LAV CUVID Decoder - High Quality Hardware decoding for NVIDIA

[RedDwarf1]

It's an interesting App but I doubt I will bother messing around using it because I don't think it's worth the effort. If I plug my power usage meter in then maybe I would change my mind. That uses 20 watts by itself so I don't keep it plugged in all the time.

[6233638]

It's an interesting App but I doubt I will bother messing around using it because I don't think it's worth the effort. If I plug my power usage meter in then maybe I would change my mind. That uses 20 watts by itself so I don't keep it plugged in all the time.Understandable. While it's working great for me to lower my system's power consumption (18W compared to CPU decoding & 55W compared to CUVID with SD) it also makes it a nuisance for 3D gaming as you will have to add each title to the P0 list for full performance.

What I think this highlights though, is that you shouldn't just assume that GPU decoding automatically means lower power consumption. (the reason I started using CUVID to begin with)

Because CUVID kicks the GPU into its P0 state, it turned out that it was using almost 40W more than CPU decoding on my system. I'd be curious to see the results from low-power HTPC cards. In many systems, you might be better off using the CPU rather than GPU if you don't want to mess with this tool.

[RedDwarf1]

Understandable. While it's working great for me to lower my system's power consumption (18W compared to CPU decoding & 55W compared to CUVID with SD) it also makes it a nuisance for 3D gaming as you will have to add each title to the P0 list for full performance.

What I think this highlights though, is that you shouldn't just assume that GPU decoding automatically means lower power consumption. (the reason I started using CUVID to begin with)

Because CUVID kicks the GPU into its P0 state, it turned out that it was using almost 40W more than CPU decoding on my system. I'd be curious to see the results from low-power HTPC cards. In many systems, you might be better off using the CPU rather than GPU if you don't want to mess with this tool.

You make some very good points and well worth considering. My GT240 is a relatively low powered card, using about 8 watts while idling and around 75 max. I doubt my CPU, a core 2 quad, would be using that much power when decoding SD Mpeg2. Therefore, using LAV CUVID for the sake of using it is probably not a good idea. I will have to find my power usage meter and test it when I can find some time. If Gigabyte DES is anything to go by then around 8 to 14 watts for my CPU while decoding an Mpeg2 TV channel means it is probably best to use the CPU rather than GPU.

[Gleb Egorych]

nevcairiel, thanks for 0.12, it fixed long channel switching in DVBViewer for me too.

[nevcairiel]

nevcairiel, thanks for 0.12, it fixed long channel switching in DVBViewer for me too.

Glad it wasnt just me, then!
Now only pankov has to confirm as well. ;)

[yesgrey]

It's not an ideal solution, but Nvidia Inspector (http://blog.orbmu2k.de/tools/nvidia-inspector-tool)'s Multi Display Power Saver tool (right-click the "show overclocking" button to access it) allows you to force the GPU into the lower clocked P8 (Video) power state rather than going into the P0 (Full 3D) power state with CUVID.
Any reason for you to not create a specific profile for your video application? This way you can control all the clocks and do not affect the gaming.

[6233638]

Any reason for you to not create a specific profile for your video application? This way you can control all the clocks and do not affect the gaming.Well, creating a video profile (P8) for MPC-HC means that the GPU is always running in the medium power consumption mode for video playback. (this is the same mode DXVA uses)

It uses significantly less power than P0, but with my card at least, the lowest-power 2D mode (P12) is fast enough to decode SD video and scale it using MadVR. (bicubic75 chroma, softcubic70 luma)

This drops power consumption with SD video to 92W from 103W, which is only 2W more than the system at idle and worth the extra effort to me.

Setting the VPU threshold to 35% seems to be the right number for my card. While most HD video was using 45%+ I did find a couple of titles that dropped to 40% when showing title cards/credits etc which caused the power state to change and video to stutter. 35% keeps SD video in the P12 state, and HD video in P8 with everything I have tested so far. Depending on your card/source material (my HD content is all Blu-ray discs) this may need to be set differently based on the demands put on your GPU.



There is a similar setting for GPU threshold rather than VPU, but I have that disabled. It seems to either go in P0 when unnecessary, or constantly switch states causing bad stuttering with games/video playback. For applications that require it, I will just add them to the list manually. Fortunately, you can right-click the list and choose from a list of programs currently making use of the GPU.

The upside to this is that I can force the card to stay in the medium power P8 state when running less demanding games, rather than having them run at full performance unnecessarily.


Using this tool does require a bit more effort (though it's really only a one-time thing) but that's why I have a PC in the first place.
If I wanted things to be simple I would just be using a stand-alone Blu-ray player for films and a console for games (or a PS3 for both) but I would rather spend a little bit more time & effort to get a much better experience. E.g. 50Hz DVDs played back at 24p, upscaled using MadVR. Games played in native 1080p at 60fps rather than the 720p30 consoles offer.

I suppose I'm also a bit more concerned about power consumption because using a PC for playback is so much more demanding than a stand-alone now. I don't know what the most efficient is currently, but Panasonic have a player that uses 11W for playback!

I still want the functionality that a high-end PC brings for gaming, video quality and getting work done, but if I can cut the power consumption as much as possible just through the use of software, I feel it's worth making the effort.

I will admit though that I do now wonder if I should have made a better effort to reduce power consumption by only putting one optical drive in the system, buying 3TB drives rather than multiple 1.5TB drives (best price/capacity at the time) and looking into low-power RAM rather than 1.5v gaming-grade stuff (turns out you're looking at less that a 5% difference in performance) choosing a lower wattage PSU (I actually think my old 350W would have been fine, this has never pulled over 200W in my testing) and possibly having gone with AMD for video. (though I'm still wary after having driver troubles in the past)

[jmone]

So how do you setup the VPU thresholds exactly?

[6233638]

So how do you setup the VPU thresholds exactly?This is one of those things that's much harder to describe with text than it ought to be.

So there are three power states, P12, P8 and P0. Low power for 2D, medium power for video playback, and full power for 3D/games.
Normally DXVA puts the card into P8, and CUDA (which CUVID uses) seems to put the card into the full power P0 state regardless of how hard it's working.


The VPU threshold option sets the point at which the card will switch between P12/P8 depending on VPU load.

With VPU threshold set to 35%, if load goes above 35% the card will switch to P8 and if VPU load goes below 35% in P8, it switches back down to P12.

So if you set the threshold to 100% for example, it will stay in P12 until load gets to 100%. However, 100% load in P12 might only be 50% load in P8, so it will then switch back down, and up in a never-ending cycle. (ideally you would be able to set independent upper/lower thresholds)


With my card, what I have found is that decoding SD video never reaches 30% VPU load in P12, and decoding HD video always requires 40% or more in P8.

So with threshold set to 35%, the card stays in the very low power P12 state with SD video, and switches to the medium power P8 state with HD.

Without this tool, it would be stuck in the full power P0 state at all times, as CUVID makes use of CUDA.



There are two alternative options if this is sounding too complicated, or doesn't work well in your setup:
Install this tool, add mpc-hc (or your player of choice) to the P8 application list, and it will stay in P8 rather than P0 with all video playback.

On my system that means 103W usage regardless of whether video is HD/SD, rather than dropping to 92W with SD.


Forget about GPU decoding and use LAV Video Decoder, letting your CPU handle things instead. This doesn't interfere with your GPU's power management at all (it should then stay in P8) but because the CPU is doing more work, it raises power consumption to 110W in my system, however this is still a big improvement compared to the 150W drawn when CUVID sets the card to P0.

[jmone]

Thanks - Where in the tool do you set the thresholds? - Eg, I've been looking for some setting like: "Set VPU Threasholds: <0% P12 <35% P8 <80% P0 <100%" but I must be blind as I can not see where this option is hidden!

[6233638]

Thanks - Where in the tool do you set the thresholds? - Eg, I've been looking for some setting like: "Set VPU Threasholds: <0% P12 <35% P8 <80% P0 <100%" but I must be blind as I can not see where this option is hidden!Sorry, the application is not very intuitive. Right-click the "show overclocking" button:
http://i.imgur.com/x9umD.png

Unfortunately there is only one VPU slider, two would be ideal.

[jmone]

Thanks - I never would have found it! Unfortunately, it looks like there is no combo that will work for me on a 550Ti (LAV CUVID and madVR) as:

Using VPU %'s:
- SD needs P8
- "Normal" HD OK on P8
- VC-1(i) HD borderline on P8
- 1920x1080 50p needs P0

Also for smooth HQ "TheaterView" in MC you need P8 (nil VPU use but GPU maxes out on P12)

So for me I'd need really need more control to set something like:
Activate by VPU %: 0% < P12 < 5% < P8 < 75% < P0 < 100%
Activate by GPU %: 0% < P12 < 35% < P8 < 75% < P0 < 100%

In other words on my 550Ti P12 is fine when doing nothing, P8 for most playback, but I need P0 for 1080/50 or 60p (and probably also 1080/50 or 60i) and even then the VPU is being hammered.

[yesgrey]

So there are three power states, P12, P8 and P0. Low power for 2D, medium power for video playback, and full power for 3D/games.
It seems you are using the auto power management feature.

There are two alternative options if this is sounding too complicated, or doesn't work well in your setup:
How about a third option:
Disable the auto power management feature, and then install NVidia system tools and create your own custom profiles? This way you won't be stuck to P12, P8 and P0. Each person can create profiles specific to their card model.

[6233638]

It seems you are using the auto power management feature.Using the VPU threshold setting is the only way I know of that will keep the card in P12 with SD video and P8 with HD video.

Using the GPU threshold setting is what caused me all kinds of trouble, with it always switching power states in the middle of watching a film or playing a game.

How about a third option:
Disable the auto power management feature, and then install NVidia system tools and create your own custom profiles? This way you won't be stuck to P12, P8 and P0. Each person can create profiles specific to their card model.I have installed these tools, but don't see anything to do with VPU usage monitoring, only GPU?

[yesgrey]

Using the VPU threshold setting is the only way I know of that will keep the card in P12 with SD video and P8 with HD video.
Now I'm getting it. You want to change the card's clocks according to the content you are watching (SD vs HD), so it makes sense to use this tool. Another option would be to use two mpc-hc installs and use one for HD and the other for SD, and apply different profiles to them, but that would be a nasty workaround.;)

Furthermore, since you can redefine the clocks of the three different states (P0, P8 and P12), it would be almost the same. The only disadvantage is that you are limited to three states, but that should be more than enough.

I will give this a try, because with this tool I can use even lower clocks for the standard 2D mode, and that would mean lower power consumption.

[pankov]

Glad it wasnt just me, then!
Now only pankov has to confirm as well. ;)
I confirm
;)
The change is quick and there is only one minor glitch - one frame from the previous channel shows up for a moment and the first frame of the new channel is shown. Can the reason for this be that I'm using Adaptive deinterlacing and "50p/60p (Video)" frame rate.
Does anybody else see this?

Nev,
will you satisfy my curiosity and tell us what was causing the delay?

[nevcairiel]

Its just some internal technical things, DVB Viewer changed the channel without explicitly flushing the decoder, and therefor it caused the decoder to not properly discard frames from the old channel.

The issue with one frame showing is probably caused by the renderer, showing the last frame it knows until the decoder outputs a new one.

[pankov]

About the "one frame showing" issue - I'm not sure I explained it good enough so I'll try one more time:
when DVBViewer changes the channel it freezes the last frame of the old channel until the re-tuning is complete by the DVB card. After this it shows a frame (or a few - can't say for sure) from the new channel and then flashes quickly a frame from the old channel and then continues with the new channel.
Obviously it has a frame from the new channel so it's "last known frame" shouldn't be from the previous channel ... at least I think so.

[CruNcher]

Now I'm getting it. You want to change the card's clocks according to the content you are watching (SD vs HD), so it makes sense to use this tool. Another option would be to use two mpc-hc installs and use one for HD and the other for SD, and apply different profiles to them, but that would be a nasty workaround.;)

Furthermore, since you can redefine the clocks of the three different states (P0, P8 and P12), it would be almost the same. The only disadvantage is that you are limited to three states, but that should be more than enough.

I will give this a try, because with this tool I can use even lower clocks for the standard 2D mode, and that would mean lower power consumption.

Clocks aren't the major Power Consumption factor Voltages are

[ryrynz]

Yup, the trick is to lower clock speeds enough that you can lower the voltage, getting the most performance from your input voltage.

[yesgrey]

Clocks aren't the major Power Consumption factor Voltages are
Is there any tool that allows changing the voltages?

[ryrynz]

You could edit the BIOS with Nibitor to do that.

[BatKnight]

This is getting offtopic.
Please go mess with your graphic cards on a newer thread. :thanks:

Bat

[roozhou]

Hi nevcairiel,
About timestamps mess of ASP, why don't you try using timestamps from upstream(demuxers)? You can store incoming PTS in a fifo and always assign decoded frames with smallest PTS. Using average duration to calculate PTS is not a good idea even for cfr content because some files may have fake fps stored in the container or bitstream.

[nevcairiel]

Its impossible to match the frames to the input timestamps, because the decoder does not let me know if it dropped a frame to avoid artifacts (after seeks, when a reference frame is missing), so it would end up in a desync.
People should just use a software codec for MPEG4-ASP.

Also, if files contain fake header infos, i consider those broken files, and will not spend time trying to fix them. :)

I'm open to suggestions, but a simple FIFO doesn't work, because of the problem mentioned above.

[roozhou]

Its impossible to match the frames to the input timestamps, because the decoder does not let me know if it dropped a frame to avoid artifacts (after seeks, when a reference frame is missing), so it would end up in a desync.
People should just use a software codec for MPEG4-ASP.

Also, if files contain fake header infos, i consider those broken files, and will not spend time trying to fix them. :)

I'm open to suggestions, but a simple FIFO doesn't work, because of the problem mentioned above.
Will it drop frames in the middle of playback, or only after seeks? If it's latter, we can assume the first frame decoded after seek has correct pts and drop timestamps before this point.

[nevcairiel]

On a healthy file it should not drop anything during playback. It just drops frames when its obvious that the frame would be corrupted - trying to decode a P/B frame without having the reference frames.

Assuming the first frame it outputs was a I frame, and B frames are present, it would still re-order the frames internally, and PTS might not equal DTS.
B frames are quite annoying.

I could do something like scanning the MPEG4 Headers myself and just drop any frames after a seek until i see a I frame, or something like that.
Wonder if that would work for VC-1 as well....

[roozhou]

Did you try AVC-in-AVI or AVC-in-MKV using VFW mode(avi_mui_gui always produces such files). I am afraid they will suffer from similar problems.

[nevcairiel]

I don't particularly care for non-spec compliant MKV files. If you create such a file, its your own fault. There is only one proper way to mux H264 in MKV, and doing it another way will not only break this decoder, it'll probably output wrong timestamps with pretty much every decoder.

For H264 in AVI, at least there is only one way to do it, even if its ugly, and you can switch the processing mode on the file extension. But its not supported by LAV CUVID, and i have no plans to change that anytime soon.

[roozhou]

I don't particularly care for non-spec compliant MKV files. If you create such a file, its your own fault. There is only one proper way to mux H264 in MKV, and doing it another way will not only break this decoder, it'll probably output wrong timestamps with pretty much every decoder.
All video codecs can be stored in MKV using VFW mode. ASP and H264 are similar, both have native mode and VFW mode. mkvmerge uses VFW mode for ASP by default, until you add the hidden option "--engage native_mpeg4".

[nevcairiel]

"Can" doesn't mean you should. I'm just saying that alot of decoders will probably choke when you feed it such a file, and therefor is not recommended, and no-one should expect it to work properly. But this is really beside the point of the original discussion. :p

[robpdotcom]

I'd like some advice if someone doesn't mind:

I'm thinking of upgrading my GPU, partly because of LAV CUVID and the ability to use HW deinterlacing with madVR. It seems that the GTS450 is a favorite. Aside from the HW deinterlacing/decoding, should I expect any other performance increase over my ATI 5750?

[Mangix]

yes. it has faster tesselation units and higher clocks.

[nevcairiel]

The GTS450 might have higher 3D performance then the 5750, however you won't notice anything of that unless you actually play games or use other 3D applications.

[Redemption80]

Good thing i read through most of these pages over the last few days, was about to get a GT520, but hearing about the poor de-interlacing performance made me rethink that.

nevcairiel, have to say thanks for this and the splitter/audio decoders, video quality has never been better or smoother.
Also made me switch from using SPDIF for audio, and back to analog so i could use TMT's dtsdecoder:D
All working flawlessly in MPC-HC with MadVR which i have just started testing, and also flawless in MediaPortal which gets used more often due to it being more girlfriend friendly :)

One thing though to you, or Madshi, if the majority of my files are full BluRay/HDDVD rips with some BBCHD H264 1080i/50 streams as well, is it worse losing the user friendliness and pretty GUI of MediaPortal and using MPC-HC and MadVR instead?

[nevcairiel]

Only you can really judge if its worth it for you.

You can get decent results if you use ffdshows high-quality RGB conversion instead of letting your renderer deal with it, and the quality difference to madVR won't be that obvious anymore.
So if you really want to use the GUI, for a minimal quality loss i would setup something like LAV CUVID for decoding and deinterlacing, then ffdshow as a post-processor with HQ RGB conversion.

Its too bad that there are so few players with a nice 10ft GUI that support madVR.

[Redemption80]

Yeah, does seem like a dumb question when it’s my eyes, just that I wasn’t really seeing a huge difference in most of the stuff I was testing, but I do need to mess with MadVR more as I while I have had it installed for a while not, only recently started to use it.

Might try that out later on with ffdshow, thanks for the suggestion, with EVR on MediaPortal I’ve got lots of breathing space performance wise anyway on both GPU and CPU.

Yeah it is too bad, but it seems one side of the HTPC scene is going for all out quality, and the other side is prioritising ease of use and convenience.

[robpdotcom]

The GTS450 might have higher 3D performance then the 5750, however you won't notice anything of that unless you actually play games or use other 3D applications.

It won't be used for games, but how about upscaling via madvr?

Currently, I get tons of dropped frames when I upscale 720p60 to 1080p60, unless I use exclusive mode. And even in exclusive mode, I can forget about subtitles. Would the 450 perform better in that situation?

[nevcairiel]

I would've thought the 5750 would be fast enough for that as well.
You can easily test if the GPU is the bottleneck by just lowering the scaling algorithms, and see if the problem goes away.

That said, my 450 can downscale 1080p60 when in window mode without frame drops. I don't have a way to test upscaling, really.

With ATI/AMD, it might also be a driver thing, i remember reading some thread about 60fps issues...

[jmartinr]

All video codecs can be stored in MKV using VFW mode. ASP and H264 are similar, both have native mode and VFW mode. mkvmerge uses VFW mode for ASP by default, until you add the hidden option "--engage native_mpeg4".

Would you happen to know a way to remux h246 (VFW) in AVI to it's native mode in MKV or MP4?

[sneaker_ger]

Just remux with mkvtoolnix for mkv.

[Xaurus]

I am also struggling with 1080i. Using a 570 GTX. The problems arise when there is "scrolling" in the film, for example a great test part is the 1080i version of the last episode of Game of Thrones, around 6-7 minutes when the Stark mother is in the woods and you can see lots of trees scrolling infront of her, from left to right. They are jerky and I have yet to find a proper way of making this smooth.


sample ?
I have now made a sample using tsMuxeR 1.10.6 which I found to be a nice and easy tool to use. Feel free to download the sample and tell me about how your setup tackles the "evil trees". :D

Download here (http://home.halden.net/mordor/evil_trees.7z)

29.97, 1080i (if you don't remember our discussion).

[CruNcher]

I got it smooth :)

So first im using Aero on Win7 2nd im using Intel HD2000 currently so no Nvidia currently the setup that got me the smoothest motion result is

MPC-HC 64 3706

Lav splitter->Lav Audio->Lav Video->EVR (Normal) all MPC-HC Internal Sync functions disabled

http://img3.imageshack.us/img3/3503/smoothmotion.png

I didn't achieved smooth hardware playback yet

[nevcairiel]

You get 24.27 fps on a 29.97 movie, and you call that exact frame rate? :D

The file is really tricky though. It seems to have its interlacing or telecine completely screwed up, EVR doesnt manage to deinterlace it properly.

[Xaurus]

You get 24.27 fps on a 29.97 movie, and you call that exact frame rate? :D

The file is really tricky though. It seems to have its interlacing or telecine completely screwed up, EVR doesnt manage to deinterlace it properly.
Hi nev,

If the file is improperly encoded it would be a shame though...
I mean, if you regard it as unsuitable for deinterlacing/panning tests?
Perhaps something happened in the process of me extracting this part for a small clip.
Is there a way for me to give you these details from the main file? It is 5,4 GB in size...

I can tell you that with CUVID I have absolutely zero combing with the 25/30 fps setting and just no/minor combing with 50/60 setting.

In any case... do you experience the "panning" issue too? The trees "lagging" (I am not sure which word to use) even if there are no frame drops and the rest of the contents are smooth (to the eye at least)?

edit: My setup is added to my signature now.

[Xaurus]

Apparently the signature doesn't get added after a post is made (some boards do this). Signature should be visible now.

[CruNcher]

You get 24.27 fps on a 29.97 movie, and you call that exact frame rate? :D

The file is really tricky though. It seems to have its interlacing or telecine completely screwed up, EVR doesnt manage to deinterlace it properly.

Yes i know but the tree pan is smooth and it still seems sync (and it goes more towards 25 fps even might be indicating that it was actually recorded @ 25 fps from the start, nothing uncommon to use a 25 fps cam and later convert to 24 or then because of broadcasting to 29) ;) i made a video of it but it's not as smooth as in reality on screen (recording overhead)

But im surprised that i dont get smooth results with the MS DTV-Decoder and DXVA :( CPU utilization is much lower @ arround 4% but the smoothness is more bad compared to the 10% Cpu Utilization LAV Video needs for this but absolute smooth on the 4 I-5 Cores and Power Consumption isn't that much higher i would have expected even higher latency because of the multithreading but Win7 seems to be very efficient here compared to XP also the Power Config (Balanced) doesn't interfere in anyway Performance wise :)

@nev
it seems the smoothnes on EVR (normal) compared to EVR Custom comes from the Interlace Flag sending in Lav Video turning it on on EVR causes problems for the tree scene turning it off results in that 24.xx smooth playback :) though on Evr custom you have to turn it on again and then it fails with smooth playback resulting in some 39.xx fps It's interesting on EVR you dont seem to need to send anything it will get correctly deinterlaced and telecined by the hardware regardless. EVR custom seems to lose that ability currently this could maybe also explain the behavior on my sample.ts (which is also telecined it seems) :)

Also my major guess is you see those 24.xx fps on EVR because Intels Hardware is doing some good adaptive work here ;)

[Xaurus]

nevcairiel,

I ran the MPEG2Repair program on the original file just to log it (not repair it):

Sequence Frame 79601(0-I) / Time 0:54:35 :
Info: End of MPEG2 sequence

Sequence Summary:

File Size Processed: 5.13 GB, Play Time: 00h:54m:35s
1920 x 1080, 29.97 fps (24.30 fps Telecine), 18.00 Mbps (12.66 Mbps Average).
Average Video Quality: 63.60 KB/Frame, 0.25 Bits/Pixel.
AC3 Audio: 3/2 Channels (L, C, R, SL, SR) + LFE, 48.0 kHz, 384 kbps.
Dialog Normalization: -24.0 dB, Center Mix Level: -3.0 dB, Surround Mix Level: -3.0 dB
0 of 79601 video frames found with errors.
0 of 102358 audio frames found with errors.
0 corrupted video bytes in file.
0.000000 seconds of video timestamp gaps.
0.000000 seconds of audio timestamp gaps.

End of Log

[CruNcher]

Hehe i found a way to get it smooth with Lav Video (Send Interlace Flags) also on EVR Custom being telecined correctly :P this is really funny but forcing DirectVobsub to be loaded makes this possible

[Gleb Egorych]

Hi, nevcairiel

Regarding deinterlacing issue (http://forum.doom9.org/showthread.php?p=1519634#post1519634): I've uploaded a sample -> http://www.mediafire.com/?4cilzfa418ztey5
MediaInfo says it's progressive ("scan type") but it also has "scan order" property.