LAV Filters - DirectShow Media Splitter and Decoders

[DragonQ]

You can force your GPU to stay in P1 when using CUVID with certain tweaker programs but this may not provide enough power for lower-end cards aimed at the HTPC market. Using DXVA2 Native is better in terms of power usage.

[aufkrawall]

You can force your GPU to stay in P1 when using CUVID with certain tweaker programs but this may not provide enough power for lower-end cards aimed at the HTPC market.

With those cards, P0 state may consume less power than P12 state with big Fermis. ;)


Using DXVA2 Native is better in terms of power usage.
Doesn't work with madVR and QS still consumes less power.

[shadowdogg]

I have only noticed this to happen in HP1 ultimate edition but it keeps wanting to restart to the beginning part way through the track? Then randomly fix itself later.. it doesn't do that with TMT5 or FFDshow audio decoder

[egur]

I'm curios for QuickSync deinterlacing.
With 1080p VC-1 interlaced my GPU still has to switch to P0 state with shader deinterlacing, very annoying. :(
HQ deinterlacing via QS "for free" would really be a nice thing.
I really got exited about QS in general. :)

In QuickSync, the deinterlacing is ASIC not shaders for superior power/performance (same goes for everything you see in the driver's Media control panel).
My decoder has already basic support for it so LAV + QS DI is not far.
The price will be more memory copying, but the next feature will be output of DXVA frames. I'll also add the option to copy the DXVA frames so subtitle could be added on them. I'll ask for suggestions/requirement when it's relevant.

[aufkrawall]

In QuickSync, the deinterlacing is ASIC not shaders for superior power/performance (same goes for everything you see in the driver's Media control panel).
My decoder has already basic support for it so LAV + QS DI is not far.

Will the quality be about the same as CUVID or madVR? :)

[egur]

Will the quality be about the same as CUVID or madVR? :)

I think MadVR uses DXVA deinterlacing which will be the same.
CUVID will be different and we need some test clips. My guess is that for most clips, they will produce similar results. I'm not familiar at all with the DI quality of CUVID.

[nevcairiel]

It uses the same DXVA deinterlacing. I was hoping aufkrawall realized by now that all deinterlacing done is the same GPU deinterlacing. :P
Of course CUVID uses NVIDIAs deinterlacing.... :)

[Fullmetal Encoder]

That's interesting.
So, Nvidia, AMD and Intel have a dedicated function on the chip to do HQ deinterlacing?

Edit: Or at least in the driver.

Define "high quality."

While hardware de-interlacing I've seen is better than nothing it still leaves obnoxious distortions in the image. So far, nothing I've seen can compare to QTGMC. It really is a shame that we can't abstract QTGMC into an algorithm that could be implemented in such a way as to run on a GPU.

[aufkrawall]

I think MadVR uses DXVA deinterlacing which will be the same.

Alright. :)

Define "high quality."

I just lent that term from the LAV decoder GUI.


While hardware de-interlacing I've seen is better than nothing it still leaves obnoxious distortions in the image. So far, nothing I've seen can compare to QTGMC. It really is a shame that we can't abstract QTGMC into an algorithm that could be implemented in such a way as to run on a GPU.
I haven't tried QTGMC yet, but now that you mention it, I will do so.
That weird tearing artifacts are really totally fugly.

[DragonQ]

I can't see any deinterlacing artefacts on a VP4 GPU. I'm sure the algorithm is better than older GPUs though because my VP2 GPU's deinterlacing doesn't look as good in fast moving scenes.

[aufkrawall]

I see kind of tearing when he is moving his hand quickly:
http://forum.doom9.org/showpost.php?p=1274753&postcount=15

[ajp_anton]

That only applies to uncompressed video.How else would I compare the two? Both double the uncompressed data, so they both probably slightly-less-than-double the compressed data.
I was countering the "4:2:0 sucks on colored edges" argument.

[kolak]

I haven't tried QTGMC yet, but now that you mention it, I will do so.
That weird tearing artifacts are really totally fugly.

QTGMC and deinterlacing in LAV decoder is like day and night :)
I have not seen anything, which comes close to QTGMC- including DigitalVision DVO filters. They have 20 years experience in motion/video aspects, but even so QTGMC is better. Maybe Teranex box can match it, but I would not be so sure.
Problem with QTGMC is speed :(

[MasterNobody]

How else would I compare the two? Both double the uncompressed data, so they both probably slightly-less-than-double the compressed data.
I was countering the "4:2:0 sucks on colored edges" argument.
2160p 4:2:0 quadruple uncompressed bitrate not double as 1080p 4:4:4. And as was mentioned already chroma samples in 4:2:0 are not aligned with luma so you can't simply downsize luma to get 4:4:4.

[xv]

Why shouldn´t it be possible to create 4:4:4 by downsizing? You just need to know chroma position and use correct subpixel-cropping during resize. AviSynth-resizers support that (you can use float parameters as source position):
...Resize(clip clip, int target_width, int target_height, float src_left, float src_top, float -src_right, float -src_bottom)
I think you will have a position change of 0.5 pixel towards the bottom for standard 4:2:0 chroma position (at the bottom 0.5 pixel are missing and at the top 0.5 pixel are added by the resizer)

Any news for DV sound problem? Problem reproduceable with my sample?

[STaRGaZeR]

How else would I compare the two? Both double the uncompressed data, so they both probably slightly-less-than-double the compressed data.
I was countering the "4:2:0 sucks on colored edges" argument.

Compressed obviously, because that's what would be distributed eventually.

Regarding the bitrate, that's not true at all. Encoders do a much better job discarding info when compressing than just blindly removing 3/4 of the color samples. For a given bitrate, if you feed the encoder with 4:4:4 and 4:2:0, 4:4:4 will look much better without sacrificing luma quality.

The same thing happens with audio. Just feed the encoder with your 48kHz source and let it decide what is important and what is not, instead of lowpassing the source to 24kHz and compressing that. At a given bitrate, the 48kHz one will sound much, much better.

[ajp_anton]

2160p 4:2:0 quadruple uncompressed bitrate not double as 1080p 4:4:4. And as was mentioned already chroma samples in 4:2:0 are not aligned with luma so you can't simply downsize luma to get 4:4:4.I never mentioned 2160p. I specifically said increase the resolution so that the bitrate doubles. It doesn't give the same chroma resolution as simply going 4:4:4, but it also increases the luma resolution
Encoders do a much better job discarding info when compressing than just blindly removing 3/4 of the color samples. For a given bitrate, if you feed the encoder with 4:4:4 and 4:2:0, 4:4:4 will look much better without sacrificing luma quality.That's true... I was somehow thinking lossless compression. Yes, a (good) lossy encoder should be able to decide where to keep the chroma at high quality and where not to.

[Fullmetal Encoder]

QTGMC and deinterlacing in LAV decoder is like day and night :)
I have not seen anything, which comes close to QTGMC- including DigitalVision DVO filters. They have 20 years experience in motion/video aspects, but even so QTGMC is better. Maybe Teranex box can match it, but I would not be so sure.
Problem with QTGMC is speed :(

Yes, I'd really like to see a direct comparison between Teranex and QTGMC. Though I doubt Teranex could beat it. And yeah it's slow but if memory serves the last time I tested it I was able to get it to run with Preset="slow" in real time without video freezing and that's with a Core i7 860 at stock speeds. Hell even at medium preset it's still better than the deinterlacing on my Radeon 5850. Makes me wonder what will be possible with a Haswell "E" processor at end of next year which is probably what I'll be upgrading to.

The only way I can think of that you could find better deinterlacing than with QTGMC would be with a manual frame-by-frame approach.

We need to get -Vit- and Didée jobs with Teranex so they can straighten out their act.

[Andy o]

For media converter from ArcSoft it works fine. For madVR, not so much. I have to use Egur's trick of setting up a fake display. But I have CUVID now so I don't much care about it for madVR. QS is freaking *FAST* for transcoding blu-rays to mp4's for my Xoom, though. Like 7 minutes fast. It's insane.

How about if you connect the display to the Intel output, and use Virtu to access 3D graphics from NV instead? I think such a solution would work for me, if I could have an NV card with Intel HDMI audio, if Intel doesn't have the "silent stream" thing.

[SamuriHL]

How about if you connect the display to the Intel output, and use Virtu to access 3D graphics from NV instead? I think such a solution would work for me, if I could have an NV card with Intel HDMI audio, if Intel doesn't have the "silent stream" thing.

Wouldn't you wanna do the opposite then? :) I believe Nev said connecting Intel to the receiver for audio and the nVidia card to the display for video works fine. I haven't tried it cause I don't know where the hell I put my spare HDMI cables. But I'm seriously considering it because the 680 has massive overscan issues when going through the Pio Elite. I have to set a custom res to get it to work and it makes me unhappy. I'm replacing the TV sometime this year with a plasma screen, but, that's a few months away at the earliest.

[Andy o]

overscan issues now? Damn this never ends. But what I meant was why do you need to connect the cable to the NV output? If I'm not mistaken, most of what you guys said involved connecting the cable to NV, and either fool the Intel output into thinking it's active, or using Virtu for QS decoding. What I'm saying is connect the cable to the Intel output, and then you can use Virtu with the NV card for games and CUDA?

[SamuriHL]

Yea, you can do it that way and it works ok with a small performance decrease in games. My thought is that if the Intel HDMI connection is connected to the receiver and the NV is connected directly to the display (which I'll need to do anyway once I upgrade my TV for 3D since the Pio Elite isn't 1.4a) that both QS and CUDA should be active. Besides, the NV has better output IMO than the Intel.

[wanezhiling]

It uses the same DXVA deinterlacing. I was hoping aufkrawall realized by now that all deinterlacing done is the same GPU deinterlacing. :P
Of course CUVID uses NVIDIAs deinterlacing.... :)

CUVIDs deinterlacing is very nice on a not so weak or old card.

Do you have any plan to implement the GPU deinterlacing for AMD/ATI cards? I mean cb/native mode. :)

[RBG]

Do you have any plan to implement the GPU deinterlacing for AMD/ATI cards? I mean cb/native mode

Deinterlacing should work in native mode on ATI/AMD and on every dxva compatible gpu. In CB mode either EVR or MadVR can do deinterlacing just fine.

[wanezhiling]

Deinterlacing should work in native mode on ATI/AMD and on every dxva compatible gpu. In CB mode either EVR or MadVR can do deinterlacing just fine.

In cb/native mode, the deinterlacing job is just taken over by the video renderer(madVR,EVR,VMR..), lav video decoder does nothing on this.

[ianken]

I cannot get VSFILTER to load in WMP.

I searched. Googled. Binged. Bingled. Saw one post asking the same question followed by a "I figured it out" wihtout any info. Grrrrr.

Graphstudio works fine. The only bits on the box are LAV and vsfiler (xy-vsfilter 3.0.0.8).

MPC-HC of course works fine. But sometime users hit the regular player in Media Center rather than going in through Media Browser which is configured to use MPC-HC.

[RBG]

In cb/native mode, the deinterlacing job is just taken over by the video renderer(madVR,EVR,VMR..), lav video decoder does nothing on this.

In both cases deinterlacing is done via DXVA DDI function callback, and it works fine on AMD/ATI cards. I understand it now, you asked about deinterlasing implementation, where deinterlaced frames are returned to the RAM. I doubt that it possible to make it work this way in DXVAc/b mode, and I am pretty sure it is impossible in DXVAnative mode.

[wanezhiling]

I doubt that it possible to make it work this way in DXVAc/b mode.
CUVID implemented it, dunno if cb also can.
and I am pretty sure it is impossible in DXVAnative mode.
:thanks: I agree.

[shadowdogg]

Nobody wants to help. Fine. Cheers anyway

[Cyber-Mav]

MPEG4 ASP works fine if you have a 300 series driver.

thats good news, i will give it a try.

[Sebastiii]

Nobody wants to help. Fine. Cheers anyway

hi,
I was converted to PCM :) with arcsoft dll.

[aufkrawall]

thats good news, i will give it a try.
May I ask if you really need it?
Probably your PC will consume much more power with it than with CPU processing, even slow CPUs aren't stressed much.
imho on most PCs just H.264 and VC-1/WMV3 acceleration are useful.

[Octo-puss]

hi,
I was converted to PCM :)
Better than converting to Islam I guess.

[Pat357]

In both cases deinterlacing is done via DXVA DDI function callback, and it works fine on AMD/ATI cards. I understand it now, you asked about deinterlasing implementation, where deinterlaced frames are returned to the RAM. I doubt that it possible to make it work this way in DXVAc/b mode, and I am pretty sure it is impossible in DXVAnative mode.

On AMD even DXVA c/b with copying the deinterlaced frames back to the RAM would be for sure impossible due the slow copy performance from AMD to copy from the surface that holds the interlaced frames.
Maybe for NV it's possible, but what would be the sense ?
With the current implementation DVXA cb works well with interlaced content because the most important renderers (EVR, EVR-Custom, MadVR) will do deinterlace while rendering.

[Midzuki]

hi,
I was converted to PCM :) with arcsoft dll.

I see, Soaron has been enforcing the will of the Machine. :D

[nussman]

On AMD even DXVA c/b with copying the deinterlaced frames back to the RAM would be for sure impossible due the slow copy performance from AMD to copy from the surface that holds the interlaced frames.
Maybe for NV it's possible, but what would be the sense ?

You could use some processing (scaling, sharpen, ...) with ffdshow/avisynth for interlaced content (together with high quality gpu deinterlacing).

But i dont know if it's possible.

[aufkrawall]

You could use some processing (scaling, sharpen, ...) with ffdshow/avisynth for interlaced content (together with high quality gpu deinterlacing).

But i dont know if it's possible.
Why not? :)
Using LAV CUVID deinterlacing via Avisynth is possible.

[nussman]

Why not? :)
.
For dxva copyback? Its not very fast.
copyback 50i <> copyback 50p (adaptiv deinterlacing).


Using LAV CUVID deinterlacing via Avisynth is possible.
LAV CUVID => decoding, deinterlacing, copyback => avisynth?
Or what do you mean?

[RBG]

CUVID implemented it, dunno if cb also can.


CUVID is a different api, working only on nvidia hardware. AFAIK as it was designed, it had this functionality from the beginning. Anyway, sooner or later quick sync will have deinterlacing with frames return to the RAM.



Maybe for NV it's possible, but what would be the sense ?
With the current implementation DVXA cb works well with interlaced content because the most important renderers (EVR, EVR-Custom, MadVR) will do deinterlace while rendering.

Well that was originally my point, but as far as I understand wanezhiling would like to apply some kind of postprocessing to an interlaced video. :)


For dxva copyback? Its not very fast.
copyback 50i <> copyback 50p (adaptiv deinterlacing).
.....
LAV CUVID => decoding, deinterlacing, copyback => avisynth?
Or what do you mean?

I think he wrote about CUVID, and I can assure you postprocessing is working like a charm with CUVID, even in a hardest cases like SVP. For example, I did real time frame interpolation with SVP on PAL video to match 60hz refresh rate of my display. :)

[DragonQ]

Hmm I haven't found any video conversion/encoding software that can do good deinterlacing. Most of the time I prefer to leave files in interlaced format but occasionally I've thought it would be useful to deinterlace something. Are there any decent GUIs that support CUVID deinterlacing (to 50/60p) via avisynth?

[Diegotimmy]

I asked this question over at the AVS forums but didn't receive a reply:

Wondered if anyone here could help.

How can I get the lav splitter to prefer the secondary audio stream by default rather than the hearing impaired stream?

Annoyingly the hearing impaired stream is 256kbps compared to the default of 192kbps in my .WTV recordings, so i guess this is why it is being selected by default. In addition to this, playback freezes until I select the correct stream in MPC HC.

Thanks

[wanezhiling]

http://www.mediafire.com/?htq1ahkue7pqbre


http://i.imgur.com/YpNG6.png
lav splitter + lav cuvid + evr cp

http://i.imgur.com/7P6C2.jpg
lav splitter + lav sw/qs/cb/native + evr cp

[egur]

Nev, try QuickSync decoder r54.
Got Deinterlacing working (including telecine) as well as Detail and Denoise.
Let me know if this passes your regression.
Note that on soft telecined content, the output frame rate will change to 23.976 - no frame replication. If you want frame replication as a feature, let me know.
This works with and without my decoder changing the time stamps.

[nevcairiel]

Soft-telecine needs to set the repeat flags on the frames, but it should not convert it to hard-telecine. So if its not doing that, thats OK for me.
I'll check it when i have some time.

[dukey]

I got a problem with some VC1 files and CUDA. Using lav splitter and CUDA decoder. The stream ends up quite broken.

http://i.imgur.com/i6K9h.jpg

I'll upload a sample

[nevcairiel]

I'll upload a sample

Don't bother, its not like anyone here can fix any problems with CUDA decoding, only NVIDIA can do that. Send it to them.

[dukey]

Okay ..

[aufkrawall]

You could upload it anyway and I will report it to someone else who maybe can Nvidia let it fix this. :)

[sumawo13]

I made an encode from a Fraps video recorded in RGB mode, using x264 version 2197, my command line was as follows:

x264-10 --demuxer lavf --force-cfr --fps 60 --output-csp i444 --crf 16 --preset slower -o vid.mp4 "SoF2MP 2012-05-19 09-50-24-37.avi"

Here is the original file.
http://ompldr.org/tZHRmaQ (http://ompldr.org/vZHRmaQ)

Here is the encoded file.
http://ompldr.org/tZHRmeA (http://ompldr.org/vZHRmeA)

If you look you can notice a slight difference in color, I believe this to be a colorimetry problem caused by decoding with LAV Filters, which I use.

----

It seems adding --colorprim bt470bg --transfer bt470bg --colormatrix bt470bg decodes properly, but this seems like only a temporary fix.

http://ompldr.org/tZHRnZA (http://ompldr.org/vZHRnZA)

[nevcairiel]

If you look you can notice a slight difference in color, I believe this to be a colorimetry problem caused by decoding with LAV Filters, which I use.

1024x768 already qualifies as a "HD" resolution, so if the color matrix is not explicitly tagged, it'll use the BT.709 matrix.
If your content is not BT.709, and you have a resolution that would otherwise qualify as HD, you need to explicitly tag the matrix like you did in the command line you quoted in your "temporary" fix.

If the file does not tell LAV which matrix was used, all we can do is guess. And every guess will be wrong eventually.
Its working as intended, from where i'm standing.