Improved Multi-GPU Video Acceleration Support?

[quamt]

Hello @rigaya,

I saw that Intel is implementing Improved Multi-GPU Video Acceleration Support for FFMpeg encoder.

`2023Q1:
FFmpeg:

ffmpeg-qsv and ffmpeg-vaapi added hstack_qsv, vstack_qsv and xstack_qsv filter support
ffmpeg-qsv and ffmpeg-vaapi added gopconcat bsf support. This filter will concat gop from different streams to make a new stream, which helps to run encoder on multi-gpu platform
ffmpeg-vaapi support RGB* format clips for vp9 decode`

I wanted to know if this would be possible to implement. https://github.com/intel/cartwheel-ffmpeg/releases/tag/2023q1

[rigaya]

This is actually already discussed here.

As it is not part of oneVPL and requires complicated implementation, I'll not try to focus into that. For iGPU and dGPU case, we already have --hyper-encode, and currently I don't think it is useful to think of a case of multi Arc GPU environment.

[quamt]

Ok. I didn't see that in that specific topic.

'--hyper-encode' is slowing the encodes down, at least for me, and doesn't work with all settings applied, unfortunately.

[rigaya]

If you don't get speed up from --hyper-encode, that will be the same with the implementation you've mentioned, as the concept of speed up is the same.

It seems to require following conditions for speed up.

• dGPU Arc A770 or A750 (A380 is known to not get speed up from --hyper-encode)
• Small GOP (below 60)
• Large async depth (avobe GOP size, around 60)

Check the guide from oneVPL for detailes.

[quamt]

The test system I tried this on is like below:

Device #1: Intel Arc A770 Graphics
Device #2: Intel Arc A770 Graphics
Device #3: Intel UHD Graphics 770

API 2.08: hw(mfx-gen), Acceleration d3d11
API 2.08: hw(mfx-gen), Acceleration d3d11
API 2.08: hw(mfx-gen), Acceleration d3d11

QSVEncC (x64) 7.39 (r3079) by rigaya, May 10 2023 09:16:12 (VC 1934/Win)
 Intel Media SDK API v2.09
 reader: raw, avi, avs, vpy, avsw, avhw
Environment Info
OS : Windows 11 x64 (22621) [UTF-8]
CPU: 13th Gen Intel Core i5-13600 [4.71GHz] (6P+8E,14C/20T)
RAM: Used 10798 MB, Total 64329 MB
GPU: Intel Arc A770 Graphics (512EU) 300-2400MHz (31.0.101.4335)

I also checked the information provided but I don't see as to why it should slow down using --hyper-mode.

I was looking for the Multi-GPU support as the system would fulfil all the requirements as far as I can see.

[rigaya]

As I have said earlier, there is no plan to add multi-GPU Acceleration.