Intel QuickSync Decoder - HW accelerated FFDShow decoder with video processing

[CruNcher]

No release notes for R2 nothing i guess it's just the IVY update to 1.3 API, like it was floating around in the beta driver

[egur]

Originally Posted by CruNcher

No release notes for R2 nothing i guess it's just the IVY update to 1.3 API, like it was floating around in the beta driver

There's a release notes document with the MSDK download.
Changes from the 2012 gold version are:
* Three new H.264 profiles and 6 (all defined in standard) constrained flags were defined. This change should simplify Media SDK usage by providing direct access to constrained flag configuration.
MFX_PROFILE_AVC_CONSTRAINT_SET0/1/2/3/4/5
MFX_PROFILE_AVC_CONSTRAINED_BASELINE
MFX_PROFILE_AVC_CONSTRAINED_HIGH
MFX_PROFILE_AVC_PROGRESSIVE_HIGH
* View output mode was added for MVC encoder mfxExtCodingOption::ViewOutput. This flag instructs encoder to output each view in separate bitstream buffer and format them according to Blu-ray* and AVCHD* Format requirements.
* MVC encode now supports interlace coding mode.

[nevcairiel]

Hey Eric,

i found a sample with quite an odd behaviour today.

http://files.1f0.de/samples/h264-qs-timestamp-issue.mkv

When i play this file with LAV in QS mode, it plays too fast (should be 24p, plays at 30p), running out of sync.
I checked the timestamps of the file, and they are fine. Plays fine with software or other hardware codecs too.

For some reason, the QS decoder insists on changing the timestamps, even though i turn off the timestamp mangling flag that you expose.

I'm running driver 2696, happens both on my Ivy and Sandy systems.
I just tried 2656, and the problem does not occur anymore.

The driver dependency makes me believe its the MediaSDK itself causing this, and not your code directly.
Can you maybe check if its something you can turn off, or maybe if the problem is just gone in a newer driver which i don't have yet?

Edit:
2712 is also showing the problem.
Either its a bug in the newer MediaSDK libraries which you should forward, or its a feature thats on by accident?

[egur]

Originally Posted by nevcairiel

Hey Eric,

i found a sample with quite an odd behaviour today.

http://files.1f0.de/samples/h264-qs-timestamp-issue.mkv

When i play this file with LAV in QS mode, it plays too fast (should be 24p, plays at 30p), running out of sync.
I checked the timestamps of the file, and they are fine. Plays fine with software or other hardware codecs too.

For some reason, the QS decoder insists on changing the timestamps, even though I turn off the timestamp mangling flag that you expose.

I'm running driver 2696, happens both on my Ivy and Sandy systems.
I just tried 2656, and the problem does not occur anymore.

The driver dependency makes me believe its the MediaSDK itself causing this, and not your code directly.
Can you maybe check if its something you can turn off, or maybe if the problem is just gone in a newer driver which i don't have yet?

Edit:
2712 is also showing the problem.
Either its a bug in the newer MediaSDK libraries which you should forward, or its a feature thats on by accident?

Odd behavior - managed to reproduce. Still no root cause. I'll update when I have a fix.

I've root caused the problem:
* H264 SPS header doesn't contain a frame rate (num_units_in_tick = time_scale = 0).
* If I manually set the frame rate for the MSDK to 23.976 it will behave correctly.
* This happens even when VPP is off and no time stamp calculation/manipulation.

Nev, make sure the version of the MSDK HW DLL is 368 (from 2696) not 369 (from 2712).

I've committed a patch that also fixes some DI issues along with a workaround for the above issue (r51).

Here's my commit log:

Experimental build - not stable enough for production SW.
* Update to MSDK 2012 R2.
* Workaround for MSDK bug - zero frame rate in H264 SPS header would cause a wrong frame rate.
* Deinterlacing now works on frames stored as progressive but marked as interlaced content (many PAL TV broadcasts).
* Telecined content not supported by DI (yet).

[NikosD]

Originally Posted by egur

I'll try the other clips on Sunday.

Eric,

if you try the 4K clips posted by Wanezhiling and me, you can use PotPlayer's internal built-in codecs, because it has an option to "force" HW decoding above 1080p.

The tab key shows CPU utilization of PotPlayer process alone and overall CPU utilization at the same time, separated by a "/".

This is in case you want to check playback performance, not benchmarks of course.

PotPlayer has an extra advantage:
It has direct/ native internal DXVA playback and your QS decoder as an option, too.

So you can check both (direct DXVA and QS decoder) in an easy way, until Nevcariel fixes LAV filters playing 4K in direct mode.

[egur]

Tested 5 4K clips using LAV and ffdshow.
LAV and ffdshow showed very similar results so I just wrote LAV-QS score.
The CPU% is how much the player's process took, not the entire system. All clips maxed out turbo at 3.4GHz. Performance wasn't constant during playback as the clips varied in bitrate - OldTownCross ran at 280mbps, peaking at 500 (!) mbps which is outside H264 spec.

CrowdedRun 29% (up to 34%)-3.4GHz
DucksTakeOff 36% (up to 42%) -3.4GHz
InToTree 32% (up to 36%) -3.4GHz
OldTownCross 36% (up to 42%) -3.4GHz
ParkJoy 29% (up to 31%) -3.4GHz

LAV DXVA-CB – half the CPU% of QS but drops a lot of frames (visually, EVR reports everything is OK). Clips are not playable.
LAV DXVA-native – not working – uses libavcodec, maxes out CPU.

[nevcairiel]

Originally Posted by egur

LAV DXVA-native – not working – uses libavcodec, maxes out CPU.

Native DXVA is limited to 1080p right now, because there are quite some issues with DXVA of 4K on AMD and NVIDIA.
Need to re-visit this in a while when everything has new drivers. Its not like 4K content is important right now.

I probably should improve the logic to allow Intel, but will do that in one go when i check it again.

[egur]

Originally Posted by nevcairiel

Native DXVA is limited to 1080p right now, because there are quite some issues with DXVA of 4K on AMD and NVIDIA.
Need to re-visit this in a while when everything has new drivers. Its not like 4K content is important right now.

I probably should improve the logic to allow Intel, but will do that in one go when i check it again.

I'm considering 4K video a preview to a new technology - give us and the HW makers time to clean up the act before it becomes mainstream.

[hajj_3]

4k video will probably be using HEVC codec though, i'm not sure if we'll see much 4k video in h.264. I'm hoping that Intel adds dedicated hardware in Haswell for HEVC hardware decoding.

[nevcairiel]

Originally Posted by hajj_3

4k video will probably be using HEVC codec though, i'm not sure if we'll see much 4k video in h.264. I'm hoping that Intel adds dedicated hardware in Haswell for HEVC hardware decoding.

Doubtful that it'll be in Haswell already.
The HEVC standard isn't even finalized yet, and content providers will take years to adopt it anyway.

[hajj_3]

there is already an ARM chip that has HEVC support and that chip was released a while back. The final draft spec is out in july and ratified in january. Chip makers already have chips that can do HEVC hardware decoding such as qualcomm snapdragon and ziilabs: http://www.youtube.com/watch?v=vZfsULKG2Zo http://www.eetimes.com/electronics-n...ll-Android-SoC

So i'm pretty sure that intel could have support in haswell if they wanted to.

Yes it will take a while for software support for HEVC but if they have hardware that can support it then your pc will be future proofed (video-wise) with a haswell cpu.

[egur]

Originally Posted by nevcairiel

Doubtful that it'll be in Haswell already.
The HEVC standard isn't even finalized yet, and content providers will take years to adopt it anyway.

Let me correct that, not doubtful - impossible. Spec closes in 2013 about the same time Haswell hits the market...
Intel would need a time machine to pull this off

Originally Posted by hajj_3

there is already an ARM chip that has HEVC support and that chip was released a while back. The final draft spec is out in july and ratified in january. Chip makers already have chips that can do HEVC hardware decoding such as qualcomm snapdragon and ziilabs: http://www.youtube.com/watch?v=vZfsULKG2Zo http://www.eetimes.com/electronics-n...ll-Android-SoC

So i'm pretty sure that intel could have support in haswell if they wanted to.

Yes it will take a while for software support for HEVC but if they have hardware that can support it then your pc will be future proofed (video-wise) with a haswell cpu.

This ARM SoC has a DSP processor to do the decoding. DSPs are programmable so it's not an ASIC solution. This is not far from CUVID in concept. They can always rewrite the DSP code. It will suck a battery dry in no time...

[NikosD]

New intel beta drivers for Windows 8, for both Sandy and Ivy.
They work also for Win 7 (I tried 32bit version with Core i5 and Win 7 SP1 x86)

They include OpenGL 4 drivers for Ivy.

You can download the HD Graphics driver v2927 (15.28.0.64.2729 (9.17.10.2729)) from this page

64bit:
http://downloadcenter.intel.com/Deta...21180&lang=eng

32bit:
http://downloadcenter.intel.com/Deta...21235&lang=eng

[CharlieCL]

Some GPUs of Ivy Bridge Core i5 are HD3000, can it play 4K?

[nevcairiel]

Originally Posted by CharlieCL

Some GPUs of Ivy Bridge Core i5 are HD3000, can it play 4K?

Ivy Bridge is either 2500 or 4000, none are 3000.
Both 2500 or 4000 should be able to play 4K.

[CruNcher]

does it also fit performance wise 2000(GT1),2500(GT1),3000(GT2),4000(GT2?) ?

i doesn't looked @ a compare of 2500 (Ivy,GT1) vs 3000 (SB,GT2) yet

though actually Ivy bridge has been renamed back to GT1 in that line so i guess it doesn't fit to continue the GT naming indeed as we talking about 6 (2500) vs 12 EUs (3000) and getting the same performance with half the EUs i doubt that (on the 3D part).

And the speculation about Haswell also sound a little crazy 40 Eus with the Performance of what 80 or even 160 of the current EUs

Originally Posted by NikosD

New intel beta drivers for Windows 8, for both Sandy and Ivy.
They work also for Win 7 (I tried 32bit version with Core i5 and Win 7 SP1 x86)

They include OpenGL 4 drivers for Ivy.

You can download the HD Graphics driver v2927 (15.28.0.64.2729 (9.17.10.2729)) from this page

64bit:
http://downloadcenter.intel.com/Deta...21180&lang=eng

32bit:
http://downloadcenter.intel.com/Deta...21235&lang=eng

Outch
Normally i don't give much about Benchmarks but Microsofts Performance Index Shows a Performance Decrease of 1.1 Points on my HD 2000 for the Game Graphics Value before it was 5.8 (Aero Performance result stayed the same @ 5.2) that's a heavy drop before this drop it was constantly going up per driver release climaxing @ 5.8 with 2719 with 2729 now it dropped down to 4.7 (and becomes the final system result now compared to 5.2 before).

This might indicates some big change here (maybe IQ wise ?)

Have todo some real Engine tests to see if there is something realizable outside of Benchmarks

I could swear Aero feels even snappier (this also constantly improved over driver releases) though (lower latency) even showing no real Benchmark improvement in the Performance Index.

Yep definitely a good latency test is the Aero to Windows Basic change it's much much faster then it was in every previous driver (the mode change is almost fluid now, though Win 8 will be totally fluid WDDM 1.2 )

@egur

Both Lav Video Quicksync as well as FFdshow Quicksync stopped working (DXVA works)

[egur]

Windows 8 drivers have either broken MSDK DLLs or they are missing altogether. The Win8 drivers are beta and not fully functional.

[diizzy]

Hi,

I might be completely off in this matter but I can't enable QS on my i7-3770 (non K) and I can't find out why, LAV Filters lists it as "Not available" and it doesn't show up in ffdshow at all.

Motherboard: MSI Z77A-GD65
OS: Windows 7 Pro 64-bit SP1
Driver: 8.15.10.2712 (link: http://downloadcenter.intel.com/Deta...adType=Drivers )
ffdshow: ffdshow_rev4441_20120430_xhmikosr_icl12.exe (I can give your build a go if you want to)
LAV Filters: 0.50.2

Does it require lucid or something?

//Danne

[wanezhiling]

Originally Posted by NikosD

Using PotPlayer latest version and ATi hardware in DXVA mode I couldn't play both of them at the signature system.

Playback of the MPEG-2 clip is awful - I can't see even one frame correct.

Playback of the H.264 file is better - the only problem is a green horizontal stripe at the top of the film.

BTW, WMP12 plays both clips perfect in DXVA mode.

PotPlayer 1.5.33425 should have fixed both.

[egur]

Originally Posted by diizzy

Hi,

I might be completely off in this matter but I can't enable QS on my i7-3770 (non K) and I can't find out why, LAV Filters lists it as "Not available" and it doesn't show up in ffdshow at all.

Motherboard: MSI Z77A-GD65
OS: Windows 7 Pro 64-bit SP1
Driver: 8.15.10.2712 (link: http://downloadcenter.intel.com/Deta...adType=Drivers )
ffdshow: ffdshow_rev4441_20120430_xhmikosr_icl12.exe (I can give your build a go if you want to)
LAV Filters: 0.50.2

Does it require lucid or something?

//Danne

You need Lucid if you don't have a display connected to Intel GPU. Direct3D/DXVA requires a display for GPUs to be enumerated. You can also try faking a connected display. See a previous post