vpp_qsv tonemapping and color space conversion does not change metadata

[Nathan Fendley]

Summary of the bug:

When you use these options in vpp_qsv on any Intel GPU:

out_color_matrix  <string>     ..FV....... Output color matrix coefficient set
out_color_primaries <string>     ..FV....... Output color primaries
out_color_transfer <string>     ..FV....... Output color transfer characteristics
tonemap           <int>        ..FV....... Perform tonemapping (0=disable tonemapping, 1=perform tonemapping if the input has HDR metadata) (from 0 to 1) (default 0)

the metadata of the output is not changed to match the new color information. On the other hand, tonemap_vaapi does this properly. The video outputs look identical when played in VLC between vpp_qsv and tonemap_vaapi but not identical to playing the original HDR input, so something more than just incorrect metadata may be going on.

How to reproduce:

Use a system with an Intel media engine that has both the VAAPI and QSV APIs working properly, and use any HDR video that's possible to decode with the media engine.

# Does not work as intended, outputs a video that *says* it's BT.2020 with HDR metadata but the video still comes out right
% ffmpeg -hwaccel qsv -hwaccel_output_format qsv -i HWDEC_HDR_INPUT.mkv -vf vpp_qsv=scale_mode=hq:format=p010:tonemap=1:out_color_matrix=bt709:out_color_primaries=bt709:out_color_transfer=bt709 -c:v hevc_qsv -global_quality:v 26 -an -sn OUTPUT_BAD_META.mkv

# Works as intended, produces correctly tagged BT.709 SDR
% ffmpeg -hwaccel vaapi -hwaccel_output_format vaapi -i HWDEC_HDR_INPUT.mkv -vf tonemap_vaapi=format=p010:m=bt709:p=bt709:t=bt709 -c:v hevc_vaapi -rc_mode ICQ -q:v 26 -an -sn OUTPUT_PROPER.mkv

ffmpeg version n7.1 Copyright (c) 2000-2024 the FFmpeg developers
  built with gcc 14.2.1 (GCC) 20250207
  from Arch Linux repos

Relevant mediainfo from source:

Color range                              : Limited
Color primaries                          : BT.2020
Transfer characteristics                 : PQ
Matrix coefficients                      : BT.2020 non-constant
Mastering display color primaries        : Display P3
Mastering display luminance              : min: 0.0050 cd/m2, max: 1000 cd/m2
Maximum Content Light Level              : 186 cd/m2
Maximum Frame-Average Light Level        : 147 cd/m2

from vpp_qsv tonemap (metadata I believe to be false):

Color range                              : Limited
Color primaries                          : BT.2020
Transfer characteristics                 : PQ
Matrix coefficients                      : BT.2020 non-constant
Mastering display color primaries        : Display P3
Mastering display luminance              : min: 0.0050 cd/m2, max: 1000 cd/m2
Maximum Content Light Level              : 186
MaxCLL_Original                          : 186 cd/m2
Maximum Frame-Average Light Level        : 147
MaxFALL_Original                         : 147 cd/m2

from tonemap_vaapi:

Color range                              : Limited
Color primaries                          : BT.709
Transfer characteristics                 : BT.709
Matrix coefficients                      : BT.709

[nyanmisaka]

Tone-mapping in the vpp_qsv filter requires at least an Intel 11th gen CPU (Xe/UHD7xx), not just any Intel GPU.

​https://github.com/FFmpeg/FFmpeg/blob/52eb0e18db27bf608403675b0a618880009a301f/libavfilter/vf_vpp_qsv.c#L405

[Nathan Fendley]

Replying to nyanmisaka:

Tone-mapping in the vpp_qsv filter requires at least an Intel 11th gen CPU (Xe/UHD7xx), not just any Intel GPU.

​https://github.com/FFmpeg/FFmpeg/blob/52eb0e18db27bf608403675b0a618880009a301f/libavfilter/vf_vpp_qsv.c#L405

I'm running this on an Intel Arc Battlemage B580 - in addition to that, tonemap_vaapi does in fact work correctly on my laptop, which has an i7-8650U/UHD 620. It must have some sort of software/GPU compute fallback with VAAPI.