LAV Filters - DirectShow Media Splitter and Decoders

[roytam1]

Originally Posted by nevcairiel

I mostly finished the implementation of the new RGB converter, roughly modelled after ffdshows HQ implementation, but with two key enhancements:

- Faster and more accurate chroma upscaling (ffdshow assumes MPEG-1 Chroma siting, while most content is using the MPEG2/H264 siting)
--- I do plan to support the MPEG1 version as well at some point in time, but its no pressing issue
- Increased processing bitdepth to natively support 9/10bit content

The code will always use dithering, because i've found that disabling dithering doesn't yield a worthy performance increase.
My implementation is not yet multi-threaded, but i'm not sure if thats actually needed. I'll see about that.

Below some performance numbers from my i7 2600k, decoding and processing a 1080p 4:2:0 H264 movie

- Decoding w/ YV12 output: 253 fps
- Old low-quality RGB conversion: 235 fps
- Old high-quality RGB conversion: 40 fps
- New RGB conversion: 180 fps
- New RGB conversion without dithering: 183 fps

And for fun:
- New RGB conversion in Debug mode: 20 fps.
Don't watch stuff with a debug version, people!

Speaking of multi-threading, anyone know a simple threadpool implementation that does not require me to use boost?

This one uses pthread. But I wonder if it will work in windows.
http://code.google.com/p/cppthreadpool/

Or maybe this one:
http://www.codeproject.com/KB/threads/threads.aspx

Or this one:
http://code.google.com/p/spserver/do...ar.gz&can=2&q=

In short, people like having their own wheels.

[Qaq]

Originally Posted by pirlouy

Except that it does not work in my case. Even if I add ffdshow audio processor (in MPC-HC "prefer" external filter), it never
shows after LAV audio. I don't understand. :/

Make sure you have Uncompressed - all supported in ffdshow audio - codecs.

[Superb]

Originally Posted by nevcairiel

Below some performance numbers from my i7 2600k, decoding and processing a 1080p 4:2:0 H264 movie

- Decoding w/ YV12 output: 253 fps
- Old low-quality RGB conversion: 235 fps
- Old high-quality RGB conversion: 40 fps
- New RGB conversion: 180 fps
- New RGB conversion without dithering: 183 fps

And for fun:
- New RGB conversion in Debug mode: 20 fps.
Don't watch stuff with a debug version, people!

What about ffdshow's RGB (low/high quality) conversion? How fast (fps) is it on your machine (for this file)?

[Snowknight26]

Originally Posted by fastplayer

Just select the correct speaker-setup in the Control Panel settings of your sound-card and let the mixer do the rest.

If the input channel count is greater than the output channel count then the extra channels are dropped. No downmixing occurs.

[nevcairiel]

Originally Posted by Superb

What about ffdshow's RGB (low/high quality) conversion? How fast (fps) is it on your machine (for this file)?

ffdshows conversion is multi-threaded, the comparison wouldn't work.

The only meaningful numbers:

ffdshow with YV12 output: 216 fps
ffdshow with RGB: 200 fps

For a real comparison, wait until i implemented multi-threading.

I actually just increased LAV Videos RGB performance as well.
Its now at 220 fps with RGB output, in comparison to 253 fps with YV12.

Originally Posted by Snowknight26

If the input channel count is greater than the output channel count then the extra channels are dropped. No downmixing occurs.

The DirectSound mixer on Win7 does downmixing.

On XP, that depends on your sound card drivers.

[mindbomb]

Originally Posted by nevcairiel

I will eventually offer some post-processing, at first only the filters implemented in libavfilter (mostly for yadif), anything else .. we'll see.
I would like to do some plugin type thing, maybe it'll happen. Not soon, though.

Personally, i just don't use post-processing much.

i dont use postprocessing much either, but if lav video had yadif, i imagine nearly 100% of ppl using lav video with madvr would use it.

[BloodySword]

Originally Posted by nevcairiel

ffdshow also works if you set it to libavcodec for decoding, its all the libfaad2 decoders that fail without the init data.
The more important question, which other splitter makes it work?

Oh, i see. Is the libavcodec decoder as good as the libfaad2 one? Does it decode SBR-Hints?

The splitter was the MPA-Splitter from GABEST. I can look if I can provide the version of his splitter wich I had. I think I did a backup of the old 100000 splitters before using LAV . But I never would go back to these xD.

Sorry, the file MPASplitter.ax does not have any version information.

[nevcairiel]

As far as i know, SBR is supported by avcodecs aac decoder.

I can still try to construct the proper codec setup data, not sure how much work it would be to find the right values.

[pirlouy]

Originally Posted by Qaq

Make sure you have Uncompressed - all supported in ffdshow audio - codecs.

But I don't want the ffdshow audio decoder; I only want the ffdshow processor... :/

[BloodySword]

Originally Posted by nevcairiel

As far as i know, SBR is supported by avcodecs aac decoder.

I can still try to construct the proper codec setup data, not sure how much work it would be to find the right values.

I would test it, but I don't have SBR-AAC at the moment. :/
I'll just trust and believe you! xD

ASAK the Gabset source is on sourceforge. You can take a look at it and reimplement (not copy and paste) it.

[Superb]

Originally Posted by nevcairiel

ffdshows conversion is multi-threaded, the comparison wouldn't work.

The only meaningful numbers:

ffdshow with YV12 output: 216 fps
ffdshow with RGB: 200 fps

For a real comparison, wait until i implemented multi-threading.

I actually just increased LAV Videos RGB performance as well.
Its now at 220 fps with RGB output, in comparison to 253 fps with YV12.

Wow. So you actually surpassed ffdshow already, even w/o multi-threading. :O
Amazing work.
Can't wait for the numbers after you add multi threading.

[Midzuki]

@ pirlouy, this is what works over here:

1) dis-able "uncompressed" in the audio decoder

2) en-able "uncompressed" in the audio processor

3) with Graphstudio, or RadLight Filter Manager, set the merit of the audio processor to 0x00800000 or greater

4) reboot, or logoff and logon

5) enjoy

[nevcairiel]

Originally Posted by Superb

Wow. So you actually surpassed ffdshow already, even w/o multi-threading. :O
Amazing work.
Can't wait for the numbers after you add multi threading.

First, to rule out any misunderstandings:
LAV Video already uses multi-threading for decoding, its only about the RGB converter, which is right now single-threaded.

Continuing in the text..

Well, its not 100% fair.
LAV Video will overtax your CPU, using <#cores * 1.5> number of threads, while ffdshow uses a maximum of <#cores> for decoding. So with that rule, LAV Video is faster, only because it uses more decoding threads (and of course more CPU power)

Comparing with LAV Video forced to the same number of threads as ffdshow:

8 threads:

LAV Video:
- YV12: 216 fps
- RGB output: 193 fps

ffdshow:
- YV12: 216 fps
- RGB: 200 fps

1 thread:

LAV Video:
- YV12: 55 fps
- RGB output: 49 fps

ffdshow:
- YV12: 54 fps
- RGB: 52 fps

Note that these are the number of decoding threads, ffdshows RGB converter is always multi-threaded, you really cannot turn that off for a fair 1:1 comparison.
Anyhow, multithreading the RGB converter gives a small advantage, i would say it cuts the required processing time about in half. I'll look into adding it tomorrow, should be fairly easy.

[Snowknight26]

Originally Posted by nevcairiel

The DirectSound mixer on Win7 does downmixing.

Have any more details on this? As far as I can see that doesn't happen at all. If 6-channel audio is passed to the default audio renderer, the audio renderer reports the audio stream as having nChannels = 6 and consequently (for a 2-channel device) 4 channels are dropped.

[mindbomb]

go to control panel, sound, right click on your sound device, and go to configure speakers.
then select stereo.

[e-t172]

Originally Posted by Snowknight26

Have any more details on this? As far as I can see that doesn't happen at all. If 6-channel audio is passed to the default audio renderer, the audio renderer reports the audio stream as having nChannels = 6 and consequently (for a 2-channel device) 4 channels are dropped.

No. That doesn't happen if your audio device is correctly configured. If the audio device is set to stereo, then the Windows mixer will downmix. Assuming you're not using WASAPI Exclusive or Kernel Streaming, of course.

[CruNcher]

This mixed 64/32 Bit environment is strange, i successfully registered the 64 Bit components but i have problems with the 32 Bit ones



PS: Funny the Installer fails registering the 64 Bit components, though the 32 bit components register with it crazy :P

Ahhh ok get it syswow64 and system32 are reversed

[Snowknight26]

Originally Posted by e-t172

No. That doesn't happen if your audio device is correctly configured. If the audio device is set to stereo, then the Windows mixer will downmix. Assuming you're not using WASAPI Exclusive or Kernel Streaming, of course.

Guess I was wrong. Seems that ffdshow's mixer has incorrect matrix coefficients for downmixing as they definitely don't match the Windows mixer's.


On a side note, LAV Splitter doesn't parse the 'test_audacity.wav' file in this zip.

[mindbomb]

edit: not a lav issue.

[mindbomb]

nev, can you put in the option to make lav video handle raw formats in future versions?

that way, we can, for example, use the RGB conversion with lav cuvid or cyberlink ham?