Evaluation of HEVC decoders (SW, Hybrid and HW)

[NikosD]

Originally Posted by nevcairiel

I fixed the Samsung clips, they work fine now.

Yes, it works.

But this 8bit HEVC clip, still crashes Intel's DXVA (both native and cb)

http://demo-uhd3d.com/files/uhd4k/Hi...-2014_HEVC.zip

[nevcairiel]

Originally Posted by NikosD

But this 8bit HEVC clip, still crashes Intel's DXVA (both native and cb)

http://demo-uhd3d.com/files/uhd4k/Hi...-2014_HEVC.zip

This one crashes somewhere deep in the driver, those take some time to figure out. Presumably because its missing a reference frame and Intel expects it to be present.

Ideally, Intel would just fail to decode the frame instead of crashing completely, but from past experience, this is a common theme with Intel .. they crash when something is wrong.

[Gravitator]

Originally Posted by mzso

ERROR 503 - Service Unavailable!

Привет,
I also have a problem (Firefox).
Use a manual download from > http://demo-uhd3d.com/fiche.php?cat=uhd&id=21

[NikosD]

Originally Posted by JohnLai

As for Ivybridge HD2500/4000 with latest driver, I verify that there is no HEVC DXVA support. (Tested it yesterday). There is no HEVC DXVA exposed as shown in DXVAChecker.

It seems that Ivybridge still uses 15.33 driver and not 15.36, so I don't expect Ivy to support DXVA HEVC acceleration.

But have you tried Lentoid HEVC OpenCL decoder or Cyberlink's PowerDVD HEVC OpenCL decoder with Ivy to see of there is any HW acceleration from there ?

Because Ivy supports OpenCL.

[ksio89]

I'm not sure if it goes here, be free to move if not.

I would like to ask you guys: is there any hw decoder available for HEVC? I have MPC-HC (x64 flavour) installed along K-Lite Codec Pack Full 10.9.5, and everything is default, so the default video decoder is LAV, and the renderer is EVC CP.

In LAV Video decoder menu, I choose DXVA2 (native) as the HW decoder, and ticked "HEVC" and "UHD (4K)" boxes, both on LAV x86 and x64 (even though I only use x64 version of MPC-HC).

I downloaded a 4K video sample from here:

http://www.imaging-resource.com/news...ailable-for-do,

The file content generated by MediaInfo is attached as a .txt.

I played this sample in MPC-HC, but DXVA was not activated, so the framerate was choppy. I played a Full HD MP4 video file just to test and DXVA worked fine, low CPU usage and higher GPU usage.

These are the specs of my laptop:

Dell Inspiron 15R-5537:

Intel core i7-4500U @ 2,7 GHz
AMD Radeon HD 8870M (128-bit, 2 GiB GDDR5)
6 GB DDR3L @ 1600 MHz
Windows 8.1 Pro x64

Does this HW support 4K DXVA decoding? It's my laptop that's too weak to play 4K video file, or I can't just set up HEVC DXVA decoding properly? Thanks in advance.

[NikosD]

Originally Posted by ksio89

I'm not sure if it goes here, be free to move if not.

I would like to ask you guys: is there any hw decoder available for HEVC? I have MPC-HC (x64 flavour) installed along K-Lite Codec Pack Full 10.9.5, and everything is default, so the default video decoder is LAV, and the renderer is EVC CP.

In LAV Video decoder menu, I choose DXVA2 (native) as the HW decoder, and ticked "HEVC" and "UHD (4K)" boxes, both on LAV x86 and x64 (even though I only use x64 version of MPC-HC).

I downloaded a 4K video sample from here:

http://www.imaging-resource.com/news...ailable-for-do,

The file content generated by MediaInfo is attached as a .txt.

I played this sample in MPC-HC, but DXVA was not activated, so the framerate was choppy. I played a Full HD MP4 video file just to test and DXVA worked fine, low CPU usage and higher GPU usage.

These are the specs of my laptop:

Dell Inspiron 15R-5537:

Intel core i7-4500U @ 2,7 GHz
AMD Radeon HD 8870M (128-bit, 2 GiB GDDR5)
6 GB DDR3L @ 1600 MHz
Windows 8.1 Pro x64

Does this HW support 4K DXVA decoding? It's my laptop that's too weak to play 4K video file, or I can't just set up HEVC DXVA decoding properly? Thanks in advance.

Your CPU has an iGPU HD 4400 inside, which is capable of HW acceleration of HEVC in both ways - OpenCL and DXVA.

Your discrete card Radeon 8870M has only OpenCL HW accelerated support of HEVC.

LAV and MPC-HC have only DXVA HW acceleration support of HEVC, not OpenCL.

If you enable your iGPU to handle video decoding, you 'll see [Playing H/W], which means HW acceleration, in the left bottom of MPC-HC by enabling Status[Control+5] to see the status of playback decoding.

If Radeon 8870M is the primary card for video playback, don't expect HW acceleration of HEVC using LAV, you have to use Cyberlink's PowerDVD HEVC decoder or Lentoid OpenCL HEVC decoder.

The HW acceleration of full HD MP4 file probably occurs because it's not a HEVC file, but H.264 (AVC) which Radeon 8870M can handle in DXVA HW acceleration.

[NikosD]

Intel put a table to clarify the situation of VP9 partial acceleration and HEVC 10bit partial acceleration for Haswell and Broadwell here:

https://communities.intel.com/thread/59216

Also, there is a post in the above link, referring to Nevcairiel's post here, about broken 10bit HEVC driver for HSW.

[xooyoozoo]

Chromium added DXVA VP8/VP9 decode acceleration support over a month ago. Based on the revision numbers, it looks like Chrome Dev and Canary should already have the acceleration activated.

Just FYI, in case anyone has Broadwell and wants to entertain our curiosity.

[nevcairiel]

Note that its not actually DXVA being used, but directly loading the special Intel decoder DLLs.

[NikosD]

Originally Posted by xooyoozoo

Chromium added DXVA VP8/VP9 decode acceleration support over a month ago. Based on the revision numbers, it looks like Chrome Dev and Canary should already have the acceleration activated.

Just FYI, in case anyone has Broadwell and wants to entertain our curiosity.

I would like to test it with my HD 4600, because Intel's drivers 15.36.xx family on June 2014 with beta v.3651 driver, introduced for the first time the VP8 MFT decoder even for Haswell.

The link you provided says:

The decoding is for Intel Broadwell and Bay Trail CPUs, which will include an
onboard hardware accelerated decoder.

IIRC, the QuickSync ASIC inside Haswell iGPU doesn't include VP8 as a HW accelerated codec.

But maybe Chrome uses that "Intel Hardware VP8 decoder MFT" that I see in MFT decoders installed and I could get some kind of HW acceleration using iGPU's shaders and VP8 streaming video.

Does latest Chrome v39 under Win 8.1 have this acceleration enabled ?

I threw a VP8 clip on Chrome and only CPU was used during the decoding.

I don't know if I have to enable a switch or something like that for VP8 HW acceleration in Chrome.

Is there any way to "force" youtube.com to send VP8 streams, because nowadays youtube uses VP9 mainly.
Or maybe change some settings in Chrome for youtube.com in order to send VP8 streams.

Originally Posted by nevcairiel

Note that its not actually DXVA being used, but directly loading the special Intel decoder DLLs.

I think Microsoft hasn't released VP8/VP9 DXVA decoder devices yet, so I wonder what's the meaning of this phrase from the above link:

Adds code to DXVA to try and load the Intel media decode DLLs

Is this even possible to use DXVA code without DXVA decoder device ?

[nevcairiel]

They just integrated this special Intel DLL code into their DXVA code in Chromium, the commit message is slightly confusing there, its not actually using anything from DXVA2 otherwise.

[NikosD]

I see.

What about Chrome and Haswell ?

Is there any chance that VP8 HW acceleration could be working ?

Update:

I have downloaded and tested latest Chromium v42 x86/x64 (21/01/2015) from here http://chromium.woolyss.com/ and it seems that even playing a demanding 4K VP8 file, from my disk not Youtube, it doesn't trigger any kind of HW acceleration.

Only CPU is involved in VP8 decoding with my signature system.

Someone with Broadwell processor, could tell us more about the subject.

[nevcairiel]

From AnandTech GTX960 launch preview:

When it comes to HEVC (H.265) decoding, GM204, GM1xx, and GK1xx GPUs all offer limited HEVC decoding via a hybrid process that offloads some parts to the GPU fixed function units and shaders, and other parts to software entirely. However as of GM206 NVIDIA’s fixed function HEVC decoder has been completed and rolled into this GPU, making this the first discrete GPU with full HEVC decoding support.

A full review is expected in the following days.

[NikosD]

Right GTPVHD.

I took a quick look on over 10 reviews, nobody has the knowledge or will to test its HTPC/Media capabilities.

My questions are for both H.264/H.265:

1) 10bit decoding
2) Speed, especially on 100+ Mbps 4K HEVC clips

and

3) What about VP8/VP9

[nevcairiel]

There is still no DXVA2 spec for VP8 or VP9, and I have see no mention of it in the CUVID decoder headers, so I don't see how it would magically appear.
Maybe in the CUDA SDK 7, should try to get my hands on that RC.

H.264 is not getting 10-bit decoding, since its not part of the base profile (like it is in H.265).

[nevcairiel]

NVIDIA GTX 960 Benchmarks

1.Beauty-2160p@30fps-12.3Mbps

LAV x86 DXVA2 Native, Decode, 125 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 125 fps

LAV x86 DXVA2 CopyBack, Decode, 118 fps

2.Fitness-2160p@30fps-8Mbps

LAV x86 DXVA2 Native, Decode, 148 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 145 fps

LAV x86 DXVA2 CopyBack, Decode, 140 fps

3.Ducks-2160p@50fps-4Mbps

LAV x86 DXVA2 Native, Decode, 153 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 152 fps

LAV x86 DXVA2 CopyBack, Decode, 143 fps

Because none of these samples posed any challenge to the decoder, I included the "Ultra HD Sample #1" from here: http://www.imaging-resource.com/news...ailable-for-do (the first video with the big leaves)
This video has a much higher bitrate as the others, at nearly 80Mbps.

SamsungNX1.Leaves.2160p@30fps-80Mbps

LAV x86 DXVA2 Native, Decode, 118 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 116 fps

LAV x86 DXVA2 CopyBack, Decode, 115 fps

-----

Its important to note that CPU usage in these tests was always below 1% (except CopyBack), indicating that full hardware acceleration is working perfectly.

From these 8-bit movies, I can conclude that the HEVC/H.265 hardware decoder in the GTX 960 is plenty fast for 4K HEVC content.
It seems to cap out around 150-160 fps for 4K, which is slightly below the Hybrid decoder on "simple" samples, however like any decent hardware decoder its largely unimpressed by high bitrates and retains most of its speed, while the hybrid decoder gets much much slower.

I'll work on implementing 10-bit support in LAV Video soon, and we'll see how it fares there (and hopefully EVR accepts 10-bit DXVA2-Native input)

Edit:
Added CopyBack numbers. Even those have greatly improved, and the performance difference between native and CB is very small (on my system, anyway).

[NikosD]

Thanks for the results.

Originally Posted by nevcairiel

NVIDIA GTX 960 Benchmarks


3.Ducks-2160p@50fps-4Mbps

LAV x86 DXVA2 Native, Decode, 153 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 152 fps

LAV x86 DXVA2 CopyBack, Decode, 143 fps

The low bitrate clips are definitely not in favor of pure HW decoders.

Nvidia GTX 770 DXVA is faster on this clip.

When I tested DXVA on hybrid decoders LAV x64 was a lot faster than LAV x86.

Of course now with the pure decoder, it shouldn't play a significant role, but if I were you, I would definitely check LAV x64 with pure HW decoder.

Originally Posted by nevcairiel

SamsungNX1.Leaves.2160p@30fps-80Mbps

LAV x86 DXVA2 Native, Decode, 118 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 116 fps

It's definitely shows its muscles here, but I would like to see DXVA performance of a hybrid decoder here.

Can you switch to hybrid decoder using shaders on 970 or there is no way to control it and you have to use the ASIC ?

Originally Posted by nevcairiel

Its important to note that CPU usage in these tests was always below 1% (except CopyBack), indicating that full hardware acceleration is working perfectly.

Can you check the video engine usage during benchmark tests ?
Not the GPU load but video engine load.

I use gadgets for this (GPU monitor) but I think GPU-Z shows that too.

Originally Posted by nevcairiel

From these 8-bit movies, I can conclude that the HEVC/H.265 hardware decoder in the GTX 960 is plenty fast for 4K HEVC content.

It looks fast like 4K 120fps decoder, but if the rumors are true about new 4K HEVC Bluray disks, we have to test it on more than 100Mbps clips up to 150Mbps and I think in 10bit and possibly 60 fps.

Originally Posted by nevcairiel

I'll work on implementing 10-bit support in LAV Video soon, and we'll see how it fares there (and hopefully EVR accepts 10-bit DXVA2-Native input)

Did your code work as it is or you had to change it for 960 ?

[nevcairiel]

Originally Posted by NikosD

It's definitely shows its muscles here, but I would like to see DXVA performance of a hybrid decoder here.

Can you switch to hybrid decoder using shaders on 970 or there is no way to control it and you have to use the ASIC ?

You don't get a choice, if it supports full HW decoding, thats what its going to do. But luckily I also have a GTX 980, which doesn't have the new decoder.

SamsungNX1.Leaves.2160p@30fps-80Mbps using the hybrid decoder on a GTX 980:
LAV x86 DXVA2 Native, Decode, 22 fps
LAV x86 DXVA2 Native, Playback (scaling to 1080p), 17 fps

LAV x64 DXVA2 Native, Decode, 24 fps
LAV x64 DXVA2 Native, Playback (scaling to 1080p), 21 fps

Not to mention that this mode maxes out one CPU core at 100% usage as well.
High bitrates have a steep performance cost in software and hybrid decoders.

Originally Posted by NikosD

Did your code work as it is or you had to change it for 960 ?

No changes were required for 8-bit decoding.

[NikosD]

Originally Posted by nevcairiel

NVIDIA GTX 960 Benchmarks

I'll work on implementing 10-bit support in LAV Video soon, and we'll see how it fares there (and hopefully EVR accepts 10-bit DXVA2-Native input)

Does DXVA checker show a DXVA device decoder of HEVC_Main10 for 960 ?

[nevcairiel]

Yes it does, with P010 output instead of NV12, so its a proper 10-bit decoder.