AMD, Intel and Nvidia driver issues and last recommended version

[NikosD]

CUVID works always at maximum 3D clocks, all the time and it feels like a tad faster than any other decoder.

Also, CUVID exposes more codecs than D3D11 or DXVA2 like MPEG4 ASP (divx) and others, although not very common nowadays.

[el Filou]

Originally Posted by ryrynz

The result isn't different because you haven't looked at your clock speeds at all. What do you meant by nothing changed? You're just looking at render times yeah? Trust me, you have zero reason to use CUVID.

garson looked at the clocked speeds and said they were the same, so presumably it's a situation where rendering itself maxes out the GPU even with DXVA. With such a low power card, maybe it's possible the API could make a small difference even if it's a edge case?

[garson]

Originally Posted by el Filou

garson looked at the clocked speeds and said they were the same, so presumably it's a situation where rendering itself maxes out the GPU even with DXVA. With such a low power card, maybe it's possible the API could make a small difference even if it's a edge case?

Correct.
I run GPU-Z for 30 minutes and logged results into log file. Then I compared log files, everything looked pretty much same, except for video memory usage (with D3D11 it was ~850MB and with CUVID ~900MB).

[littleD]

But in dx11 mode, do you use dx11 native or dx11 cb (by choosing card specifically)? Because second way is more memory heavy and thus might choke slower cards with low membus.

[clsid]

For a proper test you would need to force the GPU to use a fixed clock speed. NVIDIA Inspector tool allow you to do that.

[Calvi]

I have tried every possible combination I can of CUVID vs DXVA CB & DXVA Native with LAV Filter and Zoomplayer and...

EVR, MadVR, MadVR Deinterlacing ON/Off, Video Forced, Film, Centre or All, Exclusive, Windowed, DX11 Presentation On/Off and...

I cannot get DXVA to deinterlace the same as Cuvid.

Only DXVA Copyback with Software Deinterlacing (Westin or Yadif) looks similar on my Australian DTV broadcasts but this uses more CPU.

Windows also reports ~40% GPU useage for most of these tests (Much lower for EVR of course as MADVR is not scaling etc). I understand the GPU clocks might still be higher for CUVID but its also worth noting my RTX fans do not spin up for any of the tests, CUVID or not.

I can upload a test video if anyone is interested. Perhaps our broadcasts have strange mixed mode but the combing is quite pronounced if not using CUVID or Software Deinterlacing.

Maybe there is something in my setup that does not turn the VA HW deinterlacing on in the renderer?

MADVR shows deinterlacing as on (which looks the same as when its off).

[huhn]

and you are sure you are not using film mode deinterlacing?

[Calvi]

Not sure of anything at the moment.

I tried force film and force video and neither makes any difference to the combing during motion.

I also did a bunch of tests with GPU-Z tonight and found that CUVID does not set the clock to max. It does set it higher than the other modes but it still scales with workload.

If I run with Zoomplayer as a small window the clock runs at about 1400, at full screen it goes to 1900.
Also the power used is only ~50w with a small window and goes up to 175w at full screen (3840x2160).

The extra load I would mainly expect to be MadVR scaling workload. (Checking with evr will confirm this next time I have a play around)

So comparing CUVID with dxva2 (and d3d11 for that matter) all ended up around 70% gpu load, all about 160-175watts and all setting the clocks to near max to achieve the MadVR settings I have.

The only difference I am left with is CUVID is not giving me any deinterlacing artifacts.

[Calvi]

The test clip I am using is a TV show which then has the lotto numbers rolling across the bottom of the screen. These numbers break apart into combed half frames if not using CUVID.

It’s possible that the lotto overlay is a different cadence to the main show. I will need to dissect the clip some more to see what’s actually there. But CUVID definitely handles it correctly whatever it is.

Software deinterlacers also work fine on it.

[littleD]

Can you upload a sample?

The best way to test HW DXVA deinterlacing us using EVR renderer, or EVR custom if you need some statistics.

[Calvi]

Here is a sample (~110mb), at the 27sec mark the lotto numbers roll in which show combing artifacts with all but CUVID or software deinterlacing for me.

https://1drv.ms/v/s!AhrKCcZdae3QgY8E...CHmEA?e=X1c4HK

Note: My setting were:
Splitters : I tried TSREADER from Media Portal & LAV Splitter
Decoders: LAV video decoder DXVA2 CB, DXVA2 Native, D3D11 Native, CUVID
Renderers: EVR, MadVR

Player: zoomplayer.

[littleD]

Originally Posted by Calvi

Here is a sample (~110mb), at the 27sec mark the lotto numbers roll in which show combing artifacts with all but CUVID or software deinterlacing for me.

https://1drv.ms/v/s!AhrKCcZdae3QgY8E...CHmEA?e=X1c4HK

Note: My setting were:
Splitters : I tried TSREADER from Media Portal & LAV Splitter
Decoders: LAV video decoder DXVA2 CB, DXVA2 Native, D3D11 Native, CUVID
Renderers: EVR, MadVR

Player: zoomplayer.

I think its indeed mixed mode interlacing (mbaff). Using lavf (cb mode) and madvr on default do deinterlace correctly on intel. It turns on deinterlacing in madvr statistics. Even if you choose cb mode in lav, also choose specific gpu to decode. That makes you sure to use cb mode.

[Calvi]

Thanks for taking a look littleD.

I have repeated the same observations on a totally different machine (Laptop with Nvidia 330GT).
I also have repeated the same with MPC-HD instead of Zoomplayer and again only CUVID is 100% working for me.
I also tried LAV output at 25 vs 50 frames, MadVR again as film vs video, forcing deinterlacing in LAV etc.
Results are always the same on both machines even though a huge number of other things are different.


Here are three screen caps from MPC-HC Using GT330 and Internal LAV (0.70.2.1-git)



Note: My main system is RTX2070, driver 436.15 using LAV 0.74 and Zoomplayer 14.5 and this exhibits the same behaviour but the above tests are from my laptop with GT330, driver 341.02, MPC-HC & Internal LAV 0.70.2.1.

[littleD]

You did not metion this explicitely. Sorry thats why i ask. Did you set dxva CB AND choose nvidia adapter? Because You should not leave any automatic gpu in lav filters under "Hardware device to use"

[Calvi]

Yes, tried all combinations of that also. There is only one adapter on my HTPC. In Auto it chooses the RTX2070.
I also set it myself and the result was the same.

I repeated this with my Laptop which has 2 displays and I tried Auto (It chooses the GT330), GT330 #1 and GT330 #2. All the same results again.

Note: your adapter is intel so it may behave completely differently. I'll try my desktop when I get a chance as this has an ATI card.

[Klaus1189]

Originally Posted by littleD

You should not leave any automatic gpu in lav filters under "Hardware device to use"

Why not? What‘s the difference? Does this also affect systems with only one dedicated graphics card or only systems that also have an onboard intel hd or amd integrated graphics?

[huhn]

you should not change from automatic to something else if you don't know what you are doing.
and it would effect system with only one card you enter copyback mode with d3d11.

[Calvi]

I have also tried my Desktop now with ATI5700 (Yes its old) and of course cannot test CUVID but DXVA2 CB deinterlaces this clip *Almost* correctly.
It fails on the first lotto ball then manages to deal with all the middle ones one and then fails on the end one. Its as if the HW deinterlacer turns on late and off early. (Probably the best the HW deinterlacer in this card can do).

Note: Again if software deinterlacing - (eg Yadif) then all frames are correctly deinterlaced (unsurprisingly as this would be independent of Video Card, and Yadif is probably better than this old cards HW deinterlacing).

Note: The above tests were with EVR.

With MadVR the results are the same except Forcing Film Mode causes all the lotto balls to not deinterlace (as expected). If correctly set to deinterlace then results are identical to EVR so the deinterlacing in the renderer is working fine here. (As best as this card can do anyway).

All of the above is what I would expect for an ATI card (and explains why littleD's sees DXVA2 working for intel) but I am still left with a discrepancy between CUVID and DXVA2 for Nvidia cards. I would love to know whether I am doing something wrong or if there is an actual difference in the HW deinterlacing here?

Has anyone else tried my sample clip with a Nvidia card?

[el Filou]

I can confirm the issue with my 1050 Ti and I also have a 'solution' for you: in NVIDIA control panel, uncheck 'Use inverse telecine', and DXVA will have the same deinterlacing behaviour as CUVID.
The problem isn't really the deinterlacing quality, it's that it just not activates reliably.
The programme being played is 2:2 pulldown so no need to deinterlace it, and apparently NVIDIA doesn't reliably detect the small video insert being overlaid onto it with the lotto balls, while CUVID in video mode obviously does as it treats every frame as video.
Note that I say 'solution' because doing that will disable progressive cadence detection in NVIDIA's deinterlacer, but if you only watch modern 25 fps programming I'd say it's safe. The quality of 2:2 pulldown material will be slightly diminished however (you could see shimmering or other artifacts sometimes).
On my Radeon, the GPU detects the interlaced overlay when the 2nd ball appears (which is acceptable IMO) and correctly deinterlaces the rest after that except the white ball. With cadence detection enabled, NVIDIA detects the interlacing in the middle of the second ball's travel, but then loses it again pretty much immediately, and then does the same again with the 3rd ball.

[huhn]

using inverse telecine is considered the high quality option in CUVID which doesn't work anymore since win 8 or something like that.

nev may know more this all happend some time ago.