LAV Filters - DirectShow Media Splitter and Decoders

[NikosD]

Two questions:

1) Can you measure your HD 4600 with some of my 1080p H.264 clips and publicly available 4K clips in DXVA native ?

2) When are you going to release the fixed VC-1/WMV3 decoder ?

If someone has an IvyBridge, I would definitely like to see the comparison with Haswell in DXVA native 1080p and 4K, because I suspect that IvyBridge QS2 is faster than Haswell QS3!

Originally Posted by nevcairiel

You need to stop thinking in absolutes, and consider the context

I don't think in absolutes, I think in general context.

[NikosD]

Originally Posted by nevcairiel

H264 is of much more interest then MPEG2, and what all my performance numbers refer to. Even extremely low-bandwidth clips won't be significantly faster, like i said, maybe ~140.

I take this phrase as an opportunity to explain further something that most CPU developers confuse and to unveil some secrets of HW architecture.

In software decoding where only CPU is involved in decoding various codecs, in various resolutions and bandwidths, all of the previous 3 parameters are significant because:

1)A decoding algorithm can be multi-threaded or not, can be optimized with various SSEx instructions or not etc.

2) Also bandwidth - the Mbps parameter - is significant, because the larger the bandwidth the harder for CPU to decode the clip.

3) Resolution for CPU is the least important factor for decoding performance as you can see in performance Excel table.

On the other hand, HW decoding in special ASICs (called UVD, VPx, QuickSync) is almost exclusively constrained by resolution.

If you see the performance tables for every HW decoder out there, for every codec and bandwidth (from the easier WMV3 or MPEG-2 to the harder H.264) and for every bandwidth (from few Mbps up to 300Mbps) the performance is almost the same, for the same resolution for each HW decoder, unless you hit first a different ceiling.

Take for example QS3.
The decoding performance for H.264 and MPEG-2, for clips between 2Mbps up to 100Mbps in 1080p resolution is almost the same.

Same rule apply for 720p, 2160p etc

If you change the resolution from 1080p to 720p or from 1080p to 2160p for example, the decoding performance changes a whole scale, for every codec and for every bandwidth.

Also you can see the same behavior for QS1, VP5,VP4, UVD but not on the same level, because they hit some performance ceilings (bandwidth limited) earlier than the resolution constrain.

The above explanation doesn't mean that HW decoders are limited to a specific number of fps.

It is just the decoding performance that is limited due to internal hardware architecture reasons for a given resolution.

If you increase the clock of the HW engine, it is obvious that you increase the upper decoding limit for given resolution, although not proportionally for every HW decoder.

So the main performance constrain for HW decoding is resolution and not codec, or bandwidth, but not artificially.

Software decoding is affected by algorithm optimizations and bandwidth a lot more.

I hope it's clear now.

[NikosD]

Nevcairiel,

I don't know if it's a limitation of Microsoft's DXVA copy-back or LAV Video, but both QS3 and VP5 can't play H.264 in huge resolutions.

DXVA native works fine for both HW decoders.

Sample here:
http://speedy.sh/Tm7BR/4080x4080-VP5.mkv

[nevcairiel]

speedy share requires a download manager, which is usually adware or at worst spyware, not going to download that.

[NikosD]

I didn't know.

Tell me a suitable server to upload it.

[nevcairiel]

http://www.datafilehost.com/ is usually OK.

I should really finish my own sample file uploading page, maybe I'll do that over the holidays.

[NikosD]

That would help to have every sample available to everyone in an easy way.

datafilehost has no progress bar.

Sample:
http://www.datafilehost.com/d/e8cfcb8c

[nevcairiel]

DXVA2 Copy back on NVIDIA or Intel works just fine here.
On NVIDIA seems a bit too slow, but i get proper images. On Intel it runs perfectly fluid.

QuickSync decoder has some image corruption, who knows whats up with that.

[NikosD]

On my system copy-back with Intel falls back to software and of course is with no image corruption.

So the problem with Intel is that is not HW accelerated - it falls back to software.
Maybe you didn't notice.

The problem with Nvidia is that although it gets HW acceleration, the images are green at the beginning and then go normal and then again have artifacts and distortion.

For QS decoder, Eric is looking into it.

[nevcairiel]

I made sure that Intel actually uses the GPU, and NVIDIA also has no corruption here.

[NikosD]

What is your method to be sure that Intel QS HW is used ?

[nevcairiel]

LAV Video tells you which decoder it uses, and that absolutely can't be wrong, also CPU usage.
Or, use GPU-Z, while it doesn't have a GPU Video Load, i can see the power usage go up 2W and the temperature increase by ~6°C. If i use software decoding, that doesn't happen.

[NikosD]

Maybe another way is to push it to the limit by benchmarking it with DXVA Checker in both modes - Software and DXVA-CB.

From the framerate and CPU usage we could have a clear sign.

With my system using both modes (software decoding and DXVA-CB) I get the same CPU utilization and same performance.

Also with VP5 I tried to take a screenshot of the green screen image, but I only got a black screen.
I tried it with DXVA Checker and standalone LAV Filters 0.59.1.

Using MPC-HC with built-in LAV Video 0.59.1.35 everything is fine with VP5.

Maybe you are using a newer version of stand-alone filters than the one posted at the first page ?

[Xaurus]

nev,

It's been a long while since I played any interlaced material, but I am watching some old stuff now. I am using CUVID and adaptive deinterlacing (50/60), but from what I read from Ctrl+J in madvr the frame rate is the same as the video itself (25 fps).
I thought I remembered that the frame rate would be doubled with this type of deinterlacing, but maybe I am remembering it wrong?

[nevcairiel]

That depends on the DVD, and madVRs frame rate display also doesn't update in realtime, it only shows the info it got during start of playback, not what actually happens.

[v0lt]

http://media.xiph.org/video/derf/y4m/akiyo_cif.y4m - normally
http://media.xiph.org/video/derf/y4m/akiyo_qcif.y4m - crash

[DragonQ]

Originally Posted by Xaurus

nev,

It's been a long while since I played any interlaced material, but I am watching some old stuff now. I am using CUVID and adaptive deinterlacing (50/60), but from what I read from Ctrl+J in madvr the frame rate is the same as the video itself (25 fps).
I thought I remembered that the frame rate would be doubled with this type of deinterlacing, but maybe I am remembering it wrong?

MadVR's frame rate indicator doesn't reflect playback after deinterlacing is done. In fairness, a 1080i/25 clip is "25 frames per second", it's just that the frames are split into interlaced fields. Once deinterlacing is done you get 50 frames per second but that doesn't make the source video 50 fps. I honestly don't know which way of indicating it would be best.

Also, it should be painfully obvious if deinterlacing is being done correctly. Not only do you lose any combing artefacts but the frame rate doubles, resulting in much smoother motion. This assumes the video is actually interlaced to start with and not just progressive video labelled as interlaced (common with TV recordings).

[vivan]

http://i.imgur.com/fukpIBk.png
http://i.imgur.com/S0ev5Q8.png
Same with CUVID

[DragonQ]

Originally Posted by vivan

http://i.imgur.com/fukpIBk.png
http://i.imgur.com/S0ev5Q8.png
Same with CUVID

CUVID/QS are different because the deinterlacing is done by LAV Video Decoder, whereas with other decoders MadVR has to sort it out (although in the end it's always hardware deinterlacing). In the latter case, you should find it says "25 fps".

[Mangix]

Ignore this post