LAV CUVID Decoder - High Quality Hardware decoding for NVIDIA

[SamuriHL]

Haha! Well, in the end you always pay when prefering other software over mine, hehe... :devil:


Ain't that the truth! You know how many times I recommend eac3to to people?? :) I find it *INVALUABLE* for many many tasks.


If you trust mkvtoolnix then you don't need Haali. Just do "eac3to source dest.*" to demux everything. eac3to will still remove the pulldown flags for you. You can then mux with mkvtoolnix. Have you tried some of the eac3to GUIs? I've heard ClownBD is supposed to be easy to use. Not sure if it handles HD DVDs, but I think it does.

The problem I had, and why I started using MakeMKV in the first place, is that I want MKV's of all my HD DVD/BD's. None of the eac3to gui's (at the time, however, I know Another EAC3TO GUI was working on this and may have added it by now) were able to put PGS subs into an MKV container. So I started using MakeMKV which is very simple and does what I want 99% of the time. For HD DVD, though, you're right. Laziness is costing me more effort than saving. Hence, I'm doing it right now. :) As for ClownBD, yes, I use it quite a bit, but, it doesn't output MKV's. He has no interest in updating it to do so, either. A shame cause I love that tool, but, I'm done with ISO images and m2ts/ts files. MKV is IMO a better container.

[SamuriHL]

Ah, man! I have to use mkvmerge anyway since eac3to only put the video back in the box. :D Alright, I should have just stripped the streams and had mkvmerge handle muxing instead of installing haali. See?!?! This is why I prefer to be lazy! I always manage to screw it up somehow. LOL! :D

[nevcairiel]

LAV CUVID Decoder 0.4

0.4 - 2011/04/16
- Frames flagged as progressive will not be doubled anymore (Fixes playback of streams with forced pull-down flags)
- Fixed an issue that could cause a crash in media type handling

Download: 32-bit (http://files.1f0.de/cuvid/LAVCUVID-0.4.zip)

Mostly because of the crash, which didn't happen to me as often as it should have, but hey, maybe i've been lucky.

While i was at it, i changed the frame doubler to skip frames that are marked progressive, which neatly fixes playing of those VC-1 HD DVD tracks. The media type will still say its playing at some high fps (probably 60p), but it will once again just send 24 frames per second. Note that i tested this against an actual .EVO File, not a MKV.

[SamuriHL]

SWEET! Look I can be lazy again! :D Thanks, Nev!!

[Dogway]

Nvidia 9600M GT - driver 266.58
Core2Duo 2.53 T9400 1066FSB. 4Gb DDR3 533Mhz
Laptop WinXP SP3 x86

testing with yesgrey's files:

NULL Renderer
avc
ffmpeg-mt 50dfps 585fps
lavcuvid 0.2 weave 19dfps 477.5fps
lavcuvid 0.4 weave 23dfps 806.4fps (DXVA OFF)
CoreAVC 2.5.1 weave 23dfps 453fps

wvc1
libavcodec 703dfps 3200fps
lavcuvid 0.2 weave 19.0dfps 52fps
lavcuvid 0.4 weave 28.5dfps 52.4fps (DXVA OFF)

[yesgrey]

VMR7 Renderer
You should use the Null renderer. The idea is measuring only the decoder performance.

[Dogway]

Yes, I started in Null, then I wondered if I needed a renderer for a real case scenario. Let me run tests again, and edit. Any hint for wvc1 through ffdshow?

[nevcairiel]

@Dogway:

The performance you're getting seems really low. Mabye update to 270.51 driver?

[Dogway]

I know, but 266.58 is the official last version, I think that 270.51 is only for vista/seven (http://www.nvidia.com/object/notebook-win7-winvista-270.51-beta-driver.html)

[nevcairiel]

http://www.nvidia.com/object/winxp-270.51-beta-driver.html

Yes, its marked as Beta, but overall it performs so much better for me then 266.58 did.

[SamuriHL]

http://www.nvidia.com/object/winxp-270.51-beta-driver.html

Yes, its marked as Beta, but overall it performs so much better for me then 266.58 did.

And what did you create for custom profiles? Cause I wasn't able to keep it in sync.

[nevcairiel]

I had my old profiles from the 266.58 install, didnt change a thing, and it still worked.

Profiles for 23.976, 50 and 59.94

My 23.976 profile is actually for 23.978, because that gives an actual refresh rate closer to 23.976 then entering 23.976 in the box .. go figure. <.<
(I get 23.9764... out of it, just 0.0004 away from the target!)

Also, it looks like there really are no 270.51 mobile drivers for XP, i linked the Desktop version. <.<

[Dogway]

Yes, its marked as Beta, but overall it performs so much better for me then 266.58 did.

Maybe that's the problem, thanks for the link but that's the desktop version of my card.
edit: Maybe I might run some tests on my desktop system to see if that's the cause...

[yesgrey]

Any hint for wvc1 through ffdshow?
You need to select "wmv9" as VC-1 decoder in ffdshow's Codecs tab.

[SamuriHL]

I had my old profiles from the 266.58 install, didnt change a thing, and it still worked.

Profiles for 23.976, 50 and 59.94

My 23.976 profile is actually for 23.978, because that gives an actual refresh rate closer to 23.976 then entering 23.976 in the box .. go figure. <.<

Also, it looks like there really are no 270.51 mobile drivers for XP, i linked the Desktop version. <.<

Guess I can try again tomorrow. If I manage to upset my wife we can always use the stand alone panasonic player to watch Glee. :D I would really like to get that working.

[Dogway]

You need to select "wmv9" as VC-1 decoder in ffdshow's Codecs tab.

that's what I tried, it just triggered an error: "cannot connect source to decoder"

[roozhou]

Found a bug when using PotPlayer 27623 + LAV CUVID 0.3 or 0.4
Steps:
1) Play a video file in PotPlayer using LAV CUVID
2) Click Stop
3) Click Play

The screen goes black. Seeking or re-opening the video do not solve the problem. If you open another file using LAV CUVID, the problem is gone.

My spec:
WinXP + GT240 + 270.51

[nevcairiel]

Works fine for me, using PotPlayer 26392, using madVR. Can't really be bothered to update to a new version right now. :d

[mindbomb]

lav cuvid .4 and madvr .56 are working great for me on 2/3 of my computers with nvidia graphics cards.

On one computer with windows 7 64 bit and an 8600gt, I'm getting very strange results.
madvr and cyberlink decoder in sw mode work great together
lavcuvid and evr cp work great together.

but when i try lavcuvid and madvr together, it's a slideshow.
I'm using 270.51 .

The strange thing is that i have a gtx 260 as well in another computer, and that is vp2 also, but playback is perfect on that.

Is it just a vram issue?

[nevcairiel]

The 8600GT only has 256mb ram, that can be a bit on the low-side for lav cuvid decoding and madvr.

My GPU memory usage goes up around ~185MB when running timeCodec without rendering, and ~364MB when LAV CUVID decoding with madVR. (Plus the 170mb that Aero seems to be constantly using)
I would recommend at least 512mb to everyone wanting to run this setup, maybe even 768 or 1gb for some headroom.

[CruNcher]

You could also try to disable DWM save the 170mb though lose the (window crash prevention, window system memory usage, faster bitblit) and return to the old 2D GUI (lower latency) :)

DXVA Interop also gives a nice Speed increase without Deinterlacing, results to follow :)




NVidia GeForce 9800GT G92 - 270.51 drivers - XP SP1 32

DiAVC
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 180.5

Arcsoft Video Decoder (couldn't test Cyberlinks Decoder in this environment)
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 52.3

LAV CUVID 0.4 - WEAVE
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 46.4
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 54.0

LAV CUVID 0.4 - WEAVE DXVA Interop
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 57.2
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 72.5

LAV CUVID v0.4 - BOB
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 46.4
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 54.3

LAV CUVID v0.4 - Adaptive
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 45.8
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 53.6

LAV CUVID v0.4 - frame doubling - BOB
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 46.1
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 53.6

LAV CUVID v0.4 - frame doubling - Adaptive
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 45.3
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 49.0

LAV CUVID v0.4 DXVA - BOB
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 57.2
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 68.6

LAV CUVID v0.4 DXVA - Adaptive
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 57.0
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 58.7

LAV CUVID v0.4 DXVA - frame doubling - BOB
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 57.2
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 57.5

LAV CUVID v0.4 DXVA - frame doubling - Adaptive
Test_clip_avc.1080i59.94.ac3.5.1 - dfps: 56.9
Test_clip_wvc1.1080i59.94.flac.5.1 - dfps: 49.9


Some VC-1 results surprise me though but i guess that has todo with the splitter decoder combination VC-1 in MKV :P

PS: Currently retesting 4 Girls Performance vs CoreAVC, will update the DXVA thread with the new Interop results, great work nev that Nvcuvid api based switch gives a nice performance improvement :)

Hmm DXVA Interop produces strange visual results (smear artifacts) on VMR here in the actual playback

http://img840.imageshack.us/img840/6585/dxvainteropvmrprogressi.png

trying to open another video produces just a black screen after this error on DXVA interop (have to close MPC-HC and reopen to open another video with the same error result in the end, looks like the VP2 connection is unstable after this problem occurs)

It looks like to propagate from a Reference Frame

http://img810.imageshack.us/img810/5606/referenceframesproblem.png

same on madvr

VC-1 shows other artifacts

[Dogway]

I updated my test (http://forum.doom9.org/showthread.php?p=1493357#post1493357)with CoreAVC performance, it is the same as CUVID so I think Im gonna wait for drivers update, or when I have a fresh Windows install. Thanks none the less for all the effort and improvement. I'll keep my eye on.

[CruNcher]

on VMR9 Renderless though DXVA Interop is a big improvement it blasts from 3 fps and stucky video to 50 fps and fast video just those artifacts that propagate allot of times and make watching impossible are a problem over here :(

Though there is a big improvement in Jitter and Performance visible even without DXVA Interop on VMR7 windowed compared to 0.2 :) beats CoreAVC on it :) It's as fast their now as it is with DGDecNV native :)

Though still the Performance on VMR9 Renderless is very bad better then before (1-2 fps) now with 3 fps but still much lower then CoreAVC (20 fps) and if this DXVA interop artifacts wouldn't be their it would be perfect even on VMR9 Renderless almost as fast as now on VMR7 windowed (52 fps (6ms jitter), CoreAVC 45 fps)

On MadVR DXVA Interop makes the playback as fast (6ms jitter) just those artifacts damnit :(

DXVA Interop (OFF) 9-10ms

http://img52.imageshack.us/img52/9984/nodxvainteropmadvr56.png

DXVA Interop (ON) 6-7ms

http://img405.imageshack.us/img405/4626/dxvainteropmadvr56.png

I updated my test (http://forum.doom9.org/showthread.php?p=1493357#post1493357)with CoreAVC performance, it is the same as CUVID so I think Im gonna wait for drivers update, or when I have a fresh Windows install. Thanks none the less for all the effort and improvement. I'll keep my eye on.

@Dogway

do you see these artifacts on 266.58 with your 9600M GT VP2 @ actual playback with DXVA Interop ?

[roozhou]

@nevcairiel
Another bug:
LAV CUVID cannot recognize aspect ratio of BBC HD. BBC HD 1440x1080 AVC have a DAR of 20:11 but all of them are played at 4:3 using LAV CUVID.

[nevcairiel]

It detects the AR encoded in the sequence header just fine.
If it doesn't, then its not coded properly in there. You can uncheck "Use Stream AR", which will make it use the AR the source filter specified. Also check if that option was active to begin with.

You could provide a sample file that shows the problem, but i don't think there is much i can do - the hardware decoder actually parses that information, and just tells me when it changed.

PS:
I actually have some BBC HD files here, they are all 1440x1080 at 16:9, which will cause them to be displayed as 1920x1080

[Dogway]

do you see these artifacts on 266.58 with your 9600M GT VP2 @ actual playback with DXVA Interop ?
Yes, I see kind of blocking...

[roozhou]

Checked "Use Stream AR" and now it's OK. Thanks.

BTW. BBC HD is actually played at 1964x1080.

[xv]

Thanks for the update!
Stream switching still doesn´t work.
H.264 in DVBViewer works now (also it takes a bit longer than other decoders until first picture is shown), but when I switch the channel it happens that the computer freezes a few seconds, screen turns black, then it comes back with a message like "The kernel mode driver NVIDIA... was not responding and needed to be restored".

[nevcairiel]

I just played a BBC Earth documentary (BluRay, H264 1080i/60), with Adaptive, Double Framerate and DXVA on, and i didn't get a single framedrop on my GTS450. Maybe when madVR is running it puts the GPU in some higher performance mode, or something, but it worked just perfectly, even when Benchmarks claimed i can barely just reach 60fps. :)

The decoder queue was like 5-8/8 instead of the typical 7-8/8, which means its probably the maximum it can do, though.

Quality and Smoothness was so awesome. :)
Something i could previously only get with EVR and its hardware deint.

[CruNcher]

You might have to make "DXVA Interop" dynamic either based on the Card it fails on G92 VP2 and G94 VP3 or the OS 2 times it fails on XP SP3 32, not sure yet where the problem is with it. Would be nice if someone with Vista/7 and G92/G94 VP2/VP3 could post his experience with it

[nevcairiel]

Its probably XP and its DXVA1 shit. :p

[CruNcher]

Hmm does the API Documentation mentions anything about Vista/7 or DXVA2 only for the Interop option or something driver related ?

[nevcairiel]

No, but considering the limitations of DXVA on XP, it would only make sense.

[CruNcher]

yeah that artifacts are different to the used ones (to much reference frames) but they fit into the borkness off DXVA, though that this works in MadVR makes me wonder if that isn't actually DXVA now just not from the renderer side but the Decoder side forcing the renderer, the results are just amazing from 10ms to 6ms and almost drop free :P
It's so damn amazing because it even beats Cyberlinks DXVA on a DXVA surrface VMR7 windowed 6ms vs 10ms :)

Cyberlink DXVA VMR7 windowed = 10ms 0%* CPU
LAV CUVID (DXVA Interop) MadVR 056 (Default) = 6ms 4-8% CPU
Cyberlink DXVA VMR9 Renderless Bicubic Alternate Sync VMR9-Mixer Mode = 40ms 0%* CPU
Cyberlink DXVA VMR9 Renderless Billinear VMR9-Mixer Mode = 35ms 0%* CPU

*in the area of context switches not really measurable by most tools

Amazing results this is just Amazing beating VMR7 windowed with just a itzy tiny more amount of CPU, though i still wonder if the decoding errors have todo something with the transfer speedup in this case, theoreticaly though they should just be visual errors (decoding pipeline driver related) not having anything todo with the transfer path :)

This combination could easily replace every Renderer in MPC-HC by now if those artifacts wouldn't be there on XP :)

It would mean in words the flexibility of VMR9 Renderless with the speed of VMR7 windowed with just a tiny amount of CPU usage (Power Consumption) comes very close to the Overlay Mixer with DXVA in fluidity :)

This could be very well just a Preview of whats to come when MadVR gets DXVA support subjective experience is like its the Overlay Mixer with Shader support ;)

another step closer to Arcsoft,Cyberlink and Corel(Intervideo) :)

This definitely looks like the right way for XP just something minor is wrong :)

http://img715.imageshack.us/img715/3951/hmmstrangedxvainteroppr.png

[pankov]

Guys,
yesterday I found a very useful tool for NVidia users and especially the ones with multiple displays connected to one card or others that want to fine tune the clocks/voltages of their cards.
It's called NVIDIA Inspector and includes the "Multi Display Power Saver" which is able to downclock the cards running two different displays.
Here you can read more about it
http://blog.orbmu2k.de/tools/nvidia-inspector-tool
http://www.3dcenter.org/artikel/nvidia-inspector-bringt-komfortables-multi-display-power-saving
(it's in German and for those of us that German is not understandable we can use Google translate)
Now that I found this and especially after Nevcairiel's work on LAV CUVID Decoder :thanks: I can proudly say that I'm more than happy I've changed teams and moved from ATI/AMD to NVidia
:)

[CruNcher]

Though AMD/ATI has still the more efficient Hardware and with Cyberlink HAM (APP OVD) + MadVR practically the same should be possible under AMD/ATI Hardware and with less Power Consumption then on current Nvidia Hardware :)
AMD/ATI all the years just lacked on the Ecosystem and Software support but they realized that and since Fussion they throwing out a massive amount of Developer stuff to gain ground against Nvidias Developer Ecosystem, though Nvidia still owns practically Linux with VDPAU AMD/ATI support is still lacking their :)

Currently their is the trend rather in the different direction most move from Nvidia to AMD/ATI because of Price/Performance/Consumption, especialy Windows user :)

[nevcairiel]

Maybe the hardware is more efficient, but without proper drivers, its all good for nothing. Personally, i would never get ATI just because of their f'ing drivers.
Then there is also their broken DXVA, the horrible performance when copying data back from the GPU into system memory, and their questionable marketing tricks .. but thats all OT, and there are already lengthy discussions on this over at avsforum.

The only thing that would be nice to have would be 3 monitor support on one card, so that basically means display port support. The GTX590 was i think the first NVIDIA card to ever get a DisplayPort connector, can hope that it'll be ported down into the lower cards at some point. :)

[CruNcher]

For what DXVA anyways if we get so great MadVR results :) that what for us currently is LAV CUVID + MadVR is Cyberlink HAM + MadVR for AMD/ATI users and it works as efficient and with less Power Consumption on their Hardware :)
Though none of both will be able to beat a HD2000/3000 Sandy Bridge System currently anyways :P but even in those you need one of those 2 cards for at least 3D stuff ;) though Liano could do it and with the right Software support @ launch it would blast (therefore they released the whole APP Ecosystem including OVD) and adding another AMD/ATI Card you could have Crossfire though only possible under Vista/7 :)

Though i decided for Intel already, currently still Intel (CPU tasks only) + Nvidia (Video 3D tasks and General GPU acceleration (Browser HTML5 ect)) ;)

[SamuriHL]

I have both AMD and nVidia solutions. I can tell you I prefer the nVidia solution for video playback.

[nevcairiel]

I have both AMD and nVidia solutions. I can tell you I prefer the nVidia solution for video playback.

Speaking of, i today redid my custom resolutions, because my 60hz was actually 60hz and not 59.94 (doh), and i noticed that its really broken with 270.51

So, i just installed an older one, 260.99 i think (with "clean install"), setup the resolutions, and upgrade to 270.51 again - worked just peachy! :)

[SamuriHL]

Speaking of, i today redid my custom resolutions, because my 60hz was actually 60hz and not 59.94 (doh), and i noticed that its really broken with 270.51

So, i just installed an older one, 260.99 i think (with "clean install"), setup the resolutions, and upgrade to 270.51 again - worked just peachy! :)

What the??! :D I guess I can give that a try. I clean installed 260.whatever and that's what I'm using now. So if I set up custom resolutions and then upgrade to 270, maybe it'll stick this time and not give me sync problems. Even with 260 it's not flawless. I was watching Glee with my wife last night and the sync was mostly good but it would drift in and out. It was subtle and she didn't notice (thank God) but I did. Drove me nuts. I don't know why that machine refuses to stay stable at 23.976.

[CruNcher]

For XP now i had a lot of crashes with 270.51 Video wise (they work perfect 3D Performance wise) here on G92 VP2 that i cant explain system gone unstable when video was being tried to render not really render dependent @ all i guess they took place in nvcuvid stalling the system entirely. So i will try the official WHQL driver and some other pre 270.51 from the 268 branch maybe those are more stable under XP.

[Ger]

About the Cyberlink HAM thing; as soon as I switch the Cyberlink decoder from DXVA to either HAM or SW mode the deinterlacing quality plummets (ATI 5770), so that rules out HAM for me, unless there is a registry setting or some other trick to fix it. CPU usage with HAM mode is also noticeably higher than in DXVA mode, but of course lower than in SW mode.

[CruNcher]

Oh didn't knew HAM mode has higher CPU use how much higher does AMD/ATIs OVD sample application also introduces high cpu usage for you ?

Nev i checked with 266.58 WHQL now and something peeks into my eye with it Lavcuvid 0.4 is unstable closing and reloading a video on MadVR it stays black, when using Lav Cuvid 0.2 it doesn't it's always correctly loading it never fails.
I didn't tested that with 270.51 as i used mainly DXVA interop their and that always made the reload to fail with a black screen (i thought @ first that caused it). Seems the same happens without DXVA Interop at least on 266.58 too (have to close MPC-HC and reopen it) are you sure you destroy everything correctly in 0.4 compared to 0.2 ?

Im using File->Close in MPC-HC and Drag & Drop for loading

I retested all releases the behavior starts with 0.3 all 0.2 including the test releases (dxva interop,frame doubling) are still ok

[SamuriHL]

HAM mode isn't higher CPU on my machines. But, really it's offtopic for this thread. It completely depends on your hardware.

[Ger]

System wide CPU usage with the 1080i H264 NTSC cheese slice test:
Cyberlink DXVA: Around 10%.
Cyberlink HAM: Around 25%.
Cyberlink SW: Around 50%.

I only get vector adaptive deinterlacing with Cyberlink when DXVA mode is activated.

[mindbomb]

Though AMD/ATI has still the more efficient Hardware and with Cyberlink HAM (APP OVD) + MadVR practically the same should be possible under AMD/ATI Hardware and with less Power Consumption then on current Nvidia Hardware :)


It would be great if HAM actually worked with madvr, but for me atm, there is dropped frames despite relatively low cpu and gpu usage.

on my AMD machine im using HAM and evr cp with bicubic A=-1.00 and ffdshow raw video filter to do the HQ RGB conversion, and it looks okay i guess.

i would really advise ppl to buy nvidia cards atm, so they can use lav cuvid and madvr.

[CruNcher]

It would be great if HAM actually worked with madvr, but for me atm, there is dropped frames despite relatively low cpu and gpu usage.

on my AMD machine im using HAM and evr cp with bicubic A=-1.00 and ffdshow raw video filter to do the HQ RGB conversion, and it looks okay i guess.

i would really advise ppl to buy nvidia cards atm, so they can use lav cuvid and madvr.

How is the CPU usage of this http://download2-developer.amd.com/amd/Samples/OVDecodeRender.zip though it goes off topic you better post some benches of your speed and setup in the DXVA thread under Software Players there results are needed from different systems decoder and OS http://forum.doom9.org/showthread.php?t=159486 :)

[SamuriHL]

Brilliant. Thanks for that!

[nevcairiel]

I briefly did some performance tests with LAV CUVID, and it seems to be the same as 270.51 - what i expected, just wanted to be sure. ;)

One thing does puzzle me a bit though. In all these tests, why is double-framerate mode not yielding a higher overall fps? The decode engine does not have to do any extra work, its just the post-processors deinterlacing a second frame, and those are running in the 3D engine, which should have plenty power on my GTX 570.
I'll do some funny test and enable double-framerate for progressive/weave, and see whats happening!