LAV Filters - DirectShow Media Splitter and Decoders

[hoborg]

@hoborg....what is that? And where can I find it? :)
Yes it is graphstudionext - View -> Decoder performance.

[SamuriHL]

Brilliant. Thanks for that. 51-52fps on average with my Quad Core 8200, 5870 12.1a preview, W7 SP1 X64, using 32 bit versions of everything.

[nevcairiel]

Too bad, i would've hoped for faster.

Here is a second try:
http://files.1f0.de/lavf/LAVFilters-0.44-dxva2-perftest2.zip

You need the first version, and just replace the .ax file

I increased the number of buffers from 2 to 8, apparently on my Intel GPU this gives me a significant performance boost on some samples, so who knows, maybe it does something for ATI.
Otherwise, i'm not sure i know what else to try without an ATI card at hand - maybe its just everything we can get out of them...

[HeadlessCow]

1. 99% of the releases don't use this feature. And LAV Splitter supports MANY other containers much much better than Haali or MPC-HC's filters.
2. The 64 bit version of MPC-HC. Yes. We all know that 64bit player cannot run a 32bit renderer. I was asking: "why not install the 64bits components of LAV Filters?" (so that they will be available for other [64bit] programs)
3. But why? LAV Splitter does a much better job.
4. Well, wasn't this a "newbie" guide? Adding ffdshow to the filter chain would be be useless for 99.999% of the users, since they don't even know what video post processing is or how to setup it properly.
5. It doesn't mess up the MPC-HC GUI. The move to exclusive full screen is just fine (does the quick "flash" really bother anyone?). They must have forgotten exclusive mode results in a much smoother playback (which is one of the best features of madVR).

1. (As pointed out by blexley later) Because it came from a page about anime fansubs where it is (relatively) common.
2. Because it's a guide about getting MPC-HC working with madvr... The 64-bit stuff is clearly outside its target.
3. "disable everything" <--- repeated for your benefit. It then goes on to show the ones that could be enabled without breaking the setup that the guide is trying to give you.
4. "Here's a simple guide... and if you feel like doing something slightly more advanced, here's how to configure that as well." This isn't really that complicated to understand.
5. The "quick" flash is much, much slower on some hardware (or drivers) than others. Starting a video and then having 5 seconds of black screen before it actually starts to play can be annoying.


Still though (and I don't think you understand this), I'm not arguing that the guide is good/bad. Just that you were an jerk for no particularly good reason. And that you didn't actually read the post before trying to mock the poster.

[hoborg]

Nothing changed here, still ~51FPS

[clsid]

Will there be a direct rendering mode in the future? For older ATI and Intel cards.
Any plans to implement some kind of smart fallback mode that switches back to software decoding if the GPU isn't fast enough to keep up with the fps of the video?


Another question, do you happen to know the minimum required version of nvcuvid.dll that is compatible with current LAV Video? Or at least the lowest version you know that is compatible.

[adam777]

Thanks, nev.

Everyhthing is done on a mobility radeon 4570, 512MB.

http://i.imgur.com/7tybi.png

Also did a (very) quick real world testing.

1080p50 (test sample) and 1080p60, both MKV, both completely unwatchable under Zoom Player and madVR.

1080p24 seems to work fine, running smooth, though madVR's queues are a bit on the lowish side (around 7-8/12).

Downloaded YouTube 1080p clip completely unwatchable (MP4).

Let me know when will you need more detailed reports, I'll be glad to help.

* EDIT *
second test build gives similar results.
Oh, I suppose it's possible the GPU simply isn't powerful enough to handle madVR at 1080p60.

[nevcairiel]

Nothing changed here, still ~51FPS

Too bad.
If anyone feels like testing alot, you could grab VLC and test its DXVA on such a 50p or 60p clip, and see if it works better? I doubt it would, but maybe AMD hacked up the driver to do something special when VLC is detected - its certainly possible.


do you happen to know the minimum required version of nvcuvid.dll that is compatible with current LAV Video? Or at least the lowest version you know that is compatible.

Not of the dll, but it should work with all CUDA 2.0 and upwards version, which is around driver 190 or so.

[SamuriHL]

Same for me, as well. Statistically maybe SLIGHTLY slower, but, 52 on average.

[HeadlessCow]

As a final point, there's something important to remember about web forums...

While blexley might be the 100th in a long line of newbies asking stupid questions on doom9, it's the first time he's been on doom9 and asked a stupid question. Treating him(or anyone new) as if he is asking a stupid question for the 100th time is just being mean.

[CruNcher]

A better listing includes GPU utilization as well (Perftest 1) :)

Progressive 8 bit Playback efficiency 1080p 60 fps (4 girl H.264)

Arcsoft DXVA = ~2% CPU; GPU = ~20%
Cyberlink DXVA = ~2% CPU; GPU = ~20%
CoreAVC DXVA = ~4% CPU; GPU = ~20%
Intel Quicksync DXVA2 = ~12% CPU; GPU = ~23%
Nev Generic DXVA2 = ~17% CPU; GPU = ~24%
Nev Lav Video Libav = ~25% CPU; GPU = ~25%


All DXVA2 and Software results (Lav Video). Results include Parser (Lav Splitter MKV)/Audio(Lav Audio)/Renderer (EVR Custom Experimental) Overhead

[fairchild]

My results on my HD 5830 using 11.12 drivers and system in my sig benched with graphstudionext (avg 3 runs to null renderer):

Lav Video DXVA2: 52.2 fps

for reference on same system:

Lav Video software: 143 fps

Edit: with your perftest2, same exact results as above.

[CruNcher]

Too bad, i would've hoped for faster.

Here is a second try:
http://files.1f0.de/lavf/LAVFilters-0.44-dxva2-perftest2.zip

You need the first version, and just replace the .ax file

I increased the number of buffers from 2 to 8, apparently on my Intel GPU this gives me a significant performance boost on some samples, so who knows, maybe it does something for ATI.
Otherwise, i'm not sure i know what else to try without an ATI card at hand - maybe its just everything we can get out of them...

It's always better to restart the host Application it could hold up the old dll still in memory better is to first make sure the host aplication doesn't do this, and if your not sure better restart it.

[VipZ]

Here's results from a 7970

Here's the correct results,
http://dl.dropbox.com/u/19338638/Decoder%20Perf.JPG

I guess the 280fps was too good to be true :p

@ madshi, here's the results for your test, http://dl.dropbox.com/u/19338638/madnvtest.JPG

[nevcairiel]

Well, a slight increase, bringing it closer to NVIDIAs performance - its unsure if this may just be the decoder limiting now, though.

Interesting that madshi's test app still only shows 4 fps NV12 download though. :p

[SamuriHL]

Knock it off. I don't have money for a new video card! :D

[CruNcher]

Though if you compare these results with Nvidias old cards it looks not really fast those results would be in the range of a 460 GTX if i remember right ;)

[SamuriHL]

Yea, well, I also game on my machine, so, you can compare the video all you want....but once you start looking looking at gaming numbers things might be slightly....different. :D

[STaRGaZeR]

~52fps with both builds on 1080p50 Crowdrun.

CPU usage lower than the QS decoder.

1080p60 unwatchable. 1080p50 starts with a few seconds of stuttering, then it's smooth with a few drops below 50fps here and there. But it's not LAV Video's fault, Microsoft's decoder fails too, the stuttering is a bit less pronounced but that's it. The hardware decoder itself is shit, nothing to be done about that.

PS: Parkjoy is harder to decode than Crowdrun. Better benchmark material :D

[sneaker_ger]

Nothing changed here with the second version, still around 50 fps on the HD 5850.

[VipZ]

Knock it off. I don't have money for a new video card! :D

hehe. Though as much as DXVA has been improved on the 7970, the power consumption is through the roof, 2600k is massively more efficient.

CPU is about 340w from the wall for my stuff and DXVA copy back is about 390w from the wall, this is for 720p, old DXVA test was just bit over 400w.

[CruNcher]

~52fps with both builds on 1080p50 Crowdrun.

Power consumption and CPU usage lower than the QS decoder.

1080p60 unwatchable. 1080p50 starts with a few seconds of stuttering, then it's smooth with a few drops below 50fps here and there. But it's not LAV Video's fault, Microsoft's decoder fails too, the stuttering is a bit less pronounced but that's it. The hardware decoder itself is shit, nothing to be done about that.

Hmm impossible i guess you forgot that the QS Decoder doesn't get used for QFHD resolution (not performant enough) so Nev disabled it and it fallsback to Libav ;)

Are you sure you took the 1080p results ?

[nevcairiel]

I'm not sure if there is anymore to be gotten out of the AMD HW decoder, maybe i should just get a cheap card like a 5450 and see if i can improve it somehow.


Hmm impossible i guess you forgot that the QS Decoder doesn't get used for QFHD resolution as Nev disabled it and it fallsback to Libav ;)
This is all 1080p.

[SamuriHL]

hehe. Though as much as DXVA has been improved on the 7970, the power consumption is through the roof, 2600k is massively more efficient.

CPU is about 340w from the wall for my stuff and DXVA copy back is about 390w from the wall, this is for 720p, old DXVA test was just bit over 400w.

I couldn't care less about power consumption tbh. :D Heat is what I'm more concerned with. I have a pretty sweet Antec case so even that's marginal. My rig is over 2 years old now and will probably get replaced once it hits the 3 year mark in November. I need to behave until then, however. :)

[SamuriHL]

I'm not sure if there is anymore to be gotten out of the AMD HW decoder, maybe i should just get a cheap card like a 5450 and see if i can improve it somehow.


Those of us on AMD hardware would certainly appreciate it. I run both AMD and nVidia in the same house, but, I know the AMD crowd is itching for hardware assisted decoding to use with madVR.

[STaRGaZeR]

There seems to be nothing more in the hardware. Pure DXVA solutions are marginally faster.

[CruNcher]

I'm not sure if there is anymore to be gotten out of the AMD HW decoder, maybe i should just get a cheap card like a 5450 and see if i can improve it somehow.



This is all 1080p.

yep sorry added that though it's strange doesn't fit to my QS Decoder results @ all (which from the top to bottom make absolute sense) though different bitstream.

[nevcairiel]

There seems to be nothing more in the hardware. Pure DXVA solutions are marginally faster.

Luckily, there also isn't a hell of a lot content that wouldn't work. Blu-ray maximum is 60i, which should play just fine (its essentially 30p)

Anyhow, i gotta fix my HTPC now, the damn thing somehow died, wonder WTF that thing has now.
And i was hoping for a smooth ride until i replace it with IVB in a few month

I think its the stupid SSD in it, meh.

[Snowknight26]

6870 @ stock
perftest1 = 55.15fps (every time, 20 tries)
perftest2 = 55.15fps (every time, 20 tries)

This was over remote desktop no less. Surprisingly, GPU usage was at a steady 35%. Guess it's limited by the memory speed. I'd try overclocking it but the CCC doesn't open over RDP. :\

[CruNcher]

Playback and Null Rendering Efficiency Results are conform (except Nev and Lav you would expect Nev DXVA2 to be faster then Libav)

Lav Video Intel Quicksync DXVA2:

http://img7.imageshack.us/img7/188/lavvideointel.png

Lav Video Libav:

http://img27.imageshack.us/img27/9109/lavvideolibav.png

Lav Video Nev DXVA2:

http://img252.imageshack.us/img252/3248/lavvideonevdxva2.png

[madshi]

@ madshi, here's the results for your test
Thanks!

Interesting that madshi's test app still only shows 4 fps NV12 download though. :p
Yes, quite interesting! I don't really know why this happens. Will have to check that. Maybe the memcpy implementation I'm using is making problems. Funny enough, it doesn't harm with NVidia GPUs.

[STaRGaZeR]

Luckily, there also isn't a hell of a lot content that wouldn't work. Blu-ray maximum is 60i, which should play just fine (its essentially 30p)

Anyhow, i gotta fix my HTPC now, the damn thing somehow died, wonder WTF that thing has now.
And i was hoping for a smooth ride until i replace it with IVB in a few month

I think its the stupid SSD in it, meh.

Indeed. I also have doubts about the supposedly ultra fast Intel decoder. When you benchmark it, the power consumption that the CPU reports for the GPU part of the chip increases to enormous levels compared to normal 1080i30 playback. I'm talking about sub 1W vs 7W and such. That doesn't seem like a normal load PC for a little fixed function hardware decoder. The GPU also enters 3D mode during the benchmark, and not while normal playing.

Out of curiosity, is that SSD Intel by any chance?

Guess it's limited by the memory speed. I'd try overclocking it but the CCC doesn't open over RDP. :\

On my 5850 playing with GPU clocks (300-800) and memory clocks (300-1000) resulted in similar perfomance.

[STaRGaZeR]

...

Slow!

QS decoder:

http://img638.imageshack.us/img638/595/43569286.png

Software:

http://img205.imageshack.us/img205/614/softg.png

Okay, I cheat, 4500MHz on the CPU, 1500MHz on the GPU ;)

[nevcairiel]

Out of curiosity, is that SSD Intel by any chance?

Nah, OCZ.
Guess i'll just put a good ol' magnetic back in there and get a new one when i upgrade the HTPC

But i think something else is going on, too, maybe the GPU broke as well, something truely odd is going on...
Meh, put in my GT430 that i had sitting around somewhere and it boots again, f'ing great. SSD and GPU gone boom, wonder if there was a power spike or something.

[King Kong]

http://www.abload.de/img/lavfilters-0.44-dxva2e5rem.png (http://www.abload.de/image.php?img=lavfilters-0.44-dxva2e5rem.png)

Tested w/ Intel i7-2600K & AMD Radeon HD 6950 2GB.
"Crowd run" test file stutters every 1-2 secs.

Not using Intel QSD due to limitations of P67 chipset.

[bjd]

@Nev

Great work on the DXVA2 (copy back), I can even use this hardware acceleration on my 4 yeaVista based AMD Turion TL-60 Laptop with a Radeon X1250

[PeQuE]

Is it my fault, or Dolby Digital Plus (EAC3) bitstreaming in Lav Audio is broken in 0.44?

In Graphstudio I can connect Lav Splitter audio pin to Lav Audio, but then it refuses to connect to audio renderer.

EDIT: it only affects if bitstreaming is selected... If not, no problem

[bjd]

@Nev

Great work on the DXVA2 (copy back), I can even use this to play 1080p24 on my 4 year old Vista based AMD Turion TL-60 Laptop with a Radeon X1250 !!

[egur]

Using ffdshow QS with background (MT) frame copy:
http://img638.imageshack.us/img638/8950/ffdshowqs23crowdrun1080.png
i7-2600 stock speed DDR3@1333MHz

[CruNcher]

@ Eric
Right i forgot its always good to check against ffdshow-quicksync efficiency wise :)

and indeed its still better if the benchmark result is really valid (see comment bellow) :)


1st. FFdshow Quicksync:

http://img97.imageshack.us/img97/9573/ffdshowquicksyncres.png

2nd. Lav Video Quicksync:

http://img7.imageshack.us/img7/188/lavvideointel.png

Either that or the result is false because of the different timestamp handling (also rate and real time doesn't get shown, could indicate a problem getting the FPS header) though this shouldn't influence the benchmark shouldn't it (lav splitter) ?
Though these stuff is one reason i dont like this benchmark splitter and decoder differences can result in very weird behavior and benchmark results same for DXVA Checker (Elecard filter for example benching is impossible, they send back a 00000 result for everything same for Mainconcept with lav splitter).


Rechecking the Playback results i see no improvement though compared to Lav Videos Quicksync (only in the benchmark)

Progressive 8 bit Playback efficiency 1080p 60 fps (4 girl H.264)

Arcsoft DXVA = ~2% CPU; GPU = ~20%
Cyberlink DXVA = ~2% CPU; GPU = ~20%
CoreAVC DXVA = ~4% CPU; GPU = ~20%
Lav Quicksync DXVA2 = ~12% CPU; GPU = ~23%
ffdshow-quicksync DXVA2 = ~12% CPU; GPU = ~23%
Nev Generic DXVA2 = ~17% CPU; GPU = ~24%
Nev Lav Video Libav = ~25% CPU; GPU = ~25%

Software results (Lav Video). Results include Parser (Lav Splitter MKV)/Audio(Lav Audio)/Renderer (EVR Custom Experimental) Overhead

[Andy o]

I know the AMD crowd is itching for hardware assisted decoding to use with madVR.

Is that true anymore, for most? I know it's not for me, deinterlacing was the last piece of the software-decoding puzzle and madVR has fixed that beautifully. It's been a few years since H.264 progressive was very demanding on current CPUs.

[SamuriHL]

Is that true anymore, for most? I know it's not for me, deinterlacing was the last piece of the software-decoding puzzle and madVR has fixed that beautifully. It's been a few years since H.264 progressive was very demanding on current CPUs.

Yes. It's one of the most asked for things. Does it TRULY matter to most? Probably not. For my laptop it'd make a big difference. For my HTPC, not really. For a lot of people, it's something they want. Notice, I didn't say NEED. :)

[Andy o]

Oh, I don't doubt it's a good thing, just wondering if people that think they need it (or even want it) are a bit misguided, cause for instance, system efficiency (was evidenced in VipZ's for instance) can suffer in many cases, contrary to popular belief that GPU decoding is always more efficient. You don't need a super powerful GPU for that to be true, either.

[SamuriHL]

My laptop and my bedroom HTPC would not be able to play video without GPU decoding. My laptop I've been able to work around the issue as it's faster than my bedroom HTPC, but, it definitely benefits from GPU decoding. Any modern CPU, however, is more than sufficient, yes. Some day I'll be getting a new laptop and rebuiding both HTPC's, but, until that happens......

[zero9999]

Too bad, i would've hoped for faster.
Here is a second try:
http://files.1f0.de/lavf/LAVFilters-0.44-dxva2-perftest2.zip

ATI HD 5750: ~51,2 fps (CrowdRun 1080p50)
Also, crashes with Hi10P material (instead of falling back to software decoding)

[DragonQ]

Nev, if you get DXVA2 working for AMD cards, will it work for nVidia cards too? If so, what are the advantages/disadvantages compared to CUVID?


By the way I get 57 FPS on average with software decoding on a Core 2 E8400 with CPU usage at ~95%. When using DXVA2 on an AMD Radeon HD3400, I get 59 FPS but CPU usage is still at ~95%. Such a poor GPU (this is an office PC) is probably useless for DXVA2.

[wanezhiling]

Mobility HD4650 (http://we.pcinlife.com/data/attachment/forum/201201/13/174540sy8yaxcyy8spu84y.png)
AVG 52.3325fps

[nevcairiel]

Nev, if you get DXVA2 working for AMD cards, will it work for nVidia cards too? If so, what are the advantages/disadvantages compared to CUVID?

Compared DXVA2 to CUVID:

Pro:
- Uses potentially less power because CUVID forces max performance mode

Con
- No deinterlacing
- Marginally slower
- Requires D3D interaction (cannot use it with a secondary card)
- No VC-1 interlaced support (unless i manage to fix that)

Note that i also have plans for a "native" DXVA2 mode for people running EVR, which would behave similar to all other DXVA decoders out there.

[DragonQ]

Compared DXVA2 to CUVID:

Pro:
- Uses potentially less power because CUVID forces max performance mode

Con
- No deinterlacing
- Marginally slower
- Requires D3D interaction (cannot use it with a secondary card)
- No VC-1 interlaced support (unless i manage to fix that)

Note that i also have plans for a "native" DXVA2 mode for people running EVR, which would behave similar to all other DXVA decoders out there.
Thanks. Lower power consumption would be nice, especially for an HTPC, but hardware deinterlacing is a must IMO.

[nevcairiel]

If you use EVR or madVR, just let them deinterlace.
It was more important when madVR did not do deinterlacing yet.