[FFmpeg-user] 0.10.3 to 0.11 performance regression?

Carsten Mattner carstenmattner at googlemail.com
Sun Jun 3 13:33:16 CEST 2012


I cannot get 0.11 or git master to perform the same speed of
conversion of a video file into an mp3 audio track via
ffmpeg -i file.ext -ab nnnk file.mp3

0.10.3 used to take a few seconds and while converting prints
size=    1193kB time=00:01:01.04 bitrate= 128.1kbits/s
as the progress line.

real    0m15.799s
user    0m14.502s
sys     0m0.213s

0.11 and git master print some orange colored text which
are maybe warnings and the speed is many many times slower:
<snip>
[buffersink @ 0x90e2b00] No opaque field provided
[format @ 0x90e2ba0] auto-inserting filter 'auto-inserted scaler 0'
    between the filter 'Parsed_null_0' and the filter 'format'
[scale @ 0x90e2ec0] w:854 h:480 fmt:yuv420p sar:0/1 -> w:854
    h:480 fmt:rgb24 sar:0/1 flags:0x4
<snip>
[mp3 @ 0x93b59a0] Got more than one picture in stream 0, ignoring.
frame=   31 fps=8.6 q=0.0 Lsize=     865kB time=00:00:01.03
    bitrate=6849.6kbits/s
<wait 16 minutes>
frame= 8873 fps=9.0 q=0.0 Lsize=    6629kB time=00:04:56.06
    bitrate= 183.4kbits/s
video:7350440kB audio:5785kB global headers:0kB muxing
    overhead -99.909891%

real    16m24.756s
user    16m10.617s
sys     0m9.119s

I've bisected this in the git repo and the result is:
2dfea1205870c298461f08fc323c652be41888f8 is the first bad commit
commit 2dfea1205870c298461f08fc323c652be41888f8
Author: Anton Khirnov <anton at khirnov.net>
Date:   Tue Feb 28 09:52:05 2012 +0100

    mp3enc: write attached pictures (APIC).

:040000 040000 7777ee70463deb64976b9f383d8b37ce9192122d
e3e753fcbd2c3ed07152b7324249a894b422afcb M      libavformat


Another question: when bisecting I have constantly hit git revisions where
the build results were avconv other executabled instead of ff* executables.
Is that due to temporary merge commits to/from libav.git?


More information about the ffmpeg-user mailing list