On Fri, Dec 11, 2015 at 01:32:45PM +0100, Matthieu Bouron wrote: > --- > > Hello, > > The following two commits aim to add a video frame pool to AVFilter. > This feature was originally introduced by commit > 39f66edbeae5ccabefe38b2fcb25d6c242d868c0 but was removed later on when the > AVFilter API was changed to use AVFrame. > > The AVVideoFramePool implementation is similar to the FramePool implementation > from libavcodec (except it's only for video), meaning that the pool keeps > references > on the underlying buffers rather than the avframes themselves. The pool keeps > reference to "frames" with the same size, format and alignement. > > On the AVFilter side, I choose to put a video frame pool per AVFilterLink to > keep it > simple and more clear (to me at least) (and avoid looping on a list of frames > of > different sizes, like it was done before). This is indeed at the cost of > potentially > using more memory depending on the scenario. > > Having a video frame pool avoid some malloc/free calls but also benefits a > lot when > writing to memory on arm devices (see benchmark below). > > Here are some benchmark performed on a BeagleBone Black board: > > Without the video frame pool: > > ./ffmpeg -f lavfi -i "testsrc2=1920x1080,format=nv12,format=rgba" -f null -t > 1 - > nvx->rgbx took 24615.000000 > nvx->rgbx took 24617.000000 > nvx->rgbx took 24601.000000 > nvx->rgbx took 24675.000000 > nvx->rgbx took 25723.000000 > nvx->rgbx took 24535.000000 > nvx->rgbx took 24592.000000 > nvx->rgbx took 24637.000000 > nvx->rgbx took 24567.000000 > nvx->rgbx took 24641.000000 > nvx->rgbx took 25521.000000 > nvx->rgbx took 24670.000000 > nvx->rgbx took 24598.000000 > nvx->rgbx took 24547.000000 > nvx->rgbx took 24644.000000 > nvx->rgbx took 25414.000000 > nvx->rgbx took 25099.000000 > nvx->rgbx took 25146.000000 > nvx->rgbx took 25447.000000 > nvx->rgbx took 25094.000000 > nvx->rgbx took 25436.000000 > nvx->rgbx took 25393.000000 > nvx->rgbx took 25411.000000 > nvx->rgbx took 25124.000000 > nvx->rgbx took 25139.000000 > frame= 25 fps=8.4 q=-0.0 Lsize=N/A time=00:00:01.00 bitrate=N/A > > With the video frame pool: > > ./ffmpeg -f lavfi -i "testsrc2=1920x1080,format=nv12,format=rgba" -f null -t > 1 - > nvx->rgbx took 10430.000000 > nvx->rgbx took 10476.000000 > nvx->rgbx took 10428.000000 > nvx->rgbx took 10462.000000 > nvx->rgbx took 11090.000000 > nvx->rgbx took 11777.000000 > nvx->rgbx took 10845.000000 > nvx->rgbx took 10924.000000 > nvx->rgbx took 10840.000000 > nvx->rgbx took 10921.000000 > nvx->rgbx took 10892.000000 > nvx->rgbx took 10927.000000 > nvx->rgbx took 10857.000000 > nvx->rgbx took 10900.000000 > nvx->rgbx took 10878.000000 > nvx->rgbx took 10745.000000 > nvx->rgbx took 10935.000000 > nvx->rgbx took 10433.000000 > nvx->rgbx took 10470.000000 > nvx->rgbx took 10412.000000 > nvx->rgbx took 10425.000000 > nvx->rgbx took 10429.000000 > nvx->rgbx took 10832.000000 > nvx->rgbx took 10413.000000 > nvx->rgbx took 10423.000000 > frame= 25 fps= 12 q=-0.0 Lsize=N/A time=00:00:01.00 bitrate=N/A >
So this explains our wonder about why more than half of the time was spent in the writing itself and not the computation. -- Clément B.
signature.asc
Description: PGP signature
_______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
