HDR10 10-bit HEVC encoding with FFmpeg's VAAPI encoders

Hello there,

For these who rely on FFmpeg in your workflows: You can now evaluate HDR10 HEVC 10-bit encoding with the VAAPI-based encoders.

A few examples of this in use.



Commands running in the shells in 1 and 2, to the left:

(a). Shell 1:
time ffmpeg -threads:v 4 -threads:a 12 \
-init_hw_device vaapi=va:/dev/dri/renderD128 -hwaccel vaapi -hwaccel_output_format vaapi -hwaccel_device va -filter_hw_device va \
-i 'a/sm.mkv' \
-b:v 15000k -maxrate:v 30000k -bufsize:v 2503k -c:v hevc_vaapi -sei hdr -profile:v main10 -qp:v 19 -g:v 50 -r:v 23.976 -rc_mode:v 3 -bf:v 3 -bits_per_raw_sample 10 -level:v 5.1 \
-flags -global_header+low_delay -max_muxing_queue_size 1024 -fflags +genpts -vsync 1 \
-map 0 -c:a copy -c:s copy \
-f matroska -y "b/sm.mkv"

(b). Shell 2:
time ffmpeg -threads:v 4 -threads:a 12 \
-init_hw_device vaapi=va:/dev/dri/renderD128 -hwaccel vaapi -hwaccel_output_format vaapi -hwaccel_device va -filter_hw_device va \
-i 'a/su.mkv' \
-b:v 15000k -maxrate:v 30000k -bufsize:v 2503k -c:v hevc_vaapi -sei hdr -profile:v main10 -qp:v 19 -g:v 50 -r:v 23.976 -rc_mode:v 3 -bf:v 3 -bits_per_raw_sample 10 -level:v 5.1 \
-flags -global_header+low_delay -max_muxing_queue_size 1024 -fflags +genpts -vsync 1 \
-map 0 -c:a copy -c:s copy \
-f matroska -y "b/su.mkv"

See how hard the Intel integrated GPU is being hammered, from the intel_gpu_top output above.

Condensed view, showing vainfo:



I can indeed confirm that this encoder implements HDR metadata retention, and the same can be recognized by HDR10-capable devices with zero issues.

How FFmpeg was built:

https://gist.github.com/Brainiarc7/4f831867f8e55d35cbcb527e15f9f116

Intel's Neo OCL runtime may fail to build from source (bugs everywhere!) but can be safely installed from a PPA if so needed.

The likes of Plex should pick this up so they stop stripping off HDR metadata when they use hardware-based encoders. In my testing a few months ago, the proprietary QSV-based encoders in libav and FFmpeg strip this info. Will repeat the tests again to confirm if this is the case at present.