[FFmpeg-user] overlay_qsv with hardware decode and encode

[Vladimir Mishonov]

On 2025-02-28 22:00, Dragan Randjelovic wrote:

> Hi Vladimir,
> 
> I have noticed that you were strugling, try using the following:
> 
> scale_vaapi=w=-1:h=-1 ensures hardware scaling before overlay  (source 
> resolution is kept with -1)

Thank you for your reply. I'm verry sorry to say, however, that it 
doesn't help me much, if at all:

> ffmpeg -init_hw_device vaapi=hw -filter_hw_device hw \
> -i rtmp://localhost/stream1 \
> -i rtmp://localhost/stream2 \
> -filter_complex "
> [0:v]format=nv12,hwupload=extra_hw_frames=20,scale_vaapi=w=-1:h=-1[i0];
> [1:v]format=nv12,hwupload=extra_hw_frames=20,scale_vaapi=w=-1:h=-1[i1];
> [i0][i1]overlay_vaapi"
> -c:v h264_vaapi -global_quality 25 -g 50 -r 25 -an -f null -

This command works perfectly already without any "scale_vaapi" filter 
regardless of the input streams' resolutions. However, it uses software 
decoding, which is evident by the lack of "-hwaccel vaapi 
-hwaccel_output_format vaapi" arguments for inputs, and also by the use 
of the hwupload filters.

Setting extra_hw_frames=20 does reduce memory usage indeed. I've seen 
the recommendation to keep it relatively high (around 100) to reduce the 
risk of issues during encoding. But I don't much care about memory usage 
anyway since I have a lot of memory to spare.
Just in case there's been a misunderstanding, let me repeat the command 
that uses a purely hardware transcoding pipeline, and which works only 
when input streams have the same resolution:

ffmpeg \
  -hwaccel vaapi -hwaccel_output_format vaapi -i rtmp://localhost/stream1 
\
  -hwaccel vaapi -hwaccel_output_format vaapi -i rtmp://localhost/stream2 
\
  -filter_complex "[0:v][1:v]overlay_vaapi=w=500" \
  -c:v h264_vaapi -global_quality 25 -g 50 -r 25 -an -f null -

And no, putting "scale_vaapi" before overlay here does not help.

P.S. I'm wondering if this recommendation was written by AI, by the way.

> Regards
> Dragan Randelovic
> 
> On Fri, Feb 28, 2025 at 6:37 PM Vladimir Mishonov via ffmpeg-user 
> <ffmpeg-user at ffmpeg.org> wrote:
> 
>> On 2025-02-28 19:53, Greg Oliver wrote:
>>> On Thu, Feb 27, 2025 at 11:49 PM Vladimir Mishonov via ffmpeg-user <
>>> ffmpeg-user at ffmpeg.org> wrote:
>>> 
>>> Sorry - that link I referenced actually did not have the whole 
>>> thread.
>>> The
>>> piece I intended to be included was:
>>> 
>>> https://lists.ffmpeg.org/pipermail/ffmpeg-user/2024-September/058652.html
>>> 
>> 
>> Unfortunately, the -noautoscale option does not work either, I've
>> already tried that before.
>> 
>> It is, however, not surprising that it has no effect because
>> -noautoscale is an output option. And in my scenario, the output
>> resolution does not change over time.
>> 
>> As for the patch linked in that thread, I have yet to try it. Not sure
>> if it's already been merged into master; if not, it means I have to
>> compile a build myself, and setting everything up for that is likely
>> going to take a while...
>> 
>>> If that does not solve it - sorry - I do not use QSV (except in Plex)
>>> for
>>> anything - I just remembered seeing that email thread a while back, 
>>> so
>>> looked it up.  This one has a resolution, but I am unsure if it is
>>> software
>>> or hardware.
>>> 
>> 
>> No problem. What I still cannot understand is whether this is some 
>> kind
>> of fundamental limitation or an outright bug. For some reason, I could
>> find very little information about this particular scenario. Whenever 
>> I
>> search for that error message ("impossible to convert between..."), it
>> always seems to concern a completely different use case.
>> 
>>> 
>>> _______________________________________________
>>> ffmpeg-user mailing list
>>> ffmpeg-user at ffmpeg.org
>>> https://ffmpeg.org/mailman/listinfo/ffmpeg-user
>>> 
>>> To unsubscribe, visit link above, or email
>>> ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".
>> _______________________________________________
>> ffmpeg-user mailing list
>> ffmpeg-user at ffmpeg.org
>> https://ffmpeg.org/mailman/listinfo/ffmpeg-user
>> 
>> To unsubscribe, visit link above, or email
>> ffmpeg-user-request at ffmpeg.org with subject "unsubscribe".