AMD, Intel and Nvidia driver issues and last recommended version

[aufkrawall]

I've tested with LAV Filters 0.74.1 and Adrenalin 20.11.2 driver: D3D11VA VP9 decoding is definitely fixed, it's now reported via "Video Codec 0" load in Task Manager instead of encode.

[el Filou]

Does that help in certain circumstances?I would imagine it's the modern equivalent of AGP Texturing in the old days: it can allow running a game/app that has absolute minimal memory needs bigger than your VRAM, but you would incur a big performance hit compared to VRAM-only.
I may try a GPU queue of 24 when I'm back on my HTPC this evening just to check how much that has an impact. :)
IIRC, people have tried running 4K madVR on a GeForce 1030 and encountered that issue: it would work but would drop frames due to "overflowing" into system RAM.

Edit: well, render queue of 24 didn't trigger it, but decode queue of 128 with native hardware decode did. It's really weird, total used memory was around 4,5 GB but instead of using all VRAM and a bit of system RAM it used only 1.3 GB VRAM and 3.2 GB system RAM. When looking at GPU-Z it caused massive traffic over the PCIe bus but all usage was very bursty and it dropped frames like crazy. As huhn said below, never ever count on that and always make sure you don't max out dedicated VRAM.

[huhn]

I'd rather assume that it was added in upstream ffmpeg (though I didn't check your claim with upstream LAV Filters), but I can't find a corresponding commit.

The broken Navi deinterlacing is a bummer, though my impression with Polaris was that its quality was crap vs. Intel anyway. Good thing mpv has automatic copyback for yadif by default when enabling deinterlacing.

they fixed it in 20.11.2. i tested 20.10.1
AMD deint was the best in my opinion(one of the reasons i usually got an AMD card as a HTPC GPU) beware polaris(if not all AMD cards) had a similar issue for many month where a really low quality deint was used not as bad as this but they fixed it one day. they broke it in the first crimson driver if you want to test that.

yadif is sadly not frame adaptive that's why it not hard to beat.

the nightly 0.74.1.59 was the first with the work around if you want to test it with an older driver or find the commit.

@ at the share memory stuff. showing shared memory in the taskmanager doesn't mean it new this was pretty much always possible and should pretty much always be avoided in a running program

[aufkrawall]

GPU deinterlacing btw. still works on Navi with VAAPI and I'd say it looks exactly like Polaris D3D11VA/VAAPI, which I find ugly due to aliased look vs. yadif.

Well, but screw deinterlacing. Hopefully AV1 works the same as VP9 with RDNA2. Tests sound positive, but they only try browsers.

[huhn]

you mean it "works" like this: https://abload.de/img/amddeint2mji2.jpg

or not at all?

[aufkrawall]

Drop me a cartoon sample, please.

[huhn]

you can force deint on everything for testing even an image.

i can still make you sample later.

[chros]

Hopefully AV1 works the same as VP9 with RDNA2. Tests sound positive, but they only try browsers.
Coreteks showed a slide in his review that only stated av1 8k@30 decoding (and no av1 encoding). Can this be true?

[aufkrawall]

While it's "only" 8k 30fps, it's still a rather convenient surprise that they've managed to support a new codec at the same time as their competitors. Intel & Nvidia don't support AV1 encoding either and only Intel support VP9 encoding.

[huhn]

sample: https://filehorst.de/d/dqmHnyet

[el Filou]

Isn't that sample a 3:2 telecine?

Anyway you made me want to test my old HD 7870 deinterlacing under D3D11 with true 576i and I confirm that the quality is awful even on such an old card, so it looks that both:
- DXVA2 deinterlace is broken on Navi (but still good on older cards even with modern drivers)
- they didn't even bother implementing good quality deinterlacing/scaling under D3D11 VP for any cards

I also noticed when doing comparisons, that the NVIDIA chroma issue is also present under D3D11 VP which is a shame, and that its deinterlacing quality isn't better than with DXVA2. It is markedly better on AMD cards when it works.
NVIDIA also shows a lot of artifacts on interlaced SD (at least 576i), and there are more of them when copyback decoding is used which makes no sense to me but that's how it is, still some of them can be seen with native but not as much.
Anyway SDTV is on its way out even on my cable provider who makes the least effort but it's sad they never got that totally right like AMD did.

[huhn]

yes it is telecine and there is an alternative to deal with this file.

nvidia needs a couple of frames before it start working so make sure that it get's them.

[el Filou]

Yes I make sure to let it play for a bit and then I play frame by frame instead of skipping directly to frame number, but it doesn't change anything, it's still lower quality than AMD's (but at least it's not buggy).
I only tested true video 576i from TV as I hardly ever watch any NTSC, I only have one DVD with 3:2.

[huhn]

can you make me a screen plz and name the hardware.
the screen i made is literal the best i can get out of amd under any situation and i'm not aware of a situation where nvidia falls apart.

[el Filou]

Here's a zip file with all the PNGs: https://filehorst.de/d/dfxxbqnl
It's from a vertically rolling credits and horizontally rolling news ticker (this one also showing bad chroma and artifacts).
I displayed at 100% (just stretched horizontally for 16:9 SD)

Here's a comparison of a small part where differences are obvious: https://abload.de/img/comparisono4ky9.png

Hardware is HD 7870 and 1050 Ti. I've not included NVIDIA D3D11 screens as it's identical quality to DXVA2.
I wouldn't say NVIDIA is really falling apart, it's just there are random artifacts in the image that look like errors, small one pixel high contrasted horizontal lines, and there's a different amount of them depending on if copyback or native decode is used. I don't think it comes from madVR itself because this doesn't happen on AMD with identical settings but it is weird.
Obviously it's much more visible on SD than HD.
Best possible quality is AMD DXVA2 with madVR, but if it's broken on Navi I don't have much hope they will fix it seeing how slow they are at fixing even more important stuff.

[Calvi]

Try the 1050 with CUVID for comparison,
Select also adaptive/hw and hq processing.

[huhn]

HQ processing is not supposed to work on windows 10 with CUVID AFAIK.

and AFAIK HG processing is the "IVTC" option in the driver.

[SamuriHL]

Oh God what have they broken now??

I mean, oh, GOOD, a new driver! LOL

[VBB]

Something new to play with tonight. Samuri, you go first!

[SamuriHL]

Joy. LOL Yea, I'll check it after work. They better not screw this up. I got (some) LOTR movies to watch!

[kostik]

The new driver works great with RTX 3080 . Watched 4k HDR movie with all settings maxed on madVR and works great (on 4k resolution RGB 10bit 120Hz) . This 3080 RTX is a beast. Max GPU Temp reached to 61C after 1 hour of watching a movie

The only 'issue' I see is the same as previous driver. When you bring the seek toolbar up, the colors become strange (but no flashes as before), once the seek toolbar is down, the colors are back as they should. Also there are flashes when you open madVR settings and when you change a setting in madVR while playing HDR but they are fast

[Manni]

The new driver works great with RTX 3080 . Watched 4k HDR movie with all settings maxed on madVR and works great (on 4k resolution RGB 10bit 120Hz) . This 3080 RTX is a beast. Max GPU Temp reached to 61C after 1 hour of watching a movie

Apparently you might not be getting RGB at 120hz, there is a known issue in the release notes stating that any content above 100hz switches automatically to YCC 4:2:2.

I haven’t verified as I have no need for anything above 60hz, but worth checking in your case.

[kostik]

Apparently you might not be getting RGB at 120hz, there is a known issue in the release notes stating that any content above 100hz switches automatically to YCC 4:2:2.

I haven’t verified as I have no need for anything above 60hz, but worth checking in your case.

When playing HDR in madVR ?

I have 65CX and when I check in the hidden HDMI diagnostic menu , I definitely see that I am on RGB 10BIT and 120HZ . Also madVR report that I am on RGB 10bit NVHDR .

I also know for sure that I am on 120Hz because the scrolling on the browser is super smooth .

I know this issue you talk about. I had this when I had GTX 1080 and was using the the DP->hdmi 2.1 adapter from Club3D , it was always defaulting to ycbcr 4:2:0 even when choosing 10bit RGB but I had a workaround to fix this (enable or disable AMD FreeSync or Game Response and then change to 10bit RGB ). But now, there are 0 issues of this sort (with RTX 3080) . Maybe its still an issue on 10XX and 20XX series but works perfectly on RTX 30XX

I use 120Hz because I love the super fluid and smooth browsing and also for gaming of course . I just leave it on 120Hz when playing movies using madVR because I am too lazy and also see no benefit changing the display to 60 or 24Hz.

For movies, I also don't see any improvement or change in motion using 120Hz, its same as 60hz

[SamuriHL]

I can't possibly stress enough how jealous I am of that 3080 on that panel. DROOL.

[Crimson Wolf]

When you bring the seek toolbar up, the colors become strange (but no flashes as before), once the seek toolbar is down, the colors are back as they should.

I don't have HDR, but I do have a 10-bit panel, so I encountered this while using the seek bar which causes to go into 8-bit mode when using it. I know the "enable automatic fullscreen exclusive mode" isn't true exclusive mode anymore in the latest windows 10 version, but it switches into the mode extremely quickly and if you use the "show seek bar" from the exclusive mode, madvr doesn't switch from 10-bit to 8-bit while using the seekbar. Can you test if HDR mode is fine if you try this method? Just curious. The seek bar in exclusive mode is different.

[kostik]

I don't have HDR, but I do have a 10-bit panel, so I encountered this while using the seek bar which causes to go into 8-bit mode when using it. I know the "enable automatic fullscreen exclusive mode" isn't true exclusive mode anymore in the latest windows 10 version, but it switches into the mode extremely quickly and if you use the "show seek bar" from the exclusive mode, madvr doesn't switch from 10-bit to 8-bit while using the seekbar. Can you test if HDR mode is fine if you try this method? Just curious. The seek bar in exclusive mode is different.

Enable automatic fullscreen exclusive mode is disabled on my madVR setup and 'show seek bar' is ticked under rendering->exclusive mode .

What option to turn on\off to test what you ask for ?

BTW I use the latest madVR 114 ver.

[VBB]

@ Crimson Wolf - It does work like you described in both SDR and HDR modes. With proper FS exclusive mode, which you get when checking "disable fullscreen optimizations" in the properties of an executable, the panel will not switch into windowed 8-bit mode. The disadvantage is that you don't get the player's seek bar with chapters, etc. Either way works for playback, though.

In my opinion, it's best to leave the compatibility settings alone, and uncheck "automatic fullscreen exclusive mode" in madVR. FS exclusive doesn't work anymore like it used to, and has no real advantages over windowed FS.

[chros]

I just leave it on 120Hz when playing movies using madVR because I am too lazy and also see no benefit changing the display to 60 or 24Hz.

For movies, I also don't see any improvement or change in motion using 120Hz, its same as 60hz
Yes, 120 Hz is almost good for everything except for 25/50 fps content, that saves a lot of refresh rate switching.
What interval does madvr report for frame drops with 23p at 119/120Hz without SmoothMotion?

[VBB]

Latest Nvidia drivers are good. No new issues to report at this time.

[Magik Mark]

Using DSPlayer, the driver doesn't switch back to sdr after playing a HDR video.
Do you have any work around?

[kostik]

Yes, 120 Hz is almost good for everything except for 25/50 fps content, that saves a lot of refresh rate switching.
What interval does madvr report for frame drops with 23p at 119/120Hz without SmoothMotion?

1 frame repeat every 41 seconds and it grows. This is what you meant ?
If not, where do I see or how do I calculate the interval ?

[VBB]

Not owning a C9/CX myself, I would assume you want to still use refresh rate switching at 120Hz. I'd use 2160p100, 2160p119, 2160p120. madVR should then switch to 100 for 25/50 fps content, 119 for 23/29/59 fps content, and 120 for 24/30 fps content. That should cover all bases.

[VBB]

Using DSPlayer, the driver doesn't switch back to sdr after playing a HDR video.
Do you have any work around?

Mark, that's one of the ongoing issues. The workaround would be to have the player exit/close after each playback.

[chros]

1 frame repeat every 41 seconds and it grows. This is what you meant ?
If not, where do I see or how do I calculate the interval ?
Yes, that is what I meant, thanks. That's bad, how much is it e. g. after 3 minutes of playback? And as VBB says, how much is it at 119Hz?

[Asmodian]

With a 3090 and a CX I don't have a 119 Hz option, nothing above 60 Hz besides 100 and 120 Hz, both in the PC section.

Smooth motion basically required for >60Hz. :p

[VBB]

Oh, wow. That's a big oversight on LG's part to not include that in the EDID. Shouldn't come as a surprise, I guess, when all people care about these days is gaming ;)

[chros]

CRU to the rescue? :D

[Manni]

CRU to the rescue? :D

CRU doesn’t seem to work with the latest drivers. At least not in my super brief tests. Will test more after I build my Ryzen rig this week-end...

[mpchelp10]

MPC-HC version 1.9.8
madVR version v0.92.17

madVR is set to "passthrough HDR to display" with "send HDR metadata to display" checked

Monitor: Lenovo Legion Y27q-20
GPU: RTX 2070 Super
Connected via display port

Windows "Use HDR" and "Stream HDR Video" settings are both on.

When I try to play a HDR video with MPC-HC + madVR, the colors are extremely dark. Watching HDR videos in YouTube on Chrome works fine. Is this a driver related issue or did I configure something wrong?

[el Filou]

CRU doesn’t seem to work with the latest drivers. At least not in my super brief tests.I've something in the release notes like "admin rights now required for EDID override", maybe this is related, they changed something in the driver maybe? I've never used CRU so it's just an idea.Try the 1050 with CUVID for comparison,
Select also adaptive/hw and hq processing.So:
- Using CUVID just for decode and letting madVR handle deinterlacing produces the same artifacts as DXVA2 copyback
- Enabling CUVID deinterlacing in LAV produces no artifacts at all, this is the best mode but unfortunately this means I can't control deint anymore with madVR hotkeys and have to open LAV settings to switch between film/video which is a pain.
So it looks like the artifacts come from something in madVR, as they don't show up using other renderers or with CUVID deint. Maybe madshi isn't using the correct DXVA deint mode. I will try to do more tests with DXVA Checker.
In both scenarios, the deinterlacing quality itself (apart from the artifacts) isn't any better than with DXVA (no 'HQ processing' anymore in current LAV).
I've also tested the other software deint options in LAV, YADIF is horrible and has even more artifacts than madVR DXVA, probably because it's not motion adaptive. Weston is really good, I'd say better quality than NVIDIA hw deint but still not as good as AMD's. Unfortunately uses a lot of CPU on my old system, and same inconvenience as CUVID which is not controllable from madVR.
I hope when madshi implements his planned motion adaptive deint for the Envy, that he asks us to test it. :D
I could also re-encode all my SD interlaced stuff to 50p on my desktop using AMD's deinterlacing, SDTV won't last eternally, but it takes time I'd rather use for something else.

[huhn]

madshi is aware of issues with DXVA deint he simply doesn't know why they exist:
http://bugs.madshi.net/view.php?id=652

maybe forcing YV12 instead of NV12 helps or so.
but i'm not even close to what you call artifacts...

[Manni]

I've something in the release notes like "admin rights now required for EDID override", maybe this is related, they changed something in the driver maybe? I've never used CRU so it's just an idea.

Thanks, I saw that but I'm an admin so shouldn't be the reason.

I'll do more tests when my rig is up.

[VBB]

Windows "Use HDR" and "Stream HDR Video" settings are both on.

Turn those off and let madVR do the switching. Let us know if that fixes it.

[mpchelp10]

Turn those off and let madVR do the switching. Let us know if that fixes it.
I'm still getting really dark colors. Windows HDR is off and hdr in madVR has been set to "let madVR decide".
Here are my NVIDIA control panel settings & madVR info (Ctr+J)
https://i.imgur.com/2WIpVsz.png

HDR playback using my laptop connected to my monitor does work, so I'm 100% sure this is a software/driver issue...

Something interesting I noticed:
When I press ALT+Enter to make MPC-HC fullscreen, there is a brief "hang"/lag before it actually enters fullscreen. During this brief delay, I can see the actual HDR colors/brightness, but once entering fullscreen it goes back to being super dark.

[VBB]

I'm still getting really dark colors.

One thing I would change still is to set madVR to 10-bit. Your monitor seems to handle that well, so you might as well use it throughout the video chain. Also, you should enable smooth motion in madVR, since you're outputting at 165Hz. This will remove any judder.

As for why HDR is not working properly, that's a tough one. Are you on Windows 10, and if so, which version? Judging by your screenshot, LAV is set to "none" for the decoder. You could try switching that to D3D11 native (automatic) to see if that makes a difference.

Lastly, what's your Nvidia driver version? Try the latest, if you haven't already, to see if that makes a difference.

[SamuriHL]

Speaking of drivers I forgot to post that the latest Nvidia driver has proper metadata. Basically no change from the previous driver in any behavior.

Sent from my SM-G975U using Tapatalk

[mpchelp10]

One thing I would change still is to set madVR to 10-bit. Your monitor seems to handle that well, so you might as well use it throughout the video chain. Also, you should enable smooth motion in madVR, since you're outputting at 165Hz. This will remove any judder.[/spoiler]
"the native display bitdepth is:" is set to "10 bit (or higher)" under Devices->Properties

[QUOTE=VBB;1929890]As for why HDR is not working properly, that's a tough one. Are you on Windows 10, and if so, which version? Judging by your screenshot, LAV is set to "none" for the decoder. You could try switching that to D3D11 native (automatic) to see if that makes a difference.
Windows 10 Pro 19042
MPC-HC>Options>Internal Filters>Video Decoder
It was originally set to D3D11. Changing to DXVA2 (copy-back) or DXVA2 (native) don't fix anything.

Lastly, what's your Nvidia driver version? Try the latest, if you haven't already, to see if that makes a difference.
457.30
My laptop is running a much older version of NVIDIA drivers...

One more thing I noticed: When I play the HDR file with MPC-HC, it's not just MPC-HC that ends up with darker colors - my entire monitor seems to go into some darker display mode...

[Calvi]

- Enabling CUVID deinterlacing in LAV produces no artifacts at all, this is the best mode but unfortunately this means I can't control deint anymore with madVR hotkeys and have to open LAV settings to switch between film/video which is a pain.

I just leave it on Video and it seems to always switch correctly itself. i.e. if I play film it does not deinterlace but if DTV it switches on automatically (I assume decided within nvidia hw).

I find CUVID with MadVR also to have the least artifacts.

[SweetLow]

That's a big oversight on LG's part to not include that in the EDID.
Even C9 version has this modes in EDID. Why do you think that CX does not have?

[VBB]

Are you saying the C9 has 119 as an option? That's great then, and even more of a head-scratcher as to why the CX doesn't.