[FFmpeg-devel] [PATCH 2/2] lavc/vaapi_hevc: Don't require exact profiles
Mark Thompson
sw at jkqxz.net
Sun Apr 28 17:14:36 EEST 2024
On 24/04/2024 14:45, Xiang, Haihao wrote:
> On Ma, 2024-04-22 at 22:23 +0100, Mark Thompson wrote:
>> Rather than turning the constraint flags into a single profile and then
>> searching for that profile (and failing if it doesn't match any profile
>> exactly), instead search all supported profiles and use the first one
>> which supports the given set of constraint flags.
>> ---
>> This fixes decode of rext 8-bit 4:2:0; it will correctly pick Main 12 or Main
>> 4:2:2 10 or Main 4:4:4 (first one available) to use as the decoding profile
>> after this patch.
>
> sw decoding and vaapi decoding might have different bits (There is the same
> issue if applying Fei's patchset).
>
> For example:
> $ ffmpeg -hwaccel vaapi -f lavfi -i testsrc -vf 'format=nv12,hwupload' -c:v
> hevc_vaapi -profile:v rext -vframes 30 -y out.mp4
>
> 8bit ouput if using sw decoding:
> $ ffmpeg -i out.mp4 -f null -
> [...]
> Stream #0:0(und): Video: wrapped_avframe, yuv420p(tv, progressive), 320x240 [SAR
> 1:1 DAR 4:3], q=2-31, 200 kb/s, 25 fps, 25 tbn (default)
>
> 12bit output if using vaapi decoding:
> $ ffmpeg -hwaccel vaapi -i out.mp4 -f null -
> [...]
> Stream #0:0(und): Video: wrapped_avframe, p012le(tv, progressive), 320x240 [SAR
> 1:1 DAR 4:3], q=2-31, 200 kb/s, 25 fps, 25 tbn (default)
That comes from what the driver reports support for, though?
E.g. with the Intel iHD driver, I have
{
"profile": 23,
"name": "HEVCMain12",
...
"surface_formats": [
{
"rt_format": "YUV420",
"max_width": 16384,
"max_height": 16384,
"memory_types": [
"VA",
"DRM_PRIME_2",
],
"pixel_formats": [
"P012",
"P016",
],
},
So to decode a 4:2:0 8-bit input it is asking for a P012 or P016 surface.
If the driver reported that a Main 12 profile 4:2:0 8-bit input could be decoded to an NV12 surface then it would be picked by the logic in vaapi_decode_find_best_format(), since it would be a better match than the higher-depth formats.
Thanks,
- Mark
More information about the ffmpeg-devel
mailing list