madVR - high quality video renderer (GPU assisted)

[Sunspark]

OK, but even at 100% zoom in the browser I can still see the chimney cover changing size between more of a line vs more of a denser rhombus.

[huhn]

well yes scaler are different

[matbezlima]

Below are my current MadVR settings for tonemapping from HDR to SDR, but which ones are the best? Please, someone tell me. The image I get from the settings below doesn't look good enough for me. My TV is crap for HDR, this is why I want to tonemap it to SDR. I'm a total newbie with madVR.

https://i.ibb.co/0p0Zy5gj/Mad-VR-125439.png

Another question: is there any video converter that allows one to reencode a video to SDR so that the output is what MadVR renders? I would love a program that is able to use madVR for the tonemapping from the HDR input to the SDR output file. I know that MadVR in itself is just a renderer for live playback.

[huhn]

read this please: https://forum.doom9.org/forum-rules.htm

i do not tone map actually i do not use madVR for HDR at all. so i can not help you at all. other with massive tone mapping experience don't seem to be active anymore or don't want to answer a question with no answer.

[matbezlima]

read this please: https://forum.doom9.org/forum-rules.htm

i do not tone map actually i do not use madVR for HDR at all. so i can not help you at all. other with massive tone mapping experience don't seem to be active anymore or don't want to answer a question with no answer.

Okay then.

[Sunspark]

That control panel for HDR seems overly complicated for a casual user. Are hardware calibrators a requirement to use it?

[huhn]

yeah i didn't follow it. so if user have HDR tone map issue i can not help. i really hate to support stuff i know little about. i'm already often enough wrong no need to to mangle with stuff i do not know.

[tp4tissue]

On the latest build, when it goes full screen exclusive, the mouse cursor would often freeze, it's still moving in the background, but the cursor won't go away.

Tp4 is using AMD. NE1 else experience this?

[tp4tissue]

Okay then.

What exact TV do you have. We can help you with basic settings.

[VergiL]

I'm continuing the tests; here's a screenshot after setting MPCBE to maximum performance under power management in the Nvidia control panel. With the same settings, my Max Stats dropped from approximately 25ms to 3ms. Is such a difference normal?
https://i.goopics.net/800/w12sr9.jpg

[huhn]

that's just the power state changing 25 ms is not an issue.

[Amuat]

On the latest build, when it goes full screen exclusive, the mouse cursor would often freeze, it's still moving in the background, but the cursor won't go away.

Tp4 is using AMD. NE1 else experience this?

I recently mentioned in relation to MPC-VR how it takes two tries every time to get the pointer to disappear, so doesn't seem to be exclusive to madVR (sorry for the terrible pun).

edit: I use NVidia

[VergiL]

that's just the power state changing 25 ms is not an issue.

Yes but presentation glitches are no longer present below 10ms, so that at least solves that problem for me..

[huhn]

a glitch can happen with a bad timed frame.
you are frame rate matching even with 20 or 30 ms rendertimes spikes there is at least 11 ms to still present the frame in the correct vsync.

the frame times of madVR are pretty bad which was fine in old windows versions.

[Sunspark]

Why was it fine with old versions if the times are bad?

[huhn]

i don't know the glitches issues happen to start with win 10. win 7 FSE was utterly fine.

[Sunspark]

For those downscaling old film shows from 1080p to 768/720p, SSIM 1D 25 AR (no LL) is very nice. Good clean rendering when paired with Mitchell-Netravali AR chroma.

2D is marginally better, but a 100% increase in GPU load from 1D isn't worth it for marginally better on an old system.

The commonly used value of 100 is a bit over-sharp so 25 is better. Might be a different story going from 4K to 1080p or 720p but I can't test that.

[tp4tissue]

I recently mentioned in relation to MPC-VR how it takes two tries every time to get the pointer to disappear, so doesn't seem to be exclusive to madVR (sorry for the terrible pun).

edit: I use NVidia

So you're saying it might be a windows glitch in combination with new madvr changes to go FSE ?

The old Madvr version would fail to go FSE, the new version can always go FSE, but it'll freeze the cursor on screen sometimes.

I've tested mine on 7900xtx 7900xt 6950xt 6900xt

[Sunspark]

Do you have any information what the new version does differently compared to the old version where FSE is concerned? I found that FSE "seemed" to work on the old version only with D3D11 mode selected (along with fullscreen optimizations disabled checkbox), but I don't trust the refresh rate numbers it was reporting.

[tp4tissue]

Do you have any information what the new version does differently compared to the old version where FSE is concerned? I found that FSE "seemed" to work on the old version only with D3D11 mode selected (along with fullscreen optimizations disabled checkbox), but I don't trust the refresh rate numbers it was reporting.

The FSE version does work on the old version madvr, but it often doesn't trigger after you run other apps, like chrome for example.

So, you have to kill explorer.exe then when it reloads, the FSE works again.

BUhhhhhh, the problem NOW, is, FSE always triggers, but then the mouse cursor freezes "Visually" on the screen, it's still moving running in the background, but it's stuck on the screen, frozen in place..

[Sunspark]

I hate when glitches like that happen.

On my Linux setup that needs to be updated, once in awhile I will get a situation where I'm in the browser and all of a sudden the screen goes black.. the mouse cursor will still move around on it, but that's all. I will still hear audio that was already playing.. but I can't use keyboard commands to put it to sleep or whatever.. have to force a poweroff.

On my Windows setup connected to the old TV, sometimes it will wake up from sleep and only display a black screen, no cursor or anything. The TV is receiving "a" signal because it doesn't say no signal detected, but there's nothing on the screen.. another force poweroff to fix.. but when it was connected to the monitor that never happened.

[tp4tissue]

Ok, it seems like my issue with cursor freeze is due to using dx11 in lavfilter, if you change to dxva2copyback, no more cursor freeze.

This could be very problematic with some laptops. They got weird issues with dxva 2 cb, such as banding. (which is why) Tp4 switched to mostly using dx11 in the first place.

[Amuat]

edit: sorry, missed your latest post about dx11 in lavfilter, but it worked with previous iteration of madVR, correct?

So you're saying it might be a windows glitch in combination with new madvr changes to go FSE ?

The old Madvr version would fail to go FSE, the new version can always go FSE, but it'll freeze the cursor on screen sometimes.

I've tested mine on 7900xtx 7900xt 6950xt 6900xt

Well, as some of you may be aware, I'm still somewhat of an idiot in the field, but even I dare venture to say it must have something to do with not only madVR/AMD. At least in my case with VR it has been an issue for a very long time now, as I tested it briefly back in the day and immediately discarded it because of that glitch. I'm happy to say I'm still rocking 23.978 with 23 Hz without issues, so personally I don't really care at all to be honest.

[tp4tissue]

Yes, the dxva2 copy back works most of the time, but for whatever reason, on laptop, because of igpu/dgpu shenanigans, muxchip, who knows, dxva2copyback can output severe banding in HDR content. Dx11 worked fine.

If it works it works, but just raising awareness, there's something up with the latest build and DX11 in Lav on AMD, frozen cursor.

[Sunspark]

Wow.. I found a setting in the Intel HD Graphics drivers that changes display rate and composition rate # reporting with madVR in the 3D settings panel.

Before, in Windows 10, the only combination of settings that would show a refresh and composition rate of 60.015 when the display was set to that with the driver set to the default setting of "Balanced Mode", was D3D11 FSE mode (w/ FS optimizations disabled checked). All the other ones would show something like 59.91/59.955 when the display was set to 60.015 which I felt was not ideal as I could perceive subtle things such motion almost, but not quite right.

Today, I see in the 3D section of the drivers where you have a choice of Performance, Balanced Mode, Quality and Custom, if you select Custom, then you can see a Vertical Sync option with the choice of Application and Driver, and if you select Driver, then the driver forces it. (Will need to go back and check to see if Application also works--Update: Application works fine.)

Now? Now it shows 60.015 for display/composition in all modes when set to that. D3D9 windowed, overlay, etc. not just D3D11.

Why is it even necessary to force it?

(Update 2: All modes except for FSE. Even with refresh rate switching unchecked, FSE will change to a different refresh rate! Weird, but still progress!)

(Update 3: Using old path backbuffers instead of new path frames in advance stops FSE from changing the refresh rate.)

[dbcooper]

read this please: https://forum.doom9.org/forum-rules.htm

i do not tone map actually i do not use madVR for HDR at all. so i can not help you at all. other with massive tone mapping experience don't seem to be active anymore or don't want to answer a question with no answer.

What software do you use for HDR video?

[huhn]

mpcVR

[Sunspark]

Currently playing with smooth motion 24 fps on a 60 Hz display to see if I like it or not.

Wondering if it is better to use software decoding for this since it is frame blending using the CPU, or if it makes no difference and copy-back is fine?

(I'm thinking about this from the perspective of obtaining a bunch of frames once and presumably holding them, vs constantly calling for the gpu to retrieve frames to copy over into shared igpu system memory?)

Likewise, I set the frames in advance to 16 and I can see the present queue when smooth motion is on is reported as 6-9. Should I lower frames in advance to 10?

[huhn]

it runs on the GPU pretty much everything with madVR runs on the GPU entirely on the GPU.

copyback operation cost CPU and GPU. usually irrelevant.
dxva native cost extra GPU with madVR because it tries to undo the damage it does. d3d11 native cost GPU thanks to the dx11 dx11 interop. software cost only the upload anything. copyback cost up and download.

[QBhd]

For simple playback of 23p and 29p material I have never understood why anyone in today's age would use anything other than software decoding...

QB

[tp4tissue]

For simple playback of 23p and 29p material I have never understood why anyone in today's age would use anything other than software decoding...

QB

Hrrrrm..

Just tested this. (Playback PC, 9800x3d, 7900xtx)

DXVA2 (Native)

110 watt gpu

13.2 watt cpu

DX11 (native)

110 watt gpu

15.8 watt cpu

DXVA2 (copyback)

110 watt gpu

33 watt cpu

DX11 (copyback)

110 watt gpu

33 watt cpu

CPU decode (software)

110 watt gpu

50 watt cpu


Ira^ war, Qatar blown up, european LNG energy crisis.

:(:(:(:(:(:(:(:(:(:(:(:(:(:(:(

[Sunspark]

Well, in my case it's a dual-core igpu connected to the TV.. yesterday I spent way too long looking at smoothness of motion on panning with different decoders, queue depths and presentation methods. Load caused by scaling choices, threading, etc. all affect frame timing. It's good I recently replaced the fan, repasted and puttied it, because some algo choices put the gpu load at 80%. By comparison, a pure dxva native decoding and scaling choice only uses about 3-5%.

For now, I am using quicksync, backbuffers 8 and regular windowed. On current systems probably the jitter, etc. wouldn't be noticed. I should see how it is on the Windows 7 partition sometime as it has true hardware overlay and FSE mode unlike W10 which only has kinda-sorta.

mpcvr on W10 is a good directshow renderer though if you don't need features like ivtc, smooth motion, etc. and I would recommend it. On the dual-core it's interesting seeing the graph line on the stats page, it's not completely flat, there is wobble.

All my stuff is old. My newest PC is the Steam Deck. Typing on it right now actually.

I was also reading old posts yesterday to look up information. Kind of depressing that this was such a lively project and then became pseudo-abandonware because the author is focusing on hardware.

[tp4tissue]

Current US electivoltz pricing, $0.25/kilowatt, You can save $5.037 a year, if you watch an average of 1.5 hour of 4KHDR per day using DXVA2native vs CPU-decode.

OBVIOUSLY, you can save alot more money by not using a stupid gpu like 7900xtx, this is really full returd stupid.

[huhn]

Hrrrrm..

Just tested this. (Playback PC, 9800x3d, 7900xtx)

DXVA2 (Native)

110 watt gpu

13.2 watt cpu

DX11 (native)

110 watt gpu

15.8 watt cpu

DXVA2 (copyback)

110 watt gpu

33 watt cpu

DX11 (copyback)

110 watt gpu

33 watt cpu

CPU decode (software)

110 watt gpu

50 watt cpu


Ira^ war, Qatar blown up, european LNG energy crisis.

:(:(:(:(:(:(:(:(:(:(:(:(:(:(:(
use a watt meter software is very inaccurate.

[tp4tissue]

It's AMD's native adrenaline driver readout. One would assume they have enough low lvl hardware access and categorized current sensing, but yea, it could be off.

[Sunspark]

I'm not trying to save electricity. I'm saying that if the system is too loaded down motion is perceived as slightly slow or fast. On the 5090s you all have, it doesn't matter.

I think I will copy the installers over to the W7 partition this weekend and see how it is with a true hardware scanout and no dwm in the way. Madshi always did say that Win 8.1 was the best version of Windows for madvr, not 10.

[tp4tissue]

There's a framebuffer no? Does dwm affect madvr?

[Sunspark]

Yes, it does. In Windows 10 and up it forces the use of the compositor. Even checking "disable fullscreen optimisations" on the player is only a partial fix.. e.g. I can take screenshots in overlay mode. I'm not supposed to be able to because it's writing directly to the hardware, etc. FSE kinda-sorta works. Part of it will work in that you can output something like 10-bit after you have checked the disable checkbox, but you can't check disable aero composition in madvr, etc. which you're supposed to be able to do for FSE to do the timings itself.

It's not a big deal, there are many offered output paths, but all of them now go through the compositor in W10+, or are partially emulated in the compositor.

In older Windows, you can write directly to the hardware scanout with overlay if your graphics card supported it (not all did), can disable the dwm compositor and use madvr FSE to time things instead of Microsoft doing it, etc. All the rest of the paths go through the compositor as they are windowed fullscreen modes using either backbuffers (discard model) or frames in advance (flip model).

[Sunspark]

Played around in Windows 7 today. Had to switch power management to high performance so it stayed at 100% cpu, otherwise there would be continued counts of dropped and delayed frames.

Overlay and Exclusive both work with the compositor off without tearing and not that much of a difference between the two, no composition rate line. I feel the smooth motion frame pacing is slightly better in 7. Not as treacly smooth as in 10's windowed mode, but the different isn't vast, the pseudo-emulated modes are not bad. Probably I will use 7 for a bit for this purpose and see how it goes.

[tp4tissue]

Does HDR output work properly on Win7?

[huhn]

why wouldn't it. madVR uses the nvidia api windows has no saying into this only intel has issue with that.

[tp4tissue]

Hrrrrm... I guess a 6700xt would le ideal for a dedicated Madvr box come Win7 support + modern fixn's like 12gb vram, HDMI 2.1, Working 10bit color correction,

Anything higher would be kinda wasteful, because you'd need win 10/11 for gamz.


What is the advantage of Win 7, I'm still a bit unclear what removing the Compositor does, does it produce LESS frame jitter/stutter? Tighter audio sync?

Does the compositor add bit errors?

[Sunspark]

Removing the compositor doesn't improve image quality. All it does is just allow tighter control of frame timing, so a slightly flatter line on the graph.

7 was the original OS on my NUC so it's not like I went to the effort to install it, I already had it there. I don't know that I would bother installing 7 just to avoid the compositor if I had a newer system.

flossy_cake, weigh in here.. I know you're using 7 on better hardware. What benefits do you accrue?

[tp4tissue]

Tighter frame-timings, as in render times?

On some gpus, older AMD, you can lock the frequency to max, probably give you lower render times than the Latest Gpus which can't lock.

[Sunspark]

No, I'm not talking about changing the clock of the GPU, though this is something you often need to do with both an igpu and a discrete card.

It's just a lighter load to wade through.. on my NUC Windows 10 has something like 150+ processes sitting around.. on 7 it's like 47. No compositor between the renderer and the graphics driver output is one less layer of abstraction.

This is more relevant for older machines not newer ones.

[j0b0rg]

Hi all,

I'm experiencing a reproducible issue with madVR (at v0.92.17) that only started after replacing my primary monitor. Looking for insights or similar experiences.

Setup
- GPU: NVIDIA RTX 4070 (driver 595.79, tried a few previous older versions to be safe)
- Primary monitor: Gigabyte MO32U2 (32" QD-OLED, 4K, connected via DP 1.4, 120Hz, 10bpc, RGB Full), Previous was (LG 27GR93U, 4K@144Hz via DP 1.4)
- Secondary display: LG OLED TV (SSCR2, 4K@120Hz via HDMI 2.1, RGB 10bpc Full, Windows HDR and Auto HDR enabled)
- Windows 11, extended desktop
- Players: DVBViewer Pro and PotPlayer, both using madVR as a renderer
- madVR HDR setting: passthrough HDR to display
- Fullscreen windowed mode (forced by Windows HDR), also tried Exclusive and Overlay

The problem
When I close PotPlayer or DVBViewer (which renders to the secondary LG OLED in fullscreen windowed mode), the LG OLED briefly goes black, then the Windows HDR logo reappears, then it goes black again.
It enters a phase in which the screen randomly blanks from time to time. The simplest way to break the cycle is to reset the display driver via Ctrl+Shift+Win+B.

Key detail
This never happened with my previous primary monitor (LG 27GR93U, 4K@144Hz via DisplayPort). The secondary LG OLED and its HDMI connection have not changed at all. The only change was swapping the primary monitor on the DisplayPort output.
It also doesn't help to run both monitors at 120 Hz. Currently, the primary runs at 165 Hz and the LG at 120 Hz.

What I've tried
- Disabling "restore original display mode when media player is closed" in madVR → no change
- Disabling "switch to matching display mode when playback starts" → no change
- Never have used any display switches before anyway (left blank)
- The issue occurs regardless of whether the primary monitor (MO32U2) is powered on or off
- During normal use (no player running), the LG OLED is completely stable on RGB/10bpc/Full


Has anyone experienced similar behavior in a dual-display setup where swapping monitors on one output affected HDMI stability on the other? Any suggestions beyond the workaround would be appreciated.

Thanks!

[tp4tissue]

Sounds like buggy DSC.

[Phoenixia]

I’m trying to generate a madVR 3DLUT using the standalone DisplayCAL-3DLUT-maker (not inside DisplayCAL), but I’m unsure which settings are correct for my situation.


I’m using an ICC profile that I downloaded from the internet for the same monitor model it's a wide gamut monitor with ~141% sRGB volume, and I have a few questions:


For tone curve and rendering intent — should I follow the “Default rendering intent” and gamma described in the ICC profile, or the source format?
For example, if the ICC profile uses gamma 2.2 but the source is Rec.709, should the tone curve be gamma 2.2 (matching ICC) or Rec.1886 (matching the source)?

Regarding gamma mode — should I use relative or absolute, and should black output offset be 0% or 100% to best preserve the ICC profile’s intended rendering?

Should VCGT be enabled/checked when generating the 3DLUT in this standalone tool?

Are there any reliable test patterns or methods to evaluate which ICC profile is actually the most accurate without a colorimeter?


Any clarification or recommended settings for this workflow would be greatly appreciated. Thanks!

[huhn]

1. the curve you want. if you tone map you need 2 3D LUT cause the one for HDR needs to be 2.2.
2. what do you want? with a perfect black device it does not matter else it matters a lot.
3. who knows is it on by default?
4. no. meter or bust.

[tp4tissue]

ICC profiles are useless without a colorimeter.

Every TV is different. Often by an extreme amount.

You also have to know the EXACT setting/condition that ICC profile was created, because different TV modes, settings will drastically alter behavior.

But it ALL starts with buying a colorimeter, avoid the Spyder brand, they're useless for OLED, get a calibrite. Doesn't really matter which one, get the one rated for lower brightnesses, the 3000nit one has worse near black stability.

If you can find a new old stock, the OLD x-rite i1d3 (Pre rebrand Calibrite) are the Best for OLED, or CCDIS3 PRO

141%, I'm assuming you have a qdOLED.

ALSO, you don't want to use 3DLUT to watch most movies. For most movies, you just want the 2D corrections, Gamma, RGB curve, Whitepoint.

If you use 3DLUT, it will clamp to either DCIp3 or Rec709(srgb), which rarely represent the author's intent, only the edges of the container spec..

On OLED, a 3DLUT for SDR will actually function properly (Though, you SHOULD NOT use one for 99.99% of movies).. But HDR 3DLUT will not work at all, due to how ABL algorithms behave, completely unpredictably, and MADVR doesn't know what's happening. The test pattern output does not have high correlation to real-content, so you can't really characterize HDR on consumer-use OLED.


In 99.99% of cases, HDR or SDR, you want Gamma/RGB/Whitepoint, and a reasonable HDR EOTF roll off setting.

DO NOT use an ICC from the internet, it's 100% wrong, 100% of the time.

The only thing you can download are Spectral Corrections to use with your Colorimeter (once you get one), which are typically close enough for general use.

Spectral correction, done correctly reflects the optical technology stack of the panel, whereas ICC includes all the firmware and software who knows what in between.