LAV Filters - DirectShow Media Splitter and Decoders

[Liandri]

Thank you, both worked. (I feel stupid, again)

As for editing mkv, thanks but it's not a very good idea for me since the files need to be seeded.

[Damien147]

Hello!

DXVA2(copy-back) slowdown/''skips'' with HEVC.Other non HEVC sources play smoothly.
DXVA(native) plays smoothly too HEVC with EVR.
Any solutions for DXVA2(copy-back)?

lav 0.69
GPU:RX 470(amd) with latest drivers.

[nevcairiel]

Copy-Back has historically not been that great on AMD, so either it works or it doesn't, can't really change anything either way.
Especially when handling 4K 10-bit content you need a lot of memory bandwidth just to download and upload the image again, which AMD cards don't seem to like doing very much.

[nussman]

Did you see my report? https://forum.doom9.org/showthread.p...95#post1798395
I am not sure what is going on, but HEVC 4K 10bit works fine with madVR for me (AMD RX460 4GB) if I disable P010 and P016 output in LAV.
CPU load decreases from over 70% to ~35% without P0 output.

[nevcairiel]

All that would do is lower the memory bandwidth requirement on uploading the images back to the GPU, which I suppose can already help. LAV performance is not really different between NV12 or P010 output, both are similar in efficiency. If anything P010 should be slightly more efficient since its the native GPU format and doesn't require dithering. But its double in size so maybe that balances out.

Make sure LAV is using direct mode (ie. decoder is "dxva2cb direct"), but if its using P010 or NV12 output there rarely is a reason why it wouldn't use direct mode.
Of course it goes without saying to avoid any software post-processing filter like ffdshow or others, as those will eat your performance on 10-bit content.

[nussman]

"DXVA2cb direct" is used.
Double in size ... mhm sounds like a problem with the memory bandwith? I will check this.

With EVR / EVR Custom LAV outputs P010 or NV12 @default for HEVC 10bit content?

[nevcairiel]

EVR doesn't accept P010, so it would use NV12 then.

[Damien147]

Originally Posted by nevcairiel

Copy-Back has historically not been that great on AMD, so either it works or it doesn't, can't really change anything either way.
Especially when handling 4K 10-bit content you need a lot of memory bandwidth just to download and upload the image again, which AMD cards don't seem to like doing very much.

The thing is that another user with same GPU(RX470) reported that DXVA2(copyback) works for him with 10bit HEVC.I tried Display Driver Uninstaller for that reason but no luck,same result.With Madvr it's worst than EVR but the problem stays.So to conclude there's nothing you can do?

Originally Posted by nussman

Did you see my report? https://forum.doom9.org/showthread.p...95#post1798395
I am not sure what is going on, but HEVC 4K 10bit works fine with madVR for me (AMD RX460 4GB) if I disable P010 and P016 output in LAV.
CPU load decreases from over 70% to ~35% without P0 output.

I tried what you said about disabling P010,P016 but result stays the same.The point is to make DXVA2(copyback) play smoothly without these workarounds.RX470 4GB here and I can't get smooth 10bit HEVC playback.EVR with DXVA2(native) gives smooth 10bit HEVC playback.

[nevcairiel]

Originally Posted by Damien147

The thing is that another user with same GPU(RX470) reported that DXVA2(copyback) works for him with 10bit HEVC.I tried Display Driver Uninstaller for that reason but no luck,same result.With Madvr it's worst than EVR but the problem stays.So to conclude there's nothing you can do?

If it works for some users but not for you, then its most likely something in your system, so there is even less we can do.

[Damien147]

Yes,it points my system if that happens but I don't know what else to do beyond Display Driver Uninstaller.Maybe format(<---boring procedure)
Thank you very much for answering.Have a nice day.

[nussman]

Originally Posted by Damien147

The point is to make DXVA2(copyback) play smoothly without these workarounds.

The point is to understand what happens here and help the developers to fix it in right place.

[NikosD]

Originally Posted by Damien147

Hello!

DXVA2(copy-back) slowdown/''skips'' with HEVC.Other non HEVC sources play smoothly.
DXVA(native) plays smoothly too HEVC with EVR.
Any solutions for DXVA2(copy-back)?

lav 0.69
GPU:RX 470(amd) with latest drivers.

Just tried various 10 bit 4K HEVC sources with MPC-HC using LAV 0.69 and EVR-CP and had no problem with my RX 470 in DXVA Copy-back (Direct) mode.

The performance looks like the same of DXVA native, but of course I can see spikes of 100% GPU utilization and full clock speed for core/ memory (1300MHz/ 2000MHz)

[nussman]

Did you try madVR with P010 output?

[Aleksoid1978]

Originally Posted by nevcairiel

EVR doesn't accept P010, so it would use NV12 then.

On Win10 - EVR/EVR Custom accept P010. Checked on Nvidia 960/AMD RX460.

[Damien147]

Originally Posted by nussman

The point is to understand what happens here and help the developers to fix it in right place.

We are saying the same thing.

Originally Posted by NikosD

Just tried various 10 bit 4K HEVC sources with MPC-HC using LAV 0.69 and EVR-CP and had no problem with my RX 470 in DXVA Copy-back (Direct) mode.

The performance looks like the same of DXVA native, but of course I can see spikes of 100% GPU utilization and full clock speed for core/ memory (1300MHz/ 2000MHz)

Oh boy...another one with RX470 and smooth playback with 10 bit 4K HEVC and dxva2(copy-back).
Can you tell which driver version you are using?I'm gonna try that.I am with 17.2.1 WHQL that came out yesterday and I can't get smooth playback with 10 bit 4K HEVC and dxva2(copy-back).Picture slows down,framerate drops.

p.s. Γεια σου Ελλαδάρα!!

[NikosD]

17.2.1 beta, I think there are no differences with WHQL.

I don't think that really matters but my RX 470 as I've already posted has 2000MHz GDDR5 memory and 1300MHz GPU clock.

Win 10 x64 of course.

Try to use the clean install of AMD drivers which uninstalls previous drivers and not DDU.

But it seems as an issue of your system.

P.S
Δεν νομίζω να καταλαβαίνει κανένας τι γράφουμε στα Ελληνικά και υπάρχει κανόνας να γράφουμε Αγγλικά εδώ, αλλά δεν άντεξα την πρόκληση

[CruNcher]

Originally Posted by Aleksoid1978

On Win10 - EVR/EVR Custom accept P010. Checked on Nvidia 960/AMD RX460.

It would be rather bad if it wouldn't so DXVA also should work with P010 already

Though definitely some of those things that will only work with Windows 10 EVR Render Backend and you need other ways (APIs) todo it pre Windows 10 with acceleration though of course you never gonna get it to work under DWM windowed so in a browser and HTML5 you gonna fail completely with 10 Bit support under Pre Windows 10 when using the Native Microsoft APIs.

don't think that really matters but my RX 470 as I've already posted has 2000MHz GDDR5 memory and 1300MHz GPU clock.

Nvidia allways had the better DCE and aside that PCIE-Compression only VEGA is about maybe to change that finally the advances in Polaris where not enough

Also because the overhead of AMDs Driver is overall higher it can make a difference which advanced CPU you have and which instructions AMD makes use of

We should see further improvements overall here due to the tuning for Zen which is partly also gonna effect the Intel overall Performance due to the similiar nature of both compared to the old Athlons and Bulldozer

[clsid]

Perhaps it is a DPC latency issue?
http://www.resplendence.com/latencymon

[Damien147]

Originally Posted by NikosD

17.2.1 beta, I think there are no differences with WHQL.

I don't think that really matters but my RX 470 as I've already posted has 2000MHz GDDR5 memory and 1300MHz GPU clock.

Win 10 x64 of course.

Try to use the clean install of AMD drivers which uninstalls previous drivers and not DDU.

But it seems as an issue of your system.

P.S
Δεν νομίζω να καταλαβαίνει κανένας τι γράφουμε στα Ελληνικά και υπάρχει κανόνας να γράφουμε Αγγλικά εδώ, αλλά δεν άντεξα την πρόκληση

Was with 17.2.1 beta too and the problem was there.Don't update to WHQL if you want,it's the same thing. About clean install I've tried that and THEN I tried DDU to see if there is a problem.Nothing changes.I have lower clocks but yes you are right.It seems that it is my system and not in general.





P.S.
Σιγά,ένα γειά ανταλλάξαμε.Καλή συνέχεια.

Originally Posted by clsid

Perhaps it is a DPC latency issue?
http://www.resplendence.com/latencymon

Gonna have to learn to use that I guess.

[CruNcher]

Originally Posted by clsid

Perhaps it is a DPC latency issue?
http://www.resplendence.com/latencymon

I bet its a inefficiency of overall CPU optimization albeit the CPU he uses and Chipset is pre Sandy Bridge era that wont cut it anymore especially copy back with AMDs Driver overhead and not so optimized DCE.

But yeah if you think about a SOC at 7W is completely destroying it overall but only with Zero Copy with Copy Back most Socs would die also for 4K Hevc.