LAV Filters - DirectShow Media Splitter and Decoders

[ryrynz]

Nev, I updated QuickSync to 0.27 and my FPS for LAV 0.46 jumped with my test file to 850FPS from 330FPS. Using LAV 0.45 with QS 0.27 I get 1250 FPS and Eric's build of ffdshow 4322 I get 1550 FPS.

I saw the posts regarding threading and thought I'd look at what my machine was doing, hoping to see why 0.45 is doing better than 0.46, as you can see below it does look like it sets itself up differently with 2 extra QS threads there in the mix.

I hope this helps somewhat in figuring out where this performance regression I'm seeing is coming from, thanks.

LAV 0.45 + QS 0.27


LAV 0.46 + QS 0.27

[egur]

FYI, the "check" function (checks QS caps) takes several hundreds ms to complete and may offset benchmarks of short clips. In normal playback it's meaningless, but for benchmarks it might lower the scores...

I don't have a debug environment for LAV decoder so I can't check what's going on exactly.

Update:
The check function reported above is not the cause. It's a misprint caused by not having a pdb file.
Replacing the QS DLL with my another that has an attached pdb file shows that a lot of time (more than 600ms) is spent calling TestMediaType and the initialization of the QS decoder.
There's nothing much to do here except run longer clips (or ignore the first run)...

[nevcairiel]

Originally Posted by egur

FYI, the "check" function (checks QS caps) takes several hundreds ms to complete and may offset benchmarks of short clips. In normal playback it's meaningless, but for benchmarks it might lower the scores...

Thats why i recommend to benchmark clips that run longer then just 1-2 seconds.
Also, it would be recommended to not count the first iteration in the benchmark.

[wanezhiling]

http://www.gokuai.com/f/B20m09Ymn989q1n0
A VC-1 file

LAV Splitter Source 0.46 failed to get video parts.
http://i.imgur.com/UXTX2.png

MPC-HC's internal Mpeg Source is ok.
http://i.imgur.com/0CqBY.png

[egur]

I'd like to request the following enhancements from LAV splitter:
1) Seek faster in TS files and provide a stream (after a seek) that will not cause image corruption (like Haali Splitter). Affects QS and libavcodec in H264 and VC1.
2) Fix seeking in WMV files. After a seek, video plays very slowly for a few seconds.

Nev, probably both issues are related to ffmpeg, but maybe you can do something...

[egur]

Originally Posted by wanezhiling

http://www.gokuai.com/f/B20m09Ymn989q1n0
A VC-1 file

LAV Splitter Source 0.46 failed to get video parts.
http://i.imgur.com/UXTX2.png

MPC-HC's internal Mpeg Source is ok.
http://i.imgur.com/0CqBY.png

Happened to me too. Here's the fix:
Reinstall LAV 0.46 and make sure VC1 is selected.
Probably related to the installer overhaul.
Your clip plays fine (32 bit anyway) with LAV 0.46 (with or w/o QS).

[CruNcher]

@ Egur
i also have a wish seeking would be this experience for anything http://www.mediafire.com/?qr6uh0x689t61xl

It's just to funny to see the GUI can't keep up, though such speeds can only be reached with a very good combination of different things working excellent together
Nev is pretty much aware of the Seek Performance and the Frame Accuracy Problems for *.ts

[nevcairiel]

Frame Accurate seeking in TS is way too complicated to even consider doing it. Its a transport format for Broadcasts, it never was designed with seeking in mind. When Blu-ray adopted the format, they added a separate index table to improve seeking (also, Blu-ray has pretty strict standards, like 1 key frame every second)
It can surely be improved a bit at some point, but you always have to keep in mind that its not a seek-friendly format. As always in those cases, 90% can be achieved with 10% of the time, and the last 10% take 90% of the time.

[Pat357]

Originally Posted by cegy

since 0.46 i have notice a odd bug with decoding. to make sure it wasn't just that 1 build so i tryed the nightly from http://roy.orz.hm/lavf-w32-nightlies...213-b92b5da.7z which still gives me the same problem, it seems it picky on how the .ts has been saved i have uploaded a sample so you can see what the problem could be. video sample

0.46 slow


0.45 fast (normal it will be over 2x as fast on a bigger sample)

LAV v0.46 software decoding falls back to single tread decoding, while LAV v0.45 can decode this file multithreaded.
Just try to set the threads in Lav-video (0.46) to 1 : you will get the same result as with xxx threads.

The question is : why lav 0.46 can only use 1 thread, while 0.45 doesn't have this limitation ??
Seems like a regression to me.

Edit : coreavc 3 gets 670 fps on this very same sample. (12 threads)

[CruNcher]

hehe btw nev did you looked further into the Mpeg-2 Bitstream crash with Lav Video DXVA2 (Quicksync Decoder) 720p.mpg that was ?

Btw also Potplayers Decoder has issues decoding it with full VLD its causing also the exception in Intels Driver, so i guess this issue goes deep, it remembers me about the funny crash conditions with jans MPC-HC tester that also caused a exception in intels driver (though he found the issue in his code for it, rather fast)

http://img526.imageshack.us/img526/2...2bitstream.png

Intels Reference Decoder (accelerated no copy back) and any Quicksync (Lav Video,FFdshow) decoder doesn't crash seems the problem is mainly in DXVA Decoding, though Cyberlink, Arcsoft and Mainconcept DXVA work

Intel Reference (no copy back) = OK
ffdshow-quicksync (copy back) = OK
lav Video Quicksync (copy back) = OK
Cyberlink DXVA (no copy back) = OK
Mainconcept DXVA (no copy back) = OK
Arcsoft DXVA (no copy back) = OK

MPC DXVA (no copy back) = Crash
Lav Video DXVA2 (no copy back) = Crash
Potplayer DXVA (no copy back) = Crash

[nevcairiel]

Originally Posted by wanezhiling

http://www.gokuai.com/f/B20m09Ymn989q1n0
A VC-1 file

LAV Splitter Source 0.46 failed to get video parts.
http://i.imgur.com/UXTX2.png

MPC-HC's internal Mpeg Source is ok.
http://i.imgur.com/0CqBY.png

That file is broken, it has CRC errors in the PAT/PMT sections.
I did however add a VC-1 probe function that manages to identify a VC-1 stream when those sections are broken, so it should work in the next version.

[wanezhiling]

I see.

[nevcairiel]

Originally Posted by ryrynz

Nev, I updated QuickSync to 0.27 and my FPS for LAV 0.46 jumped with my test file to 850FPS from 330FPS. Using LAV 0.45 with QS 0.27 I get 1250 FPS and Eric's build of ffdshow 4322 I get 1550 FPS.

Didn't we agree to go testing higher resolution material?

I did my own peformance series on 1080p material, anything else will run extremely fast anyhow.

* LAV set to QuickSync decoder, everything else default.
* GraphStudioNext, 5 passes, NULL renderer, first result not used.
* 10 clips from http://forum.doom9.org/showthread.php?t=163110
* Windows 7 Professional 64 bit, Aero on.
* Intel driver: v2622, no Lucid Virtu
* CPU: i7-2600k (3.4GHz), power management on. HD 3000 (GT2) iGPU.
* DDR3 @1600Mhz
* All Scores are averages

https://docs.google.com/spreadsheet/...jYwampfamk3ZWc

Note that at these speeds any little activity on the PC can alter the scores, so difference of 1-2 FPS are negligible.
I'm happy to report that the speed has increased basically across the board with QS 0.27 (and the changes i did for yet unreleased LAV 0.47)

I'm still a bit confused by that Samsung clip, why does it behave to special and increase so drastically, but oh well!

For the giggles, i'll add a 0.45 + QS0.27 test set.

[egur]

Originally Posted by nevcairiel

Didn't we agree to go testing higher resolution material?

I did my own peformance series on 1080p material, anything else will run extremely fast anyhow.

* LAV set to QuickSync decoder, everything else default.
* GraphStudioNext, 5 passes, NULL renderer, first result not used.
* 10 clips from http://forum.doom9.org/showthread.php?t=163110
* Windows 7 Professional 64 bit, Aero on.
* Intel driver: v2622, no Lucid Virtu
* CPU: i7-2600k (3.4GHz), power management on. HD 3000 (GT2) iGPU.
* DDR3 @1600Mhz
* All Scores are averages

https://docs.google.com/spreadsheet/...jYwampfamk3ZWc

Note that at these speeds any little activity on the PC can alter the scores, so difference of 1-2 FPS are negligible.
I'm happy to report that the speed has increased basically across the board with QS 0.27 (and the changes i did for yet unreleased LAV 0.47)

I'm still a bit confused by that Samsung clip, why does it behave to special and increase so drastically, but oh well!

For the giggles, i'll add a 0.45 + QS0.27 test set.

Very well, you're results are better than mine, we actually have almost identical systems except mine has slower RAM (1333).

[nevcairiel]

Originally Posted by egur

Very well, you're results are better than mine, we actually have almost identical systems except mine has slower RAM (1333).

Maybe its partly the ram, or the HD3000 that i have, or that i test with LAV and not ffdshow, but its no big differences, largest is 3-5% or so.

[Superb]

Originally Posted by nevcairiel

Maybe its partly the ram, or the HD3000 that i have, or that i test with LAV and not ffdshow, but its no big differences, largest is 3-5% or so.

Could you add the latest ffdshow to the table as well?

[nevcairiel]

Originally Posted by Superb

Could you add the latest ffdshow to the table as well?

I'm working on some improvements to LAV first and will re-benchmark those, after that i could, i guess.

[egur]

Originally Posted by nevcairiel

Maybe its partly the ram, or the HD3000 that i have, or that i test with LAV and not ffdshow, but its no big differences, largest is 3-5% or so.

I got much worse results when I used LAV 0.46 (w/o updates + QS 0.27).
For the Samsung clip, I got 233fps where in ffdshow I got 273.
RAM speed has a serious effect on performance.
BTW, do you a hard drive or SSD?

Update
On my new HTPC (i7-2600K, DDR3@1600, Win7 64) my result are better but not the same as yours with respect to the Samsung clip (didn't run the whole suite):
LAV 0.46 (QS 0.27) - 255fps
FFDShow 4322 (QS 0.27) - 304fps

Nev, this is strange as is this PC is practically identical yours (but has very slow/old HDD). My HTPC is installed with the bare minimum, even without anti-virus SW.

[nevcairiel]

Originally Posted by egur

I got much worse results when I used LAV 0.46 (w/o updates + QS 0.27).
For the Samsung clip, I got 233fps where in ffdshow I got 273.
RAM speed has a serious effect on performance.
BTW, do you a hard drive or SSD?

0.46 also behaved worse for me, but i changed some things now which i called 0.47 in my tests (the bold column)

I optimized even further now, which did give quite a boost again, going to benchmark 0.47 again, and then latest ffdshow build as comparison.

I have a SSD, but i don't have those samples on it.

[nevcairiel]

There, i finished another round of optimizations in LAV, and benchmarked the QuickSync decoder again - as well as ffdshow for comparison.

Here is the link again:
https://docs.google.com/spreadsheet/...jYwampfamk3ZWc

To summarize:
With 0.27 of the QS decoder, and LAV and ffdshow now using the same settings, the upcoming LAV 0.47 and the latest nightly build of ffdshow perform basically the same, with a very minor loss for ffdshow, but nothing worth noting.

Now i can work on other things again. But first, sleep.

PS:
For anyone wondering when 0.47 will be released, i only want to finish one feature, and then i would start preparing a release, so if everything goes well, the end of the week!