures directly from software, probably so that GPUs can handle them more efficiently if they don't need to worry about that).

Thanks, I'll try to go back to DXVA2 copyback then.

I hope Madshi will support D3D11 fully at some point, it would be nice for those of us who need to detect black bars. :)

[huhn]

the issue from blackbar detection comes from it runs on the CPU not because of d3d11. it is clearly possible to do it on the GPU and planned too if i'm not mistaken. but not 100 % sure why you want that on the GPU it's not like copyback hurts notable on good GPU but a GPU version of blackbar detection is clearly not going to be free.

@nev
if i understand d3d11 copyback with a headless GPU correctly than d3d11 decoding will only effect the decoding GPU not the rendering GPU because there is no real difference for the rendering GPU between this and software decoding right?

or in short d3d11 copyback from a headless GPU doesn't effect madVR at all?

and i would say d3d11 native should not be recommended until madshi support deinterlancing with it.

[HDR]

Not really, its just "fake".

Actual dynamic HDR has mastering engineers controlling the parameters of every scene, so that intentional brightness differences are fully maintained, which can give a movie much more depth.
Those automatic algorithms have no clue about the creative intent of a movie and they just "unify" everything, which can take away a lot of depth.

For me, its in a similar boat to the extremely cranked up brightness and color vibrance, or those fake HDR encodes you can find online made from consumer SDR sources. It makes for good show floor presentation, but its far from the creative intent of a movie. If you are into that sort of thing, go nuts. But don't claim your subjective preference is the only valid way to watch anything, because thats just silly.
Considering there is no standard for tone mapping you can't say that LG's method is better or worse than any other method, it's just different.

I prefer LG's active HDR and so do many others.

Do you actually own an oled and have compared it on/off? You're acting like it turns the tv into vivid mode or something. It doesn't.

It subtly increases brightness in dark scenes to better show shadow detail, and subtly decreases brightness in bright scenes to restore highlight detail, similar to madVR's highlight recovery.

Anyway, we all have our preferences.

[huhn]

but madVR follows the bt 2390 standard.

it will change colors that can'T be shown by your display because it can't get bright enough.

if you set the target peak nit to the same of the source file or higher it will not change colors.

and if your LG plays around dark parts of the image it is clearly not different from vivid mode. because it is clearly capable to display the dark details at the creator intended brightness so there is no need to change anything.

[HDR]

https://www.avnirvana.com/threads/hdr10-active-hdr-dolby-vision-and-other-hdr-schemes.588/

From what I am told "active" HDR10 works very much like the proposed HDR10+.

Samsung has proposed a change to the SMPTE ST.2086 base standard HDR10 by adding dynamic metadata so it can perform much like Dolby Vision. Another very similar method that measures each frame on the fly and tone map the display to it's brightness capability, which is commonly called "active HDR".

LG and Sony are doing "active HDR" with the current base standard HDR10 on all of the 2017 X1 Extreme processor TVs, like the A1E OLED, Z9D, X940E and X930E, and LG employs a very similar active HDR10 processing with their SJ9500, C7, E7, G7 and W7 OLED TVs.

As I am more familiar with how Sony's active HDR10 operates I'll keep my comments to Sony's 2017 X1 Extreme models. You might notice very little difference between Dolby Vision and HDR10 on Sony's X1E equipped TVs. Sony HDR TVs don’t use the static brightness metadata in HDR10 (MaxCLL, MaxFALL) they actually measured brightness frame by frame and generate dynamic metadata for HDR10 content.

MaxFALL, stands for "Maximum Frame/Average Light Level" and MaxFALL corresponds to the highest frame average brightness of one frame from the entire content.

MaxCLL, is the Maximum Content Light Level and is an additional static HDR metadata that represents and measures the brightest pixel of the entire content.

In part this explains why the standard 10% peak luminance window test pattern may measure a lower peak luminance than what the display is actually capable to deliver when we view actual HDR content that only use a far smaller area of the screen with the HDR specular highlights that typically occupy less than 1% of the display.

This HDR anomaly is most noticeable when the content is mastered at 4k nits. In Sony's new 2017 X1 Extreme processor TVs the HDR algorithms are tuned to apply tone mapping (which reduces screen brightness and accuracy) when the brightness of the frame exceeds the TV set’s capabilities. It is then applied to ensure HDR highlights are properly displayed.

[huhn]

and now you go and mark the part in that quote that states it makes darker parts more bright.

the only reason to not measure the peak brightness is performance.
this doesn't change the max nits added by DV and HDR10+.

[Betroz]

Sorry if I'm being slow to understand this, but I need some clarification. For me to avoid Black Crush / dark image I must use the following settings :

- TV set to Black Level = Low
- Use Nvidia color settings, RGB, 8bcc and Full
- MadVr set to TV levels (16-255)

Is this correct then? My LG C8 TV have only two options for Black Level when in HDMI mode, High and Low. I have tried to set the TV to High and MadVr to 0-255, but that gives me raised blacks / lighter image instead. I'm so confused since I did not have this problem with my old Panasonic VT30 plasma, or at least I didn't notice it. Btw this is with SDR content.

[HDR]

- TV set to Black Level = High
- Nvidia - RGB, 8bcc and Full
- MadVR set to (0-255)

This is what use on my LG oled and it's correct for desktop and videos.

[Betroz]

- TV set to Black Level = High
- Nvidia - RGB, 8bcc and Full
- MadVR set to (0-255)

This is what use on my LG oled and it's correct for desktop and videos.

Like I said, I then get raised blacks / lighter image.

[kitame]

increasing the minimum of monitor range from 0-255 to 16-255 would place a 16-level bias on the blacks towards whites.
leave it at 0-255 if you want deeper blacks.

monitor setting's "black level" on the other hand is a gamma curve adjustment, "High" puts it on a higher curve (brighter) while "Low" puts it on a lower curve (darker).
i suggest disabling it entirely if possible, as you want the gamma curve to be as close to the original as possible.

-----------

is there any performance benefit for MadVR when used with Nvidia's new RTX* cards?
*Turing architecture with RayTracing and Tensor cores.
if we're looking at this purely from raw processing power (GFLOPs) standpoint, Tensor cores on RTX cards is 8times more than their regular CUDA cores.

[mytbyte]

but madVR follows the bt 2390 standard.

it will change colors that can'T be shown by your display because it can't get bright enough.

if you set the target peak nit to the same of the source file or higher it will not change colors.

and if your LG plays around dark parts of the image it is clearly not different from vivid mode. because it is clearly capable to display the dark details at the creator intended brightness so there is no need to change anything.

I'm not technical enough to understand the BT.2390 paper in full, and can't quite grasp how the rolloff algorithm works (for 200 cd/m2 max brightness on a comp monitor, the 70% I measured as being way too high, 80-100 are flat i.e clipped in the graph (probably depends on MaxL of content and mastering monitor etc). But, everything under 50% (95 nits) is more-less following the PQ curve at all times so there should be no major black crush from MadVR's tone map curve (note: I'm doing HDR->SDR)

[huhn]

Sorry if I'm being slow to understand this, but I need some clarification. For me to avoid Black Crush / dark image I must use the following settings :

- TV set to Black Level = Low
- Use Nvidia color settings, RGB, 8bcc and Full
- MadVr set to TV levels (16-255)

Is this correct then? My LG C8 TV have only two options for Black Level when in HDMI mode, High and Low. I have tried to set the TV to High and MadVr to 0-255, but that gives me raised blacks / lighter image instead. I'm so confused since I did not have this problem with my old Panasonic VT30 plasma, or at least I didn't notice it. Btw this is with SDR content.

just open this image: http://www.lagom.nl/lcd-test/black.php
set your GPU to full range RGB and check which setting makes all or at least most levels visible.

16-255 is unlikely to be correct.

[HDR]

Like I said, I then get raised blacks / lighter image.
We have the exact same setup, so I don't see how that's possible.

[ryrynz]

is there any performance benefit for MadVR when used with Nvidia's new RTX* cards?


Madshi may be able to take advantage of the tensor cores but that'll likely be quite some time away, madshi will mention something I'm sure once he's taken a good look at what can be done with them.
Apart from that the 2080Ti is about 40% faster for madVR related tasks over the 1080Ti.

[kitame]

Madshi may be able to take advantage of the tensor cores but that'll likely be quite some time away, madshi will mention something I'm sure once he's taken a good look at what can be done with them.
Apart from that the 2080Ti is about 40% faster for madVR related tasks over the 1080Ti.

thanks, though 2080Ti is quite a bit more expensive than 1080Ti.
of which, the entire RTX series are a tad more expensive than the Pascal cards, so unless they have any sizeable advantage then i doubt it'd be worth it.

[ryrynz]

No need to quote my post.
They're worth it to anyone dealing with HDR 4K and 60FPS content wanting to unlock high/very high NGU modes for the most part, the 10 series still represents excellent value for madVR tasks
and right now I can't see any point in recommending the new cards apart from the 2080Ti right now unless you want the best madVR performance you can get though the 2080 would make sense if buying new as based on what I've seen so far it's about the same price as the 1080Ti.

[Betroz]

Well I can't disable the Black Level setting in my LG C8, and all reviews online recommend to set it to Low anyways. I guess I have to take a picture of my TV later to prove it.

[huhn]

reviews don't try to opimise for HTPC usage.

if you have your GPU at full range RGB. than this answers your question: "Like I said, I then get raised blacks / lighter image."
the option with the most "risen black level" is correct. because with a full range signal you can only get a clipped black/whites or correct one(ignoring madVR setting and expecting or defaults).
just try my very easy to do test with the image.

[Betroz]

reviews don't try to opimise for HTPC usage.

if you have your GPU at full range RGB. than this answers your question: "Like I said, I then get raised blacks / lighter image."
the option with the most "risen black level" is correct. because with a full range signal you can only get a clipped black/whites or correct one(ignoring madVR setting and expecting or defaults).
just try my very easy to do test with the image.

Thanks for reply. Will come back here later when I have re-tested this again.

[mclingo]

Thanks for reply. Will come back here later when I have re-tested this again.

Hi, I follow LG OLED very closely and they all have roughly the same settings in this area, I can guarantee that is the recommended settings you've been given arent working correctely for you, you either have a fault or something setup incorrectly.

Double check the following again:

PC

Graphics card set to FULL RGB 4:4:4
MADVR set to 0-255
video player set to MADVR renderer

Try these settings as a test but also report your current settings.

TV

Black level HIGH
Gamma 2.2
Brightness 49
OLED light 50

With these settings and using the black clipping pattern you should see a black crush and will probs only see 20+ flashing.

report back what you get.

Which movie player are you using?

[iSeries]

Thanks for reply. Will come back here later when I have re-tested this again.

This all might be better off in the general thread but yes as others have said, MadVR at 0-255, GPU at 0-255, and black level high will be correct.

However, my experience with LG TVs, they all behave exactly the same when it comes to RGB - they are better near-black when fed limited range. For example, with my C7 brightness at default 50, using the pattern from https://www.avsforum.com/forum/40-oled-technology-flat-panels-general/2579537-lg-oleds-how-enhance-near-black-detail-3.html#post46798617, with a complete 0-255 chain, the first square that is meant to be visible (which is actually the second square, the first square is completely black) is crushed to black, but with MadVR at 16-235, GPU at 0-255, and TV at low, that square is visible. I've had 5 LG TVs over 10 years, and they have all acted like this.

[mclingo]

i've always had better control of black clipping feeding mine correctly, you'll get an extra conversion using limted in MADVR which can degrade the picture in other ways?

[iSeries]

There's no extra conversion with MadVR limited, GPU full, TV at limited (LG low). Also, with your brightness at default, test the pattern linked above :-)

[Manni]

the issue from blackbar detection comes from it runs on the CPU not because of d3d11. it is clearly possible to do it on the GPU and planned too if i'm not mistaken. but not 100 % sure why you want that on the GPU it's not like copyback hurts notable on good GPU but a GPU version of blackbar detection is clearly not going to be free.


I don't really mind whether the black bars detection is done by the CPU or the GPU, I would only like to be able to use D3D11 native instead of having to select D3D11 copyback and lose performance simply because I need the black bars detection.

If DXVA2 copyback performs better than D3D11 native, then I might go back to DXVA2 copyback and not care much, otherwise I'd still like to be able to use D3D11 native without losing black bars detection, as I shift the picture to the top of the screen for my manual masking (single mask at the bottom of my 16/9 screen when playing 2.35 or 2.4 content) and can't do without it.

EDIT: also I forgot we needed D3D11 to get 10bits without FSE, that's probably why I'm using it. I don't want to use FSE unless I can't do otherwise (i.e. in 3D) due to the long HDMI resync it causes here when displaying a menu on screen, so full screen windowed in 10bits is a requirement for me. Unless I'm mistaken on that point, going back to DXVA2 isn't possible.

[mclingo]

There's no extra conversion... :-)

you'll get a crushed desktop in this config though so web and youtube and other stuff will be crushed too wont it?

[huhn]

it depends on your GPU but i have a hard time to see d3d11 win at all if you care about GPU performance until madshi switches to d3d11 render pipe. i have no proper way to measure copyback cost but for a modern high performance GPU it's not much at all. but using d3d11 even the madVR OSD tells you what it costs.

i still don't understand why you care about decoder costing you GPU performance but blackbar detection is fine when it takes GPU performance.

time will tell if copyback cost more or doing blackbar detection on the GPU.
even through copyback has other benefits.

[iSeries]

you'll get a crushed desktop in this config though so web and youtube and other stuff will be crushed too wont it?

In terms of video playback, there is no extra conversion - MadVR outputs limited range, GPU assumes it's full range and just passes it through. Only when GPU is at limited range would there be an extra conversion.

And sure the desktop is crushed, but the TV has lots of presets. I use ISF bright room with the black level set to high, and ISF dark room set to black level low (all other settings kept identical between them, so switching between bright and dark the only difference is the black level setting). Mine is an HTPC so I don't really use it for anything else anyway (I use the TVs YouTube and other apps with ARC). I know you use Kodi DSplayer, and that has the option to display in 16-235 for correct UI levels with a limited>full>limited chain.

[nevcairiel]

EDIT: also I forgot we needed D3D11 to get 10bits without FSE, that's probably why I'm using it

This only requires madVR in D3D11 presentation mode, it has nothing to do with how the decoder operates.

There is only a few things that the decoder mode influences:

- CPU-only madVR algorithms (black bar detection, and maybe others?), does not work with DXVA2 Native or D3D11 Native
- Deinterlacing, doesn't work with D3D11 Native only.

Everything works with Software decoding or DXVA2/D3D11 Copy-Back (because copy-back looks like software decoding to a renderer). And between those last two, DXVA2 Copy-Back is usually preferable because its more efficient, unless you need D3D11 to access a headless GPU.

[mytbyte]

you'll get a crushed desktop in this config though so web and youtube and other stuff will be crushed too wont it?

Correct, anything that outputs full range is crushed, just confirmed on my LG HDMI monitor.

[mclingo]

I see, I'm a one stop shopper myself, minimal setup, I dont like having to change TV preset based on content. I will test it for myself though.

cheers


Please move this discussion to here from herein, its blocking MADVR issues.

https://forum.doom9.org/showthread.php?t=175769

[iSeries]

Well that's your choice, but I'm not going to settle for crushed near-black just because I don't want to press a button on the remote ;-) Like I said, I only use it as a HTPC (Kodi DSplayer also), so I never even really see the desktop. Kodi DSplayer is set to 16-235, as is MadVR, GPU to 0-255, TV to 16-235. All levels are correct, no near-black crush.

LGs simply do better when working limited range.

[huhn]

or you use 1-255.

[iSeries]

Yes, that's also an option. As is raising the TV brightness one notch. But both those options raises everything, whereas the only difference between limited>full>limited and full>full>full is that first near-black shade which is crushed with the latter. All other shades above that are at equal brightness (measured with meter).

No idea if that would be detrimental to the TVs HDR tonemapping as well. For sure, the advice is to not touch any of the default settings on the TV for HDR as it messes up the tonemapping.

[j82k]

Complete 0-255 chain vs MadVR at 16-235, GPU at 0-255, and TV at low doesn't make a difference on my C8.

Tested in a pitch black room using the linked and other black clipping test pattern.

Also it would have really surprised me if it made a difference regarding near-blacks/black crush. :rolleyes:

[Manni]

This only requires madVR in D3D11 presentation mode, it has nothing to do with how the decoder operates.

There is only a few things that the decoder mode influences:

- CPU-only madVR algorithms (black bar detection, and maybe others?), does not work with DXVA2 Native or D3D11 Native
- Deinterlacing, doesn't work with D3D11 Native only.

Everything works with Software decoding or DXVA2/D3D11 Copy-Back (because copy-back looks like software decoding to a renderer). And between those last two, DXVA2 Copy-Back is usually preferable because its more efficient, unless you need D3D11 to access a headless GPU.

Thanks, that clarifies everything. I'll compare DXVA2 CB and D3D11 CB, although I'm often using Teamviewer to control the HTPC without having to switch the PJ on or switch inputs on my 4K monitor, so if that's what you call headless GPU I'll probably need D3D11.

[huhn]

no it's not headless. native would fail if you use your whole system headless.

[nevcairiel]

The primary GPU is always available through DXVA2-CB as well, thanks to D3D9Ex. Just any secondary fully headless GPUs would not be.