[FFmpeg-devel] H.264 profiling (was: IRC smearing)

Jason Garrett-Glaser darkshikari
Sat Jan 16 08:22:36 CET 2010


L2 miss profile: http://pastebin.com/f7bca111e
L1D miss profile: http://pastebin.com/f744303c2
L1I miss profile: http://pastebin.com/f5b3a34e

(Thanks to saintdev)

Things to note:

1) The L1D misses are pretty much exactly as one would think: an
enormously large number, but entirely in motion compensation.
2) The L2 misses are bizarre: there are more misses in
ff_h264_decode_mb_cabac than anything else--a function which should
have practically no L2 misses.  This seems potentially problematic.
3) The L1I misses are also potentially problematic.  ff_h264_filter_mb
and ff_h264_decode_mb_cabac both have huge numbers of L1I misses,
suggesting that the code might benefit from some compaction.

Dark Shikari



More information about the ffmpeg-devel mailing list