DXVA benchmarks - Avivo vs PureVideo vs Clear Video

[littleD]

Regarding Mpeg-4 asp, both nvidia and ati claimed support for that format but as we know there is no dshow decoder with hardware acceleration for that. But i have read somewhere (dont remember exactly where) that AMD Fusion Media Explorer [beta] (http://sites.amd.com/us/athome/amdlive/fme/Pages/AMD-LIVE-Explorer.aspx) do that. Hint>>it can also work on Intel cpus after small hack. If anybody isn't feared of installing this software, then may test it. I cant install cause of some error.Regarding 4k support http://forums.amd.com/game/messageview.cfm?catid=260&threadid=136991&enterthread=y

[mariner]

No, I put a whole different beast than zoo.
I put the birds clip.
We are talking about a 1920 x 1080p 60 fps 30Mbps clip.
It's a lot more difficult than zoo and more difficult than my first one.
Do your benchmarks and you will confirm this.
It's a special chosen clip like every clip I have in my first post

The bird.mp4 clip has been transcoded to main@4.2 profile using HandBrake.

Can't get DZVAChecker to dsiplay any numbers when benchmarking DS decoders, so here are some screenshots of testing done on 9400IGP (VP3), W7x86 :

[NikosD]

But i have read somewhere (dont remember exactly where) that AMD Fusion Media Explorer [beta] (http://sites.amd.com/us/athome/amdlive/fme/Pages/AMD-LIVE-Explorer.aspx) do that. Hint>>it can also work on Intel cpus after small hack. If anybody isn't feared of installing this software, then may test it. I cant install cause of some error.

Asked for AMD CPU, I have C2D

Regarding 4k support http://forums.amd.com/game/messageview.cfm?catid=260&threadid=136991&enterthread=y

Same guy here:
http://www.rage3d.com/board/showthread.php?t=33965737&page=2

But I'm not interested in opinions of ATI beta testers.
I want someone to explain the official Catalyst 10.4 release notes regarding 4K x 2K support

[NikosD]

The bird.mp4 clip has been transcoded to main@4.2 profile using HandBrake.

Can't get DZVAChecker to dsiplay any numbers when benchmarking DS decoders, so here are some screenshots of testing done on 9400IGP (VP3), W7x86 :

Try different splitters. Cyberlink or Haali or Nero MP4 splitter
I had same problems solved by Haali splitter (I think it was Haali)

[changturkey]

MPC-HC lastest revision here (I can't test VLC since DXVA does not work under XP).

I just tested the X64 version of build 2561 of MPC-HC, artifacts still remain for SD H.264.

[littleD]

Asked for AMD CPU, I have C2D
Ok, you may try this (http://www15.zippyshare.com/v/18690889/file.html) and may need this (http://www.microsoft.com/downloads/en/details.aspx?FamilyID=dc614aee-7e1c-4881-9c32-3a6ce53384d9&displaylang=en#filelist).
Mhh.. and this software needs also windows indexing service enabled:/.

[NikosD]

Ok, you may try this (http://www15.zippyshare.com/v/18690889/file.html) and may need this (http://www.microsoft.com/downloads/en/details.aspx?FamilyID=dc614aee-7e1c-4881-9c32-3a6ce53384d9&displaylang=en#filelist).
Mhh.. and this software needs also windows indexing service enabled:/.

The patched version worked OK.
No DivX acceleration.
Too futuristic for me and a lot of garbage.
Sorry, but it's one of the worst programs I have ever used.
Uninstalled immediately.

[Blue_MiSfit]

Very impressive performance for the Microsoft Media Foundation decoder. Very impressive.

Derek

[NikosD]

Indeed.
But it really makes me wonder, so much work for both DirectShow and MediaFoundation DXVA H.264 decoders from Microsoft and so little support for their own VC-1, WMV ?

No bitstream DXVA support for VC-1, WMV3 in WMP12 in both DS and MFT.

It's a little weird that they don't support at the same level their own video format in WMP12.

[NikosD]

@Tetsuo
It seems that even under Catalyst 10.9a, ArcSoft decoder of the first post is capable of DXVA SD H.264 without artifacts.
But that decoder can't play L5.1 ReF 16 H.264 clips.
That's why it fails at the benchmark test of clip 2.

I think this is another proof that it's probably impossible to support DXVA SD and DXVA ReF 16 L5.1 clips at the same time in UVD2.x.

[tetsuo55]

At the very least its proof that ATI is making a mess of things.

Maybe we can leverage L5.1 in some way to produce correct results for SD

[NikosD]

After further investigation I can report for ATI codecs the following:

1) ATI Fast (DnD) decoder is a Media Foundation transcoder/ decoder.
It has both attributes of transcoder and DXVA decoder, but actually it works only as a transcoder. As a hardware decoder it's broken.
It's only installed under Win 7 and it is used by AVIVO transcoder in Catalyst.
If you try to use it in DXVAChecker as a decoder, it crashes the application with H.264 clips during enumeration.
VC-1 and WMV can be tested with AMD MFT decoder in DXVAChecker because they crash the application after the test.
The last non broken AMD MFT decoder was included in Catalyst 10.3 driver. This particular transcoder/decoder doesn't crash DXVAChecker but it is useless too as a DXVA decoder only. It crashes the application after the test.
Even if you try to make it usable in WMP12 replacing VC-1 and WMV WMDecoder MFT, it crashes WMP12 too.
So, forget AMD MFT decoder

2) The "new" ATI directshow MPEG video decoder, is not so new after all. It's in Catalys hotfix versions (like 10.7a or 10.9a). It seems that official Catalyst drivers don't include the decoder. You can find it in AVIVO optional download
It's the same decoder used in ATI cards back to 2006 and 2007!
It doesn't support DXVA MPEG4 ASP at all. Only CPU mode.
DXVAchecker cannot enumerate it in most video formats and when it does, the DXVA is not working.
So it seems a useless decoder like AMD MFT.

Conclusion:
ATI has no working DirectShow and MediaFoundation DXVA decoders.

[mariner]

nVidia GT240 PureVideo4, result after three benchmark runs.

H.264 clips
----------
Third clip:

CoreAVC CUDA: 67/70/76
MPC-HC DXVA: 63/66/70
DivX 8 DXVA: 63/66/71
Microsoft DXVA DirectShow: 63/67/74

DXVA refused to enumerate the newly-remuxed-into-MP4 video. I don't know why, but the second clip works just fine after being remuxed.


Greetings ranpha.

Just like to confirm with you that MPC can render the Panasonic 1080/60p m2ts clip smoothly in EVR/CP using DXVA.

Many thanks and best regards.

[NikosD]

@Mariner(VP3)
@Ranpha(VP4)
@Namaiki(VP2, Intel)

Hello.
Is it hard for you to benchmark all the clips of the first post as it is right now, with your video engines ?
You can put your results in a post and I can put your tests to the first post in order to have them all in one place.

Thanks

[NikosD]

Updated first post results.
Vista x86 tests with CPU and GPU usage.
DXVA1 and DXVA2 versions added.
MPEG2 assisted results.

CPU average ~3%, GPU average ~0% for every bitstream format supported (H.264, VC-1, WMV3)
MPEG2 GPU average ~0% (No GPU shaders used - Only UVD and CPU for DXVA MPEG2)

[NikosD]

@Tetsuo
I know that VC-1 and WMV3 are both members of WMV9 and WMV3 is fully compliant with VC-1.
How much work is it for a developer to make the MPC - Video Decoder capable of using the VC-1 decoder for WMV3, too?
I have no idea.

[tetsuo55]

interesting, i think directshow does treat them differently, do they work with dxva in vlc? or any of the comercial codecs?

[NikosD]

Yes, VLC decodes VC-1 and WMV3 in the exact same way.
The same is true for ArcSoft and Cyberlink in some versions.
It would be useful to have MPC-HC as an all-in-one filter for DXVA.

[NikosD]

New kid on the block - UVD 3.0

Here is a screenshot of DXVA capabilities
11686

New codecs:

1) H.264MVC_VLD

2) MPEG4ASP_VLD

3) MPEG2_VLD

The first one (BluRay 3D) works in HD 5000 series too, but without HDMI 1.4a - like in HD 6000 series - is practically useless for HD 5000.

The second (DivX, Xvid etc), although had VP4 support months ago, it is the first time that it is actually supported. The support of Nvidia was paper support.
The new DivX beta codec has ATI DXVA support.

The third has a long story...

I don't have a 68xx card in my hands, but I would really like to know H.264 performance/capabilities - UHD , 1080p60, SD

[NikosD]

Catalyst 10.10 brings:

1) Removal of H264_MVC for HD 5000 introduced in Catalyst 10.9 (But it works in PowerDVD 10 Ultra Mark II as it worked before the Catalyst 10.9)

2) WMV HD support in Win 7 (It was there from the beginning - I see no difference, another trick from ATI)

3) Improved Dynamic contrast algorithm

4) Support for Blu-ray 3D - HD 6800 series

5) Stereo 3D gaming for HD 5000, 6800 series

I don't think I'm going to redo the tests for this driver.

[tetsuo55]

They seem to have changed a lot in the driver this time, results might be different when compared to older drivers.
Retesting would be welcome

[mesostinky]

I've found this to be a very interesting thread. I guess most people aren't complaining because they have fast enough cpu's and don't even worry about dxva anymore? Anyway I just ordered a low powered amd system, a dell zino w/P340/onboard vid, and after reading here about lacking DXVA implementations and seeing DXVA benchmarks like this (http://www.bit-tech.net/hardware/graphics/2010/05/24/amd-880g-graphics-performance-review/7) I wonder if I'd made a mistake? I originally was going to get a atom/ion system but just ordered a Zino instead. So sorry for taking the thread slightly off course but does anyone have an opinion if AMD should be avoided for low end HTPC systems due to issues talked about in this thread? Thanks

[NikosD]

Well, it seems that ATI/AMD finally did it.

In Catalyst 10.12 they fixed the H.264 DXVA SD bug.

You can play your favorite anime in SD with full DXVA support in both DS and MFT decoders.

Also, the WMV HD (WMV3) and VC-1 support in Win 7 introduced in Catalyst 10.10, simply meant they added a new Media Foundation decoder with an obvious name: AMD Playback Decoder MFT capable for full DXVA decoding of WMV3 and WVC1 files.

This new MFT decoder is used automatically by WMP12 only and accelerates successfully only progressive VC-1 (not interlaced - many artifacts)

The performance of that MFT decoder is the same as of AMD Fast (DnD) MFT decoder, but without crashing DXVAchecker.

Maybe the most interesting thing I found out in Catalyst 10.12 (perhaps it was there in earlier versions too) is the HUGE performance increase (x4) in Microsoft H264 Video Decoder MFT.

The difficult clip of birds (clip 3) -1080p60 fps gave me a max of 460fps!, an average of 224 fps and a min of 64.

Check out the first page to realize that people in ATI have taken their job of improving DXVA performance seriously in the latest Catalysts.

[tetsuo55]

wow that is good news, too bad you do not have a 6xxx card to add divx results with

[hwti]

The difficult clip of birds (clip 3) -1080p60 fps gave me a max of 460fps!, an average of 224 fps and a min of 64.


It seems very strange :
- benchmark results are very bad with DS (Microsoft DTV-DVD decoder) : Avg 24 / Min 23 / Max 30
- both DS and MFT do not manage to play correctly (55fps average)

Perhaps the decoder skips frames in benchmark mode, making the result wrong.

[NikosD]

It seems very strange :
- benchmark results are very bad with DS (Microsoft DTV-DVD decoder) : Avg 24 / Min 23 / Max 30
- both DS and MFT do not manage to play correctly (55fps average)

Perhaps the decoder skips frames in benchmark mode, making the result wrong.

Try WMP12 and the built-in FPS counter of WMP12.
Give me your results.

[hwti]

Try WMP12 and the built-in FPS counter of WMP12.
Give me your results.

Most of the video is played at 43fps (it starts at 40-50fps, there is a big drop to <20fps after 3s, then it goes up and stabilises).

[NikosD]

Card?
OS?
Catalyst version?

[hwti]

HD5870
Win7 64bit
Catalyst 10.12 (8.801-101125a-109686E)

[NikosD]

Try a clean system.

I can't think of anything else

[hwti]

I checked the number of DXVA2_DecodeDeviceBeginFrame events in DXVAChecker Trace Log.
The video has 2487 frames.

WMP12 : 2322 events
So it skips a few frames, which is not surprising since it can't play at full speed.

MF benchmark : 2146 events (fps results : avg 312 min 61 max 973)
341 frames have been skipped.
The maximum fps is random, between 350 and 1000 fps, it could perhaps come from skipped frames being counted as decoded.

DS benchmark : 2487 events
So all frames are decoded


The 24fps average framerate with DS benchmark comes from the fact it wants to play at 25fps ("(@25fps)" text in blue).
It seems to come from Haali Media Splitter.
In MPC-HC, looking in the "Pin Info" tab of the decoder (ffdshow with ffmpeg-mt):
- with Haali Media Splitter, the framerate is not present on input, and is 25fps on output (but it still plays at the correct framerate)
- with internal MP4 source, the framerate is 59.94fps on both pins

So I tried the DS benchmark with the standalone "MPC - MP4 Splitter".
It works, but the fps and time stay at 0 during the benchmark (even if I enable the preview, which shows the video is correctly decoded).
But in "play" mode, the fps and time are displayed !

[NikosD]

Interesting research.

I suggest to forward your results to DXVA Checker developer. I know he was in the process of making a new version. Your results of MFT benchmarking mode would be useful.

I have a beta of DXVA checker 2.4.0.0 and the DS benchmark mode gives 52/56/60 and the MFT benchmark mode gives similar results with 2.3.0.0

In my last email to him, I was complaining about the "strange" results of built-in WMP12 fps counter he is using as reference.

I suggested to him to use FRAPS.

I suggest to you also to use FRAPS while playing the clip in WMP12 and come back with your results.

[hwti]

In WMP12 FRAPS gives 57-60fps, so some decoded frames are displayed several times, or WMP12 counter is wrong.

[NikosD]

New MFT DXVA codec for UVD3 in Catalyst 11.6

It's called AMD Hardware MFT Playback decoder and accelerates in hardware MPEG-4 ASP:

MP4S, M4S2, MP4V, DIVX, DX50, XVID

[TheElix]

Wow! Will it be possible to use it in MPC-HC for those formats?

[nevcairiel]

Wow! Will it be possible to use it in MPC-HC for those formats?

No, MPC-HC does not support MFTs, and most likely never will. Only real player that can use them is WMP.

BTW, why are there two of this thread?

[NikosD]

This is the first one created for DXVA in general.

The other is for DXVA H.264 only vs CPU.