LAV Filters - DirectShow Media Splitter and Decoders

[Xaurus]

Well, the T series are down-clocked for less-power, aren't they?
The i3-2100 seems to be even faster then yours, but of course uses more power (65W)

I can't wait for Yvy Bridge, my HTPC is up for a rebuild. :)
Yes, the 2100 is 3.1 GHz @ 65W while the 2120T is 2.6 GHz @ 35W.

In any case I use CUVID with the GTS 450 so no problem what so ever. :)

[El Topo]

According to Intel (http://ark.intel.com/products/53414), Clear Video HD Technology is disabled.

Thats right for the Hardware encoding part, but the decoding part is the same for all Sandy Bridge CPUs.

[El Topo]

Well that clearly isn't using hardware decoding, is it? MPC-HC will say "DXVA" at the bottom if it's using hardware decoding. You can also press CTRL-J to see what rendering device and decoder are being used. LAV works with nVidia's CUDA but it doesn't work with Intel's equivalent, does it?

Personally I'm going Celeron G530 + nVidia GT430 for my HTPC. Using LAV CUVID the video card can handle all the difficult stuff with ease and the CPU is plenty for other things like DivX/Xvid. It can even handle MadVR.

Of course this isn't using HW decoding, because the Intel CPU/GPU was used with LAV Video Decoder, which doesn't support HW decoding on Intel GPUs (yet).

[DragonQ]

Of course this isn't using HW decoding, because the Intel CPU/GPU was used with LAV Video Decoder, which doesn't support HW decoding on Intel GPUs (yet).

What was your point then? nevcairiel said the G530 had hardware decoding good enough for anything, not software decoding.

[fastplayer]

Thats right for the Hardware encoding part, but the decoding part is the same for alle Sandy Bridge CPUs.
Then their info (http://www.intel.com/content/www/us/en/architecture-and-technology/clear-video-hd/clear-video-hd-technology-general.html) is inaccurate because it says Clear Video HD Technology covers HW-accelerated decoding.

[El Topo]

Then their info (http://www.intel.com/content/www/us/en/architecture-and-technology/clear-video-hd/clear-video-hd-technology-general.html) is inaccurate because it says Clear Video HD Technology covers HW-accelerated decoding.

Yes, it's inaccurate. Have a look at this thread (http://www.avsforum.com/avs-vb/showthread.php?p=21015057#post21015057) and the post linked.

[El Topo]

What was your point then? nevcairiel said the G530 had hardware decoding good enough for anything, not software decoding.

While the Intel i3 and better ("CPUs" as Nev calls them :D) are capable of getting all the work done without HW decoding, the smaller Sandy Bridges (Celeron and Pentium) are too weak for the big challenges (1080p/i). Thats why I asked Nev to support HW decoding on Intel, which is on his todo list (at the very end I guess).

[nevcairiel]

While the Intel i3 and better ("CPUs" as Nev calls them :D) are capable of getting all the work done without HW decoding, the smaller Sandy Bridges (Celeron and Pentium) are too weak for the big challenges (1080p/i). Thats why I asked Nev to support HW decoding on Intel, which is on his todo list (at the very end I guess).

Not the very end, just not at the top either. :)

[jmonier]

I'm sorry, but my brain is not wikipedia, its not archived with citations and source links for every piece of information.

My comment also only applies to Win7. Its pretty easy to test - wip up some tool that just runs an infinite loop. Its not multi-threaded, yet the task manager will show multiple cores with load instead of one being maxed out.

Well, I'm not talking about test cases. I've described a real world sample that I have experience with over a long period of time and, believe me, it works just as I said. And my experience extends from XP to Win 7 (which is what I'm currently using).

[andyvt]

While the Intel i3 and better ("CPUs" as Nev calls them :D) are capable of getting all the work done without HW decoding, the smaller Sandy Bridges (Celeron and Pentium) are too weak for the big challenges (1080p/i). Thats why I asked Nev to support HW decoding on Intel, which is on his todo list (at the very end I guess).

For now you'll want to check out egur's custom build (http://forum.doom9.org/showthread.php?t=162442) of ffdshow.

[linzki]

why can't i see any difference when i use CUVID and adaptive, 50p/60p deinterlacing.

Shouldn't the difference be like when i use Splash players motion2 60p option.

i have: Intel core 2 quad CPU Q6600 @ 2.40 GHz, 4Gb ram, GeForce GTX 460, Windows 7 32-bit.

[nevcairiel]

Hi,

http://files.1f0.de/lavf/LAVFilters-0.39-47-g121fb24.zip

This build adds support for YADIF deinterlacing.
I have tested it quite a bit already, and it seems stable enough - however i cannot shake the feeling that i'm missing something.

Limitations:
- It only works on YUV 4:2:0 or 4:2:2 (8-bit only) (before output conversion)
- It does not work with the CUVID decoder (why would you want to anyway?)

Please give it a try, and if anything goes wrong, report the problems. :)

[VipZ]

Hi,

http://files.1f0.de/lavf/LAVFilters-0.39-47-g121fb24.zip

This build adds support for YADIF deinterlacing.
I have tested it quite a bit already, and it seems stable enough - however i cannot shake the feeling that i'm missing something.

Limitations:
- It only works on YUV 4:2:0 or 4:2:2 (8-bit only)
- It does not work with the CUVID decoder (why would you want to anyway?)

Please give it a try, and if anything goes wrong, report the problems. :)

Awesome :)

I have done some basic testing with forced RGB32 output, looks good to me. I got a feeling YADIF is not multithreaded or am I mistaken.

[jmone]

I've run through my Interlacted files and it works perfectly on all of them (well with the exception of VC1-(i) of course):
1) CPU Overhead seems to be low for me (only adding and extra 2-3% on my i7-920)
2) It play "nice" with madVR's (and it's deinterlacing), eg if you toggle LAV SW Deinterlacing on/off then madVR dynamically changes as well. I was using the 50/60p Video Mode though madVR still reported the stream as 25/29.970fps

I'm really pleased with the quality and the balance with the low CPU overhead.

[nevcairiel]

I got a feeling YADIF is not multithreaded or am I mistaken.

Its not, adding that would also be kinda hard.
Neither is ffdshows YADIF, fwiw.

It does feel fast enough for me to not worry about it.

[VipZ]

Its not, adding that would also be kinda hard.
Neither is ffdshows YADIF, fwiw.
I always thought ffdshow's was and also mentions this on their wiki.

It does feel fast enough for me to not worry about it.

I have a 1080i h264 ts file which often hits 12-13% CPU, with turbo up to 4.5ghz and seems like it dropping frames.

[nevcairiel]

I always thought ffdshow's was and also mentions this on their wiki.

Hm right, i just found the code, its well hidden, as everything in ffdshow. :p



I have a 1080i h264 ts file which often hits 12-13% CPU, with turbo up to 4.5ghz and seems like it dropping frames.

I don't think its performance is really all that much dependent on the complexity of the image, although if you want, try to cut a sample of such a file.
I tested a 1080i60 h264 track, decodes at around ~260 fps without YADIF, at ~180 fps with.

Maybe in the future i might try to add multi-threading support, but right now, considering the effort it would be, i don't see the need on numbers i see here.

[VipZ]

Hm right, i just found the code, its well hidden, as everything in ffdshow. :p

:) Would that code be easy to port from ffdshow or too messy and not worth the trouble?

With the toggle option for YADIF is there any reason to not have it ticked other than to use HW deinterlacing, or is there a chance it has side affects with some progressive content?

[nevcairiel]

:) Would that code be easy to port from ffdshow or too messy and not worth the trouble?

Way too messy.


With the toggle option for YADIF is there any reason to not have it ticked other than to use HW deinterlacing, or is there a chance it has side affects with some progressive content?
It shouldn't affect pure progressive content.

[jmone]

FYI, just tested on a lowly C2D Laptop - I'm seeing a big CPU improvement with LAVVideo+YADIF over FFDSHOW+YADIF (in MC's RO setting)

[VipZ]

Way too messy.

Thanks, guessed as much. Think it would be quite rare for a video to stress a single core of a SB CPU

Which deinterlacing would take precedence if NV12 to EVR is used with YADIF selected?

[DragonQ]

Hi,

http://files.1f0.de/lavf/LAVFilters-0.39-47-g121fb24.zip

This build adds support for YADIF deinterlacing.
I have tested it quite a bit already, and it seems stable enough - however i cannot shake the feeling that i'm missing something.

Limitations:
- It only works on YUV 4:2:0 or 4:2:2 (8-bit only) (before output conversion)
- It does not work with the CUVID decoder (why would you want to anyway?)

YADIF 2x (i.e. with frame doubling)?

Also, is this 32-bit only?

[STaRGaZeR]

BIG day for LAV Video ;)

Your implementation of avfilter is so clean, me likes. Using Yadif already, will report any problems. Very good perfomance in seeks. So far the only issue I see is that it requires a video restart (or the repeat function of MPC) to apply the changes in the output mode (same/double framerate), pressing the apply button doesn't work.

And now that I'm "suffering" this, a nice feature for the future would be configurable keyboard shortcuts to enable/disable stuff on the fly ala ffdshow, without having to open the properties page all the time.

EDIT:

- It's me or LAV does a better job at yadif than ffdshow? I have to test and compare more, but quality seems a lot better in some places like fast horizontal letters in DVB commercials. It looks way better than ATI's VA deinterlacing to my eyes, while ffdshow looks worse.
- It seems that there is no way to completely disable deinterlacing. If Yadif is off, hardware takes over. There was an option to send flags downstream (on/off), right? This is needed to avoid deinterlacing on samples that are properly flagged as interlaced but the content is actually progressive, like in most DVB movies here.

[linzki]

Hi,

http://files.1f0.de/lavf/LAVFilters-0.39-47-g121fb24.zip

This build adds support for YADIF deinterlacing.
I have tested it quite a bit already, and it seems stable enough - however i cannot shake the feeling that i'm missing something.

Limitations:
- It only works on YUV 4:2:0 or 4:2:2 (8-bit only) (before output conversion)
- It does not work with the CUVID decoder (why would you want to anyway?)

Please give it a try, and if anything goes wrong, report the problems. :)

Thanks a lot Nevcairiel. Now it's working :)

[Mikey2]

What is the difference between "LAV Splitter" and "LAV Splitter Source" ?

I cannot find anything in my research, and I have switched between the two multiple times and I do not see any differences.

[jmone]

It is two versions of the splitter, but one includes a "Source Filter", the other will connect to the std Source Filter already on you PC as a part of the std windows install.

[nevcairiel]

YADIF 2x (i.e. with frame doubling)?
Frame doubling is really the wrong term for that.
The two modes work like this:

- Create one frame for every field (results in 50 or 60 fps)
- Create one frame for every 2 fields (results in 25 or 30 fps)

It doesn't double anything, it just takes all of the original information and creates frames from that.
This means it keeps the full temporal resolution.

Anyway, yes its supported, in the small YADIF box set output to 50/60p (which is the default)

Also, is this 32-bit only?
64-bit is supported just fine, the test build i provided is 32-bit only, though.

The release will be 32 and 64, as always.


So far the only issue I see is that it requires a video restart (or the repeat function of MPC) to apply the changes in the output mode (same/double framerate), pressing the apply button doesn't work.

Making the apply button work with every setting is not a big priority. For some settings it would be easier then for others though, so maybe i can enable it for some.
For this particular feature, seeking should re-initialize it too.


- It's me or LAV does a better job at yadif than ffdshow? I have to test and compare more, but quality seems a lot better in some places like fast horizontal letters in DVB commercials. It looks way better than ATI's VA deinterlacing to my eyes, while ffdshow looks worse.

I noticed that scrolling texts looked really good, which did surprise me as well. NVIDIAs HW deinterlacing still looked better on other kind of content, though. I didn't compare to ffdshow.


- It seems that there is no way to completely disable deinterlacing. If Yadif is off, hardware takes over. There was an option to send flags downstream (on/off), right? This is needed to avoid deinterlacing on samples that are properly flagged as interlaced but the content is actually progressive, like in most DVB movies here.

Is that really an required option?
I removed it because its so easy to mess the configuration up and get interlacing artifacts, while running a deinterlacer on progressive content is only a minor issue. I can re-add the option, i suppose.

DVB is really constantly messed up, eh. Interlaced content flagged progressive, progressive content flagged interlaced....

[STaRGaZeR]

Making the apply button work with every setting is not a big priority. For some settings it would be easier then for others though, so maybe i can enable it for some.
For this particular feature, seeking should re-initialize it too.

Yep, seeking reinits it.

I noticed that scrolling texts looked really good, which did surprise me as well. NVIDIAs HW deinterlacing still looked better on other kind of content, though. I didn't compare to ffdshow.

The funny thing is that ffdshow uses vanilla ffmpeg's yadif filter_line and related functions, but multithreading and other stuff are custom. 100% vanilla really looks better, and awesome overall, dunno why.

Is that really an required option?
I removed it because its so easy to mess the configuration up and get interlacing artifacts, while running a deinterlacer on progressive content is only a minor issue. I can re-add the option, i suppose.

DVB is really constantly messed up, eh. Interlaced content flagged progressive, progressive content flagged interlaced....

Indeed it is. PAL DVDs are the same, and with SD content quality suffers, quite a bit more than with HD. This can be easily managed with an option to disable hardware deinterlacing like LAV previously had (we already have the Yadif one), and shortcuts to enable/disable them fast on the fly as the content requires it. It's unfortunately the only way as the filter doesn't know the actual nature of the content, only humans can do that.

[nevcairiel]

Indeed it is. PAL DVDs are the same, and with SD content quality suffers, quite a bit more than with HD. This can be easily managed with an option to disable hardware deinterlacing like LAV previously had (we already have the Yadif one), and shortcuts to enable/disable them fast on the fly as the content requires it. It's unfortunately the only way as the filter doesn't know the actual nature of the content, only humans can do that.

I can surely reintroduce the option, however regarding keyboard shortcuts - i'm somewhat hesitant to add those. For one there is no clear way to give feedback to the user that the key was pressed and received, except if i add some sort of OSD - secondly, i would really prefer such things be handled by the player, not some filter it happens to use (like thats going to happen, huh?)

It would probably also need some kind of configuration for the shortcuts, all alot of work for broken content.
I'll think about it, but don't expect keyboard shortcuts soon.

If anyone wants to contribute a component that takes keyboard shortcuts and changes the options on the fly, feel free. :)

PS:
If you happen to use madVR, it has shortcuts to disable its hardware deinterlacing.

[Mercury_22]

Your filters have made WMC a viable player / option so adding keyboard shortcuts (to be used with a remote control) for changing (turning on / off) audio and subtitles will definitely make WMC my preferred player ( cause whatever other may say WMC has a very good and nice interface) for my HTPC :)

[nevcairiel]

Your filters have made WMC a viable player

Damn, how i hate myself now.

[STaRGaZeR]

I can surely reintroduce the option, however regarding keyboard shortcuts - i'm somewhat hesitant to add those. For one there is no clear way to give feedback to the user that the key was pressed and received, except if i add some sort of OSD - secondly, i would really prefer such things be handled by the player, not some filter it happens to use (like thats going to happen, huh?)

It would probably also need some kind of configuration for the shortcuts, all alot of work for broken content.
I'll think about it, but don't expect keyboard shortcuts soon.

If anyone wants to contribute a component that takes keyboard shortcuts and changes the options on the fly, feel free. :)

PS:
If you happen to use madVR, it has shortcuts to disable its hardware deinterlacing.

The feedback issue is a tough one. In ffdshow the OSD is just a subtitle, passed to the subtitle filter. Maybe when you do something about issue 137 :p

Regarding the never ending debate of where to put features, yadif and any other option that is LAV specific has no place in the player, renderer, etc. IMHO.

I don't use madVR, nor HW deinterlacing unfortunately.

BTW I just tested some filters in LAV. Deband works just fine (needs gcc > 4.6.2) but unsharp is slow as molasses as it lacks SIMD optimizations, unusable for real time playback.

[linzki]

I use ffdshow's unsharp all the time and with LAV's latest filters and MadVR it works perfectly.

I have all queue options set on max in MadVR and they stay full.

---------
specs: Intel core 2 quad CPU Q6600 @ 2.40 GHz, 4Gb ram, GeForce GTX 460, Windows 7 32-bit and Acer
H5360 projector

[nevcairiel]

I use ffdshow's unsharp all the time and with LAV's latest filters and MadVR it works perfectly.

I have all queue options set on max in MadVR and they stay full.

---------
specs: Intel core 2 quad CPU Q6600 @ 2.40 GHz, 4Gb ram, GeForce GTX 460, Windows 7 32-bit and Acer
H5360 projector

Noone was talking about ffdshow here, though. :)

[Thunderbolt8]

unsharp is slow as molasses as it lacks SIMD optimizations, unusable for real time playback.is it the same filter as ffdshows unsharp mask? would be great if speed could be improved then, I still use for all my HD stuff, because my screen size is not 1080p.

then I could finally ditch ffdshow for good ;)

[nevcairiel]

There are certain priorities here, and improving the speed of ffmpegs post-processing filters is not high on the list. Hey, right now i won't even offer them. :p

[Thunderbolt8]

hm by reading that statement from stargazer I thought you had already integrated it. I guess he was refering to unsharp mask in ffdshow then in combination with lav video?

[Keiyakusha]

BTW I just tested some filters in LAV. Deband works just fine (needs gcc > 4.6.2) but unsharp is slow as molasses as it lacks SIMD optimizations, unusable for real time playback.
in LAV? These are some up & coming features or what? I hopes Nev doesn't plans to add crappy ffdshow's deband...

[nevcairiel]

You should all stop with the assumptions. :)
I'm not copying anything from ffdshow, and i'm not yet adding any processing other then yadif.
If you know how to modify code however, the (crude draft of a) infrastructure is now there to use libavfilter filters.

[Gleb Egorych]

nevcairiel,

Tested 0.39-47 and can confirm:
1. Hang bug (http://forum.doom9.org/showthread.php?p=1538967#post1538967) is fixed.
2. Black screen bug (http://forum.doom9.org/showthread.php?p=1538127#post1538127) is fixed too.

Thanks for your great work!

[nevcairiel]

nevcairiel,

Tested 0.39-47 and can confirm:
1. Hang bug (http://forum.doom9.org/showthread.php?p=1538967#post1538967) is fixed.
2. Black screen bug (http://forum.doom9.org/showthread.php?p=1538127#post1538127) is fixed too.

Thanks for your great work!

Great!

[STaRGaZeR]

in LAV? These are some up & coming features or what? I hopes Nev doesn't plans to add crappy ffdshow's deband...

Yes, in LAV. And no, that has nothing to do with LAV's future.

[Keiyakusha]

I'm not copying anything from ffdshow, and i'm not yet adding any processing other then yadif.Yes, in LAV. And no, that has nothing to do with LAV's future.

I see... BTW, don't get me wrong, I do want to see deband in LAV at some point, its just the one in ffdshow doesn't deserves to be in LAV cause of his poor quality and occasional brokenness.

[pankov]

nev,
this is a capture I've just made from my local cable TV.
http://www.mediafire.com/?2ub8kb1pdtlbt5f
It shows a problem with the field order which happens from time to time.
From what I see the first part is BFF while the end (and the usual for our PAL country) is TFF. It's just this news story (and rarely others like it) that is inverted.
I can manually fix by changing TFF or BFF but sadly its not very convenient and I was hoping the decoder can sense this and act accordingly.
Is there any hope or I should forget about it and simply blame the producers/broadcasters for this mess?

Edit:
Or could it be that the DVB App that I'm using (DVBViewer) messed it up?

[nevcairiel]

I was hoping the decoder can sense this and act accordingly.
Is there any hope or I should forget about it and simply blame the producers/broadcasters for this mess?

If the auto fieldorder setting doesnt work, then there isnt much hope.

[STaRGaZeR]

I see... BTW, don't get me wrong, I do want to see deband in LAV at some point, its just the one in ffdshow doesn't deserves to be in LAV cause of his poor quality and occasional brokenness.

The one in ffdshow is the one in avfilter. What brokenness? Don't blame avfilter when gcc breaks :p

BTW, what's the best deband filter out there? Speed?

[Keiyakusha]

The one in ffdshow is the one in avfilter. What brokenness? Don't blame avfilter when gcc breaks :p
But why deband is always the 1st thing that breaks and many other things continue to work as always? Or this is the only gcc-compiled part in ffdshow?
BTW, what's the best deband filter out there? Speed?
I don't think I tried all possible solutions but what I personally like is flash3kyuu (f3kdb) (https://github.com/SAPikachu/flash3kyuu_deband/tree/1.4.2), which is more optimised, available for 32 and 64 bit, supports up to 16bit bitdepth (input and output), has dithering to go from higher to lower bitdepth and gives visually much nicer results. EDIT: full-chroma colorspaces probably supported too, didn't actually tried

[pankov]

If the auto fieldorder setting doesnt work, then there isnt much hope.
It doesn't work for me
:(
Does it work for anybody?

[STaRGaZeR]

But why deband is always the 1st thing that breaks and many other things continue to work as always? Or this is the only gcc-compiled part in ffdshow?

Always? Sounds like deband is completely broken :p. After I changed it to use avfilter's code instead of the old, slow and non x64 friendly code there have been no issues except this one and it was caused by gcc, not by the code itself.

Why deband? You have to ask that question to the gcc devs, there was a bug that was triggered by something in gradfun's code and not by anything else. Everything ffmpeg related is compiled with gcc in Windows (LAV, ffdshow, madVR...). x264 has suffered gcc bugs in the past too.

[THX-UltraII]

@Nevcairiel:
Are you also planning to support frame doubling/smooth video for 1080p23,976 content in the future?