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]

Originally Posted by 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.

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

Originally Posted by DragonQ

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]

Originally Posted by aufkrawall

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]

Originally Posted by egur

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]

Originally Posted by aufkrawall

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]

Originally Posted by aufkrawall

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]

Originally Posted by egur

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

Alright.

Originally Posted by Fullmetal Encoder

Define "high quality."

I just lent that term from the LAV decoder GUI.

Originally Posted by egur

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?...3&postcount=15

[ajp_anton]

Originally Posted by STaRGaZeR

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]

Originally Posted by aufkrawall

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]

Originally Posted by ajp_anton

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]

Originally Posted by ajp_anton

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]

Originally Posted by MasterNobody

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

Originally Posted by STaRGaZeR

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]

Originally Posted by kolak

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]

Originally Posted by SamuriHL

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]

Originally Posted by Andy o

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.