[MPlayer-dev-eng] xvidix threading

[Howard Chu]

(Resending after subscribing...) I've been messing around with my Dvico FusionHDTV tuner card and 
running into performance problems. Playing standard-definition channels is easy enough, but playing 
a full 1920x1080i channel was just too much. Using xvidix/radeon for output...

oprofile showed that the bulk of the execution time is in memcpy, from vidix_draw_slice:

CPU: AMD64 processors, speed 2600 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit mask of 0x00 (No unit
mask) count 50000
samples  %        image name               app name                 symbol name
179887   40.7671  mplayer                  mplayer                  fast_memcpy
49657    11.2536  mplayer                  mplayer                  swScale_MMX2
29557     6.6984  mplayer                  mplayer                  mmxext_idct
20101     4.5554  mplayer                  mplayer                  hcscale_MMX2
15602     3.5358  mplayer                  mplayer                  slice_intra_DCT
15397     3.4894  mplayer                  mplayer                  MC_put_o_16_mmxext
14941     3.3860  mplayer                  mplayer                  mpeg2_slice
11475     2.6005  mplayer                  mplayer                  motion_fr_frame_420
9799      2.2207  mplayer                  mplayer                  get_non_intra_block

CPU: AMD64 processors, speed 2600 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit mask of 0x00 (No unit
mask) count 50000
samples  %        image name               app name                 symbol name
-------------------------------------------------------------------------------
    142       0.4817  mplayer                  mplayer                  vidix_draw_slice_420
    29337    99.5183  mplayer                  mplayer                  fast_memcpy
29479    22.3475  mplayer                  mplayer                  fast_memcpy_SSE
    29479    100.000  mplayer                  mplayer                  fast_memcpy_SSE [self]
-------------------------------------------------------------------------------
    73        0.9778  mplayer                  mplayer                  mpeg2_slice
    7393     99.0222  mplayer                  mplayer                  slice_non_intra_DCT
7466      5.6598  mplayer                  mplayer                  get_non_intra_block
    7466     100.000  mplayer                  mplayer                  get_non_intra_block [self]
-------------------------------------------------------------------------------

I tried scaling it down to 960x540 which was enough to bring the CPU usage down to around 95% or so,
but it was still maxing out occasionally. Since I've got a dual core processor and the other core
was idle, I tried moving the draw_slice function into a separate thread. This turned out to work
pretty well; instead of maxing out at 100% of one core I can now manage higher resolutions without
any frame losses (using both cores). (I also tried to use the -dr option but it only worked at a few
choices of image sizes. Not sure what the story is there.)

The attached diff (against current svn) is just for your consideration, for real use it would need a
commandline switch to toggle it on/off. By the way, it seems to have some sync problems when used 
without the scale filter.

-- 
    -- Howard Chu
    Chief Architect, Symas Corp.  http://www.symas.com
    Director, Highland Sun        http://highlandsun.com/hyc/
    Chief Architect, OpenLDAP     http://www.openldap.org/project/

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: vidixdif.txt
URL: <http://lists.mplayerhq.hu/pipermail/mplayer-dev-eng/attachments/20070916/4c04af73/attachment.txt>