Trying to use the GPU on a headless server to decode and encode

[twilson90]

trafficstars

I have a server with a Intel® Xeon® Processor E3-1245 v5, which has an onboard GPU, Intel® HD Graphics P530 with QuickSync.: https://www.intel.com/content/www/us/en/products/sku/88173/intel-xeon-processor-e31245-v5-8m-cache-3-50-ghz/specifications.html I'd like to use the GPU when possible to decode and/or encode video using MPV. I've tried running some tests using the latest FFMPEG and after much trying I managed to decode some video with the GPU but it required a very precise set of parameters. What parameters are necessary to use a GPU with QuickSync for decoding and encoding in MPV, and do I need to build it with any particular libraries enabled to get it working?

[CounterPillow]

On a Linux system, you're unlikely wanting to go through the QuickSync layer but instead want to go through va-api instead, which pokes precisely the same hardware but through a more commonly used path.

You can specify to use hardware decoding through vaapi with --hwdec=vaapi. You can manually specify a device with --vaapi-device, which defaults to /dev/dri/renderD128. If this is not the right one for you, you can select one of the other devices in that subdirectory.

As for encoding with vaapi, --ovc=<codecdname>_vaapi might suffice, see --ovc=help for a list. You can set libavcodec (i.e. ffmpeg) options through --ovcopts, see --ovcopts=help for a list.

Please note that mpv's encode mode bypasses mpv's rendering last I checked, so if you're after video rendering quality then you're not getting any advantages over ffmpeg that way (you can use libplacebo as a vf in either ffmpeg or mpv though).

[twilson90]

Thanks, that's a very helpful response. I'm using MPV primarily as a way of encoding a dynamic playlist for streaming. Not really interested in enhancing the video if that's what you mean. In my build (forked from about 2 years ago) I'm not seeing h264_vaapi in my list of codecs, I'm assuming I need to build it with an additional library? Any idea how I would do that?

[CounterPillow]

meson setup -Dvaapi=enabled build to create a meson configuration with vaapi forcibly enabled in ./build. Usually, it is auto-detected if present. You'll also need to have libva installed, and I believe the ffmpeg you build mpv against needs to enable vaapi as well.

In my build (forked from about 2 years ago)

Speaking of meson, that may have been pre-meson, in which case you need to see what the corresponding waf switch was (though in that case it was also autodetected)

[twilson90]

ffmpeg builds ok with vaapi enabled but mpv is complaining: You manually enabled the feature 'vaapi', but the autodetection check failed.

Log: mpv build log.txt

vainfo:

error: XDG_RUNTIME_DIR not set in the environment.
error: can't connect to X server!
libva info: VA-API version 1.14.0
libva info: Trying to open /usr/lib/x86_64-linux-gnu/dri/iHD_drv_video.so
libva info: Found init function __vaDriverInit_1_14
libva info: va_openDriver() returns 0
vainfo: VA-API version: 1.14 (libva 2.12.0)
vainfo: Driver version: Intel iHD driver for Intel(R) Gen Graphics - 22.3.1 ()
vainfo: Supported profile and entrypoints
      VAProfileMPEG2Simple            : VAEntrypointVLD
      VAProfileMPEG2Main              : VAEntrypointVLD
      VAProfileH264Main               : VAEntrypointVLD
      VAProfileH264Main               : VAEntrypointEncSliceLP
      VAProfileH264High               : VAEntrypointVLD
      VAProfileH264High               : VAEntrypointEncSliceLP
      VAProfileJPEGBaseline           : VAEntrypointVLD
      VAProfileJPEGBaseline           : VAEntrypointEncPicture
      VAProfileH264ConstrainedBaseline: VAEntrypointVLD
      VAProfileH264ConstrainedBaseline: VAEntrypointEncSliceLP
      VAProfileVP8Version0_3          : VAEntrypointVLD
      VAProfileHEVCMain               : VAEntrypointVLD

[Andarwinux]

hwdec=vaapi-copy

[CounterPillow]

@twilson90 Looks like your mpv version is too old to have separated vaapi from X11. Current mpv distinguishes between vaapi on X11, vaapi on DRM and vaapi on Wayland.

@Andarwinux your suggestion is completely unrelated to a build issue.

[twilson90]

I've got it building with vaapi enabled on an updated version of my fork now, but it's untested. Is there no way to build my older fork with vaapi then?

[twilson90]

Issues with new build:

root@ns3053860:/opt/mpv-build/mpv/build# ./mpv --hwdec=vaapi --vaapi-device=/dev/dri/renderD128 --o=out.mkv --of=matroska --ovc=h264_vaapi --ovcopts-add=profile=high --ovcopts-add=preset=veryfast --ovcopts-add=level=4 --ovcopts-add=b=4000k --oac=aac --oacopts=ac=2,b=160k "input.mkv"
 (+) Video --vid=1 (*) (h264 1620x1080 30.000fps)
 (+) Audio --aid=1 (*) (aac 2ch 44100Hz)
[hwupload] no hw context
[hwupload] hardware format not supported
[autoconvert] Failed to create HW uploader for format yuv420p
[autoconvert] can't find video conversion for yuv420p
Cannot convert decoder/filter output to any format supported by the output.
Could not initialize video chain.
[vo/lavc] error encoding at EOF
Video: no video
[ao/lavc] Opening encoder: AAC (Advanced Audio Coding) [aac]
AO: [lavc] 44100Hz stereo 2ch floatp

[philipl]

You do need to take @Andarwinux 's suggestion. If you are transcoding with no vo, the non-copy hwdec cannot be initialised - it also doesn't offer any benefits, as the whole point of non-copy hwdecs to get zero-copy output to the display.

[twilson90]

The only benefit I'm after is reducing the load on the CPU and onto the GPU. I've tried --hwdec=vaapi-copy and --vo=gpu (although this should be set by default) and still no luck.

EDIT: I've managed to get ffmpeg transcoding with vaapi, still not sure how to do this with mpv: /opt/mpv-build/ffmpeg_build/ffmpeg -hwaccel vaapi -hwaccel_device /dev/dri/renderD128 -hwaccel_output_format vaapi -i input.mkv -c:v h264_vaapi -b:v 4M -maxrate 4M -rc_mode CQP output.mp4

[philipl]

Then you should just keep using ffmpeg. You can basically do anything you want that way, and it's actually intended for transcoding, vs mpv, for which the encoding support is a highly questionable feature.

You can use -vf libplacebo with ffmpeg to get the advanced post-processing you might imagine wanting from mpv.

[twilson90]

It might be possible to pass the job of encoding to FFMPEG, but not possible for decoding. As I stated earlier, I'm using a single instance of MPV as a sort of dynamic playlist containing various media files, which is transcoded and uploaded to an RTMP server in real-time. I need to be able to change volume, seek, change video / audio filters and lots more, on the fly while transcoding, seamlessly. MPV is a very robust video player and it can do all of this, FFMPEG cannot (without a lot of work). It's currently running and this is the front-end interface, if you're interested:

[philipl]

When used for encoding mpv always uses a special encoding vo. This means that no feature that depends on vo=gpu or vo=gpu-next is going to work. And the zero-copy hwdecs require state provided by the vo. While it's possible to make vo=lavc (the encoder) provide that state (which is basically what ffmpeg is doing), that functionality doesn't exist today, so you will be limited to using vaapi-copy. In practice, that will prevent you from doing a full GPU transcode without bouncing through system memory. But that is unlikely to prevent your workflow from working.