[FFmpeg-devel] [PATCH 1/3] libavcodec/vaapi_encode: Change the way to call async to increase performance

Mark Thompson sw at jkqxz.net
Mon Dec 27 20:31:50 EET 2021


On 27/10/2021 09:57, Wenbin Chen wrote:
> Fix: #7706. After commit 5fdcf85bbffe7451c2, vaapi encoder's performance
> decrease. The reason is that vaRenderPicture() and vaSyncSurface() are
> called at the same time (vaRenderPicture() always followed by a
> vaSyncSurface()). When we encode stream with B frames, we need buffer to
> reorder frames, so we can send serveral frames to HW at once to increase
> performance. Now I changed them to be called in a
> asynchronous way, which will make better use of hardware.
> 1080p transcoding increases about 17% fps on my environment.
> 
> Signed-off-by: Wenbin Chen <wenbin.chen at intel.com>
> ---
>   libavcodec/vaapi_encode.c | 41 ++++++++++++++++++++++++++++-----------
>   libavcodec/vaapi_encode.h |  3 +++
>   2 files changed, 33 insertions(+), 11 deletions(-)

The API does not allow this behaviour.

For some bizarre reason (I think a badly-written example combined with the Intel driver being synchronous in vaEndPicture() for a long time), the sync to a surface is to the /input/ surface of an encode rather than the output surface.

That means you can't have multiple encodes outstanding on the same surface and expect to sync usefully, because the only argument to vaSyncSurface() is the surface to sync to without anything about the associated context.

Therefore trying to make it asynchronous like this falls down when input surfaces might appear multiple times, or might be used in the input of multiple encoders, because you can't tell whether your sync means the thing you actually wanted to finish has finished.

(The commit you point to above as having decreased performance fixed this bug, since it became much more visible with decoupled send/receive.)

So: put this change after the switch to syncing on output buffers (since that operation does make sense for this), and leave the existing behaviour for cases where you have to sync on the input surface.

- Mark


More information about the ffmpeg-devel mailing list