LAV Filters - DirectShow Media Splitter and Decoders

[nevcairiel]

I just found out that my GPU Ati 4850 with 512MB Ram cannot handle DXVA-CB on

Older AMD cards suck at Copy-Back. Anything before the HD7000 series is basically unusable with Copy-Back decoding.
Especially if you try to use high framerate material, you'll really notice how the card bottle-necks the whole system.

[NikosD]

My Radeon 5750 card on PCI-E x16 has almost same performance using DXVA-CB or Native.
The problem is the high framerate.

60fps are unreachable even with HD7000 series - I haven't tested myself, I have read it though.

[mhourousha]

My HD7770 can handle 1080p60 clips with bitrate around 10Mbps.
to marsovac:you can create a custom EVR Presenter that support both DXVA2_Native and xySubfilter.

[NikosD]

And what about the clip posted above or any other AVCHD clip which usually is ~25Mbps ?

You can handle the 10Mbps 1080p60fps clip in both modes DXVA-CB and native, or native only ?

[mhourousha]

And what about the clip posted above or any other AVCHD clip which usually is ~25Mbps ?
I tested 'ducks take off(1080p version 100Mbps)'clip some months before,the decode performance is around 45fps IIRC.Native mode

[NikosD]

Thanks for the info, but "the ducks" clip is not a 1080p60fps, it's not an AVCHD clip.

Can you try the clip posted above or this one
ftp://helpedia.com/pub/multimedia/x264/testvideos/2011%20-%2002%20-%20H.264%20CPU%20DXVA%20codec%20comparison%20-%20Core2Duo%20vs%20UVD%202.2/6.Cat-1080p60fpsRef4-25Mbps.m2ts

[fagoatse]

Older AMD cards suck at Copy-Back. Anything before the HD7000 series is basically unusable with Copy-Back decoding.
Especially if you try to use high framerate material, you'll really notice how the card bottle-necks the whole system.

Shouldn't APUs be better at that task though?

[Soukyuu]

If you're talking about HSA, I'm not sure it will work out of the box. I pretty much expect the code to duplicate data in memory because it doesn't expect both the GPU and CPU to have access to the whole memory.

[mhourousha]

Thanks for the info, but "the ducks" clip is not a 1080p60fps, it's not an AVCHD clip.

Can you try the clip posted above or this one
ftp://helpedia.com/pub/multimedia/x264/testvideos/2011%20-%2002%20-%20H.264%20CPU%20DXVA%20codec%20comparison%20-%20Core2Duo%20vs%20UVD%202.2/6.Cat-1080p60fpsRef4-25Mbps.m2ts
avg 76 fps for both DXVA2_N and DXVA2_CB mode

[NikosD]

Shouldn't APUs be better at that task though?

If you're talking about HSA, I'm not sure it will work out of the box. I pretty much expect the code to duplicate data in memory because it doesn't expect both the GPU and CPU to have access to the whole memory.

Yes they should, but only in DXVA-CB, not DXVA native.
Because memory copies are faster, so I'm pretty sure that DXVA-CB should be faster than discrete cards for the same generation - at least.

But DXVA native should be the same, because basically it depends on the HW decoder.

avg 76 fps for both DXVA2_N and DXVA2_CB mode

Thanks!
It seems that the performance of HD7000 series is on par with Nvidia VP4 and my overclocked UVD2.2

Can you see the clock of GPU/Memory during benchmarking of your HD 7770?

[mhourousha]

Thanks!
It seems that the performance of HD7000 series is on par with Nvidia VP4 and my overclocked UVD2.2

Can you see the clock of GPU/Memory during benchmarking of your HD 7770?
GPU load is 30%,and the clock of core/mem is in full speed(1100/1200) during benchmark.

[NikosD]

@mhourousha

One last question:

Can you benchmark those two 1080p120fps files ?

1) http://120hz.net/hypermatrix/120fpsvideo/bf.mp4

2) http://120hz.net/hypermatrix/120fpsvideo/bf1.mp4

Can you decode them in realtime using HD 7770 (CB or N)?

[mhourousha]

@mhourousha

One last question:

Can you benchmark those two 1080p120fps files ?

1) http://120hz.net/hypermatrix/120fpsvideo/bf.mp4

2) http://120hz.net/hypermatrix/120fpsvideo/bf1.mp4

Can you decode them in realtime using HD 7770 (CB or N)?
both clips cannot reach full speed as expected
first clip:avg 77 fps
second clip: avg 79 fps
DXVA2_N

[NikosD]

It could go higher because these are easier to decode clips than AVCHD clips, VP4 has an average of 90fps.

Thanks for your time.

[marsovac]

http://120hz.net/hypermatrix/120fpsvideo/bf1.mp4

With DXVA native full speed.

With DXVA CB, extremely laggy. Like 50% dropped frames.

Phenom II X4 3Ghz, 8GB DDR3 1666 C8
ATi HD4850 512MB

[Soukyuu]

I'm pretty sure I saw someone reporting the same before and the answer was that ati's 4000 series don't have high enough memory bandwidth to cope with copying data back and forth.

[nevcairiel]

Its a problem with the architecture more then anything, copying from the GPU to system memory is just dead-slow on anything before HD7000. Its not a common use-case for a graphics card tbh, unless you're using it for compute - and that wasn't very big back then.

[marsovac]

So as I was told: if I want to get DXVA native I need to connect LAV video to a compatible renderer without any filters in between.

And I did that. Connecting directly to EVR.

Enum of all filter in my graph:
Renderer: EnhancedVideoRenderer
Default DirectSound Device
LavVideo
LavAudio
LavSplitter


I'm doing this:

if (lavVideoSettings.CheckHWAccelSupport(LAVHWAccel.HWAccel_DXVA2Native) != 0)
{
hr = lavVideoSettings.SetHWAccel(LAVHWAccel.HWAccel_DXVA2Native);
hr = lavVideoSettings.SetHWAccelResolutionFlags(LAVHWResFlag.SD | LAVHWResFlag.HD | LAVHWResFlag.UHD);
}


And connecting renderer and video:


hr = m_graph.Connect(DsFindPin.ByDirection(_video, PinDirection.Output, 0),
DsFindPin.ByDirection(_renderer, PinDirection.Input, 0));


Result: no luck:
http://i.imgur.com/qeOoRl1.jpg

I tried with various files which should work with DXVAn, none worked.

Any other suggestions?

EDIT: After removing my custom presenter from the EVR DXVAn is applied.
What are the requirements to have DXVAn and a custom EVR presenter?

[NikosD]

You have a very old version of LAV video.

Is it on purpose ?

I don't know though if a newer version solves your problem.

[marsovac]

You have a very old version of LAV video.

Is it on purpose ?

I don't know though if a newer version solves your problem.

It seems that a custom presenter that I put on the EVR makes it fall back to avcodec.

Without a presenter i get dxvan but of course I cannot make that useful in any way :D

PS. I'm using LAV 0.60.1, but I don't know why the dialog says otherwise.

[nevcairiel]

I haven't actually written a presenter myself, but I assume it needs to somehow handle the DXVA2 mode. I'm afraid I cannot help you there.

PS:
The property page thing is a common issue if you use a unregistered/private copy of LAV Video, and having another version of it installed in the system - if thats what you're actually doing. The way the code looks up the COM object for the property dialog, it'll always use the system registered page.

[marsovac]

The property page thing is a common issue if you use a unregistered/private copy of LAV Video, and having another version of it installed in the system - if thats what you're actually doing. The way the code looks up the COM object for the property dialog, it'll always use the system registered page.

That's what I am doing. Loading the COM by using the class factory in the DLL.

It probably uses the system registered version.

About the presenter... I'll review the code from MPC-HC and see what I am missing...

Thanks.

[mhourousha]

What are the requirements to have DXVAn and a custom EVR presenter?
refer to the following tutorial
http://msdn.microsoft.com/en-us/library/windows/desktop/bb530107%28v=vs.85%29.aspx
if you want to display subtitle,you can use xySubfilter.
Create a filter that implement ISubRenderConsumer declared in http://madshi.net/SubRenderIntf.h.connect it to xySubfilter,then you can request subtitle img(when the Presenter queue a video sample) using ISubRenderProvider provided by xySubfilter, receive and store the img data in a queue ,when the Presenter Draw a queued video sample, fetch the subframe with corresponding timestamp from the queue, copy it to a Texture ,blend with video frame.that's all.

[marsovac]

refer to the following tutorial
http://msdn.microsoft.com/en-us/library/windows/desktop/bb530107%28v=vs.85%29.aspx
if you want to display subtitle,you can use xySubfilter.
Create a filter that implement ISubRenderConsumer declared in http://madshi.net/SubRenderIntf.h.connect it to xySubfilter,then you can request subtitle img(when the Presenter queue a video sample) using ISubRenderProvider provided by xySubfilter, receive and store the img data in a queue ,when the Presenter Draw a queued video sample, fetch the subframe with corresponding timestamp from the queue, copy it to a Texture ,blend with video frame.that's all.

Very good info mhourousha.
Might get it done in a few days if C++ doesn't create problems :)

Thank you!

[mzso]

Hello!

This little test video causes a crash with latest LAV video and mpc-hc or potplayer. With or without HW acceleration. With Potp internal decoder and MPV it plays fine.

https://drive.google.com/file/d/0ByfdfPvnoDuzSlZqY1JxeDdXNkU/edit?usp=sharing

[JEEB]

mzso, Did you try current git HEAD or the latest release? At least the 32bit current git HEAD seems to work fine for me with a standard MPC-HC setup (splitter->decoder (LAV Audio does downmixing to 2ch)->EVR|DirectSound).

[mzso]

mzso, Did you try current git HEAD or the latest release? At least the 32bit current git HEAD seems to work fine for me with a standard MPC-HC setup (splitter->decoder (LAV Audio does downmixing to 2ch)->EVR|DirectSound).

Hmm....
Could it be that it's a LAV Video+madVR issue?
Apparently changing either circumvents the issue.

[nevcairiel]

The crash is inside madVR. I believe I reported a similar issue to madshi before, but to be safe you should tell him again.

[marsovac]

Thanks to mhourousha I got DXVAn to work with my presenter.

Now the big step, integrate xyVSFilter into the presenter instead of attaching to LAV...

One strange thing I find out is:

- In DXVA native the bf1 video lags badly (bf1.mp4 (http://120hz.net/hypermatrix/120fpsvideo/bf1.mp4))
- IN LAV software mode it works without any trouble

Is this supposed to happen for 120 FPS videos?

I have my both build (dxva, software) in this zip. Can anyone test and see if they have the same issue?

http://marino.boletus.hr/TestBuilds.zip

You can check that dxva is enabled in the lav config dialog in the tray icon.

PS. I tried to play the video in MPC-HC and it lags in DXVAn, even tough a little less then in mine.

[mhourousha]

no AMD GPU can handle 1080p120 in Hw decoding mode currently，so slow-down will occur.
but your player produce some additional lag IMHO.the display of stats flash all the time,and player crushed near the end of clip :(
I suggest you try using seperate device,seperate thread for rendering(UI,frame,stat etc).the D3D9EX has 'sharing resource between device'feature.so EVR Presenter could only draw Frame to a shared RenderTarget,then the rendering thread and device use this RT to draw the front end.

[Aleksoid1978]

no AMD GPU can handle 1080p120 in Hw decoding mode currently,so slow-down will occur.

ATI 7750 play perfect at 1080p120 in DXVA 2.0 mode.

[mhourousha]

ATI 7750 play perfect at 1080p120 in DXVA 2.0 mode.
My HD7770 simply can't.

[wanezhiling]

My HD7770 simply can't.

Win7 or Win8/8.1?

Try to disable aero if is win7

[NikosD]

ATI 7750 play perfect at 1080p120 in DXVA 2.0 mode.

I think you are the first ever reporting this.

Are you sure you are not using your CPU and/or you are actually decoding 120fps ?

Do you play 4K H.264 video too, in DXVA and 7750 ?

[mhourousha]

Win7 or Win8/8.1?

Try to disable aero if is win7
My OS is win7.aero is already disabled when benchmarking.
LAV DXVA2N:avg~79fps
MS Decoder MFT: avg~81fps

[NikosD]

Check your Catalyst video settings and make sure that everything - besides "Automatic Deinterlacing" - is disabled.

[marsovac]

no AMD GPU can handle 1080p120 in Hw decoding mode currently，so slow-down will occur.
but your player produce some additional lag IMHO.

Ok that explains what I am having with HD4850.

The additional lag I'm having is exactly what you said.

I am actually drawing the EVR surface into a separate render target, and then I'm passing that into my application over to a WPF D3DImage, which is where the image is presented in the UI.

There is more overhead then just drawing the surface on a form, but the main issue is thet the whole UI in WPF is GPU accelerated, so my UI introduces even more lag on the alredy stressed GPU.. You can see that when hovering in and out of the window during the play of that video. The transparency animation effect is lagging also.

But i'm not concerned about that. I don't want to make a player in GDI or native DX since it requires too much work to make the UI nice :)

About the player crash: I may have some leakage left in those builds since I was experimenting a lot of things to make the presenter handle DXVA2. Will check it.

[mhourousha]

Check your Catalyst video settings and make sure that everything - besides "Automatic Deinterlacing" - is disabled.
turn off all the post processing eff will lower the GPU load(30%->4%),but also lower the GPU core clock(1100Mhz->400Mhz in most time),so there is a small drop of the fps(80fps->76fps)

[NikosD]

Strange.

In normal playback that would be normal.

But in benchmarking mode, the driver should clock the GPU as high as possible.
400 MHz is the standard clock of every UVD decoder, since Radeon 5000 series (UVD2.2)

I don't know if there is a way to find the real UVD clock, instead of GPU clock, although it should be the same if the GPU is using one clock for both.

[wanezhiling]

It is still a long way for GCN to achieve NV VP5 capability in decoding, let alone the decoding *monster* Intel, sigh.

[nevcairiel]

I have a NVIDIA 750 here for testing (with the new decoder, VP6 if you want to call it that), didn't install it in a system yet though, we'll see how well it performs and blows away AMD even more.
NVIDIAs claims would put it on par with Intel, but we'll see what really happens.

[NikosD]

Nice!

I bet the performance of 750 video decoding will be half yours Haswell iGPU.

[DragonQ]

Quite interested in results for the Maxwell GPUs; nVidia's claims are rather outlandish (8-10x faster video decoding, 2x faster video encoding) and the new range of cards should be really nice with their much lower temperatures and significantly reduced power consumption.

[NikosD]

I have read that claim of x8-x10 video decoding performance and I think there is a HUGE misunderstanding.

The x8-x10 claim is referring to real time decoding, not the previous generation VP5.

VP6 (?) is x2 - x2.5 faster than VP5, which translates to 4K60fps video decoding capability.

My bet.

[DragonQ]

Yeah makes more sense. 2180p @ 60 fps AVC would be great but it's a shame they only have partial HEVC acceleration. Most 2180p content will likely be HEVC in the near future.

I'd be interested to see if they've updated their deinterlacing algorithms at all but I guess with the eventual move to 2180p they don't need to worry much about it any more.

[nevcairiel]

First testing shows its around 3x as fast as VP5, reaching close to 400 fps in various 1080p samples, or 80 fps on 4K. But that was only a quick and dirty test with very limited samples, more later, and maybe I can also squeeze more out of it by modifying the decoder a bit to be more aggressive.

[NikosD]

Not a surprise for me.

I was thinking about 4K@70 fps, but Nvidia decided to give it more room just in case.

They have fully covered HDMI 2.0 which is 4K@60.

Still 4K@80 fps is about half decoding performance of Haswell iGPU.

[nevcairiel]

My Haswell did 100 fps on the 4K sample I tested (and its usual 500 on the 1080p samples), so hardly half.

[NikosD]

Have you disabled all VPP functions in drivers ?

Because even the hardest 4K ducks@370Mbps can be decoded by Ivy HD4000 at about ~130 fps.

So your 20 EU's iGPU should go higher.

In renderless mode using decoder only could go more than 150fps.

[GTPVHD]

Thanks nev for testing the new Maxwell hardware decoder, seems to be a big jump from 120FPS in VP5 to 400FPS & from 4K24P to 4K80P in Maxwell. Intel's 4K120P decoder is still pretty ridiculous in performance.