Evaluation of HEVC decoders (SW, Hybrid and HW)

[Yups]

Interesting for Haswell and Broadwell.

- Improved video playback through partial hardware acceleration support for the VP9 video format.
- GPU accelerated decode of HEVC video file format including both 8-bit and 10-bit support. This will provide improved video playback capabilities on the platforms.
https://downloadcenter.intel.com/Det...&DwnldID=24596

[NikosD]

I tried that driver and doesn't look to have a DXVA decoder device for VP9.
Unless it's not named as VP9 or maybe DXVA Checker doesn't recognize it.


I also tried 10bit HEVC clips with latest LAV Video 0.63.41 HEVC DXVA decoder and I only get a black screen with no decoding.

[Yups]

What was your previous driver?

[NikosD]

v. 3960.

Why ?

Are you looking for comparisons between drivers ?

[Yups]

Because I noticed some differences from my previous driver (4029).

This video: http://demo-uhd3d.com/files/uhd4k/De...vely_Swiss.zip

With driver build 4029 GPU usage was low and CPU usage 80-100%. With driver 4080 I get 40-50% CPU usage and 90-100% GPU usage in Potplayer (colour issues unfortunately), in MPC Home Cinema there is only a black screen.

This video: bbb-3840x2160-cfg02.mkv

GPU usage is higher than before, 99-100% usage now. Previous usage was ~80%.

[wanezhiling]

driver 10.18.14.4080

Just like a slide show...

http://i3.tietuku.com/21a83a47c35363e0.png

[JohnLai]

10 bit HEVC dxva decoding by intel? I not sure if intel Haswell HD IGPU has the hardware to decode it in first place.

This one probably refers to upcoming Broadwell.

^.^ If this 10bit dxva for haswell igpu really work, somebody please inform NVIDIA that their competitor already has 10bit dxva working. AMD too.

[nevcairiel]

From what I've seen its still kind of pointless, as it only supports 8-bit output, so while it may be able to decode it partially, it then probably rounds the 10-bit output to 8-bit, which degrades the quality.
I need to check in more detail if its actually possible to get it to output 10-bit.

[NikosD]

Originally Posted by JohnLai

10 bit HEVC dxva decoding by intel? I not sure if intel Haswell HD IGPU has the hardware to decode it in first place.

This one probably refers to upcoming Broadwell.

^.^ If this 10bit dxva for haswell igpu really work, somebody please inform NVIDIA that their competitor already has 10bit dxva working. AMD too.

We are not talking about fixed-function HW decoder, like QuickSync, but for partial GPU assisted (via shaders) HW decoding.

So, it's possible to decode 10bit HEVC using a hybrid decoder.

Originally Posted by nevcairiel

From what I've seen its still kind of pointless, as it only supports 8-bit output, so while it may be able to decode it partially, it then probably rounds the 10-bit output to 8-bit, which degrades the quality.
I need to check in more detail if its actually possible to get it to output 10-bit.

But why the black image and no decoding without fall-back to SW using LAV Video ?

If all modifications are inside drivers, I think it could or should work instantly, without modifications of LAV video.

Or at least fallback to SW mode.

[nevcairiel]

They probably broke something else.

[NikosD]

Originally Posted by Yups

With driver build 4029 GPU usage was low and CPU usage 80-100%. With driver 4080 I get 40-50% CPU usage and 90-100% GPU usage in Potplayer (colour issues unfortunately), in MPC Home Cinema there is only a black screen.

GPU usage is higher than before, 99-100% usage now. Previous usage was ~80%.

Originally Posted by nevcairiel

They probably broke something else.

PotPlayer seems to use DXVA for 10bit clips, but with artifacts and colour issues.
But it uses DXVA.

Also, PotPlayer plays fine in DXVA those Samsung NX1 clips posted above http://www.imaging-resource.com/news...ailable-for-do, when LAV DXVA native crashes immediately.

Also, LAV DXVA copy-back doesn't crash, but falls back to SW decoding using the above NX1 clips.

[nevcairiel]

Originally Posted by NikosD

PotPlayer seems to use DXVA for 10bit clips, but with artifacts and colour issues.
But it uses DXVA

I deliberately disabled 10-bit from even attempting to use DXVA2 because broken decoding is worse than no decoding. LAV will never knowingly produce wrong output.

HEVC 10-bit has a special second decoder device called HEVC_VLD_Main10 (or {107AF0E0-EF1A-4D19-ABA8-67A163073D13}). If any GPU supports proper 10-bit decoding, I expect it to offer this device. Its the only way to really know if it supports it or not, without hard-coding a list of GPUs, which I would be very reluctant to do.
In fact, the DXVA2 HEVC specification clearly states the the HEVC_VLD_Main device shall only be used for the "Main" profile, and HEVC_VLD_Main10 for the "Main10" profile.

So far I have not seen this new device on any GPU, even with Intels new driver which presumably claims 10-bit support.

[NikosD]

OK, but Samsung NX1 clips are not 10bit, they are 8bit.

[wizziwig]

Originally Posted by nevcairiel

So far I have not seen this new device on any GPU, even with Intels new driver which presumably claims 10-bit support.

Only on Broadwell. Haswell is limited to 8-bit only.

https://communities.intel.com/message/273816#273816

10bit hybrid HEVC decode is enabled on BDW U and H sku. Not available on HSW.

8bit hybrid HEVC decode is available on both HSW U sku and above and BDWU and H sku.

Thanks
Robert

[Yups]

Ok that makes sense. No wonder there are playback issues on my HSW with 10 bit. Looks to me that the playback improvements refer to Broadwell.

[nevcairiel]

Thats unfortunate, as I'm unlikely to get a Broadwell-U CPU. Maybe once a desktop Broadwell comes out, or maybe the rumors are true and NVIDIA has 10-bit support in the 960, and working on it there may work on Intel as well.

[NikosD]

Originally Posted by wizziwig

Only on Broadwell. Haswell is limited to 8-bit only.

https://communities.intel.com/message/273816#273816

Originally Posted by Yups

Ok that makes sense. No wonder there are playback issues on my HSW with 10 bit. Looks to me that the playback improvements refer to Broadwell.

Thanks wizziwig.

I really like the speed of posting in Intel forums by Yups

The Broadwell thing crossed my mind, but I didn't want to think that, that was the reason.

Probably VP9 HW acceleration refers to Broadwell only as well.

Originally Posted by nevcairiel

Thats unfortunate, as I'm unlikely to get a Broadwell-U CPU. Maybe once a desktop Broadwell comes out, or maybe the rumors are true and NVIDIA has 10-bit support in the 960, and working on it there may work on Intel as well.

Nobody with a decent Haswell is going to get a Broadwell-U.

Personally, I'll move directly to Skylake, skipping even desktop Broadwell, even if Skylake is not going to be released along with desktop Broadwell as Intel says.

In my opinion, Skylake will come after desktop Broadwell due to the well known Intel's delays, but that's only a speculation.

[Yups]

Broadwell-K might be interesting to see how GT3e performs but Skylake-S with its Gen9 has a newer Gfx architecture, it should support HEVC decode in hardware finally. This is much more appealing.

Originally Posted by NikosD

In my opinion, Skylake will come after desktop Broadwell due to the well known Intel's delays, but that's only a speculation.

Broadwell-K and Skylake-S are both scheduled for H2 2015. Broadwell-K might come 2-3 months earlier though.

[JohnLai]

Originally Posted by wizziwig

Only on Broadwell. Haswell is limited to 8-bit only.

https://communities.intel.com/message/273816#273816

>.< Deep inside my heart, I really hope for Haswell 10bit support.....I guess this is it...I can only wait for either AMD or Nvidia gpu product with fixed hardware acceleration to come out.

BTW, nobody wanna report the mis-exposure of 10bit HEVC DXVA interface for Haswell IGPU at Intel forum?

[nevcairiel]

Originally Posted by NikosD

OK, but Samsung NX1 clips are not 10bit, they are 8bit.

I fixed the Samsung clips, they work fine now.

For 10bit..
Intel broke their drivers, similar to the problem that NVIDIA had quite recently.

For some reason, EVR tries to accept a 10-bit/16-bit input (P016 in this case), but it fails at it, resulting in overall playback failure.

NVIDIA had the same problem a few weeks back, when a driver update caused P010 to be accepted by EVR, but EVR did not handle it properly and failed.

Someone should bring this to Intels attention.
You can work around this problem by disabling P010 and P016 output in LAV if you use EVR.