[Feature]: DeepLink HyperEncode for Linux

[bezirg]

What Feature?

Reading from the following source, it seems that there is only support for Windows for Deep Link Hyper Encode:

FFmpeg/FFmpeg@5002829
https://github.com/oneapi-src/oneVPL-intel-gpu/blob/main/doc/HyperEncode_FeatureDeveloperGuide.md

Since oneVPL is supported for Linux, will also support for HyperEncode in Linux come later?
Or is there some outside dependency that makes this currently infeasible (e.g. Vulkan or VA-API ?)

What's the usage scenario would be benifited?

Transcode for media delivery, Content Creation

What impacted?

No response

Do you want to contribute a patch to develop this feature?

None

[DenWolf]

@daleksan - please help to clarify the plan of enabling HyperEncode on Linux if any

[daleksan]

Hello @bezirg,

Sorry for delay with response. At this moment there are no plans to enable this feature on Linux. Focus on current implementation improvements.

[mswdev]

+Support for this to come to Linux. Seems like a no-brainer for this to be supported on Linux for improving transcoding with the iGPU and Intel Arc GPUs working in tandem with Deep Link.

[sauravbandyopadhyay]

Hello @bezirg,
Can you share with us more details and "use cases" for your requirement on "DeepLink HyperEncode for Linux"? May we request you to respond to the other questions left blank i.e. "What impacted?" and "Do you want to contribute a patch to develop this feature?"? Thanks.

[mswdev]

Hello @bezirg, Can you share with us more details and "use cases" for your requirement on "DeepLink HyperEncode for Linux"? May we request you to respond to the other questions left blank i.e. "What impacted?" and "Do you want to contribute a patch to develop this feature?"? Thanks.

I'm not @bezirg , but support for Deep Link on Linux would be particularly useful for realistically all the same reasons it is useful for Windows.

Particularly, I would find it useful for transcoding media content on Linux for my personal media library. I have a home server running Unraid which is unix-like Slackware. The next Unraid release 7.0.0 coming this fall supports Intel Arc and with that being added it would be particularly useful to use deep link to pair an Intel iGPU with an Intel Arc GPU.

Aside from my reasons, gaming on Linux is growing quite rapidly with the help from Valve and their project Proton which brings compatibility to Linux for games that are exclusive to Windows. I could also see Linux Deep Link support being useful for creators who produce content or even streamers.

While support for Intel Arc on the Linux Kernel is relatively new I see Intel Arc being a rapidly growing category for Linux especially with the advent of Deep Link being considered.

[PilaScat]

Is useful for people like me that use Frigate with multiple 4K streams, Plex and tdarr simultaneously, all of them use ffmpeg hw accelerated

[kallaballa]

Hi!

Would someone be so kind to elaborate a bit on the challenges on porting this? Would make things a lot easier since I might actually do it. I'm in good practice when it comes to porting to linux and i already know the intel gpu ecosystem. e.g.: intel/compute-runtime#166

[nyanmisaka]

Hi!

Would someone be so kind to elaborate a bit on the challenges on porting this? Would make things a lot easier since I might actually do it. I'm in good practice when it comes to porting to linux and i already know the intel gpu ecosystem. e.g.: intel/compute-runtime#166

@kallaballa The problem is that the only VPL runtime that supports HyperEncode is part of Intel's GPU driver on Windows, which is obviously closed source. So there is no publicly available runtime code to reference, just VPL header files. So to achieve this, you have to implement them from scratch in vpl-gpu-rt based on the HyperEncode documentation and how it behaves in Windows. This can be challenging but not impossible to do, but you must be well versed in how the QSV encoder and FFmpeg work.

IMO it is necessary to add a mechanism to vpl-gpu-rt so that it can automatically create a sub-device for HyperEncode internally and distribute the YUV/RGB frames submitted by the user (on host or device memory) as encoding tasks to the main device and sub-devices. The caveats are to use closed/strict GOP and ensure that the order of the encoded h26x data packets (DTS & PTS) is not disrupted, and to find the most effective way for encoders to access each other between host and device memory, otherwise PCIe bandwidth may become a bottleneck for high resolution videos. You will also need to prepare two Gen12 devices for testing, one from the iGPU (UHD7xx or Xe) and the other should be an Arc GPU/DG2 or DG1 dGPU.

See also, the architecture diagram between i915/xe kernel driver, vpl-gpu-rt, media-driver and applications such as FFmpeg. There is also an abandoned FFmpeg patch that implements it using a hacker method.

[kallaballa]

wow. that is a very good start. i guess i'll start reading and see what i can do. it sounds interesting :)

thx!

[kallaballa]

I guess you mean Gen12+? Because I have a Gen12, a Gen13 and a A770.

[nyanmisaka]

GenX means graphics, it’s not Xth generation processors. There’s no Gen13.

[kallaballa]

I understand. Thank you very much. Still peaked my interest and I am gonna look a bit into it.

…

On 2024-12-01 04:37, Nyanmisaka wrote: GenX means graphics, it’s not Xth generation processors. There’s no Gen13. -- Reply to this email directly, view it on GitHub [1], or unsubscribe [2]. You are receiving this because you were mentioned.Message ID: ***@***.***> Links: ------ [1] #279 (comment) [2] https://github.com/notifications/unsubscribe-auth/AACGEIX6SWZGE5B3FGYVQHT2DKABBAVCNFSM6AAAAABMQ6HL4KVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKMBZGU2TEMBSGM

[kallaballa]

For now the project is a bit to big for me, since I am well versed with FFMpeg but not at all with QSV. But i definitely would like to return and give it a try in time. also if others want to contribute i'd be happy to chime in.