[mythtv] [patch] XVMC_decode_mb

Daniel Thor Kristjansson danielk at cat.nyu.edu
Tue Aug 24 10:39:58 EDT 2004


So my XVMC_decode_mb function is almost twice as fast as the current
CVS version. By necessity this patch also includes the calc_cbp patch.
I used one SDTV stream and one HDTV stream, and tried both MC and
IDCT acceleration.

Profile Data (IDCT acceleration, HDTV content)
cvs XVMC_decode_mb
IDCT HDTV 3878      4.2047  libmythavcodec-0.15.so.0.15.0 XVMC_decode_mb

new XVMC_decode_mb
IDCT HDTV 553       0.6085  libmythavcodec-0.15.so.0.15.0 XVMC_decode_mb
IDCT HDTV 384       0.4225  libmythavcodec-0.15.so.0.15.0 calc_cbp
IDCT HDTV 346       0.3807  libmythavcodec-0.15.so.0.15.0 handle_p_b_block
IDCT HDTV 314       0.3455  libmythavcodec-0.15.so.0.15.0 setup_context
IDCT HDTV 238       0.2619  libmythavcodec-0.15.so.0.15.0 setup_pmv
IDCT HDTV 146       0.1606  libmythavcodec-0.15.so.0.15.0 set_block_pattern
IDCT HDTV 117       0.1287  libmythavcodec-0.15.so.0.15.0 handle_intra_block
                   --------
                    2.3084%

The fourth column is percentages of time spent in mythtv. It's probably the
easiest to compare, and smaller is better. Everything is compiled for
pentium4 and SSE2. The benefit is less with MC acceleration because that
path spends much of it's time doing the IDCT transform, so XVMC_decode_mb
is less important overall.

The full profiles are available here:
  http://www.mrl.nyu.edu/~danielk/mythtv/xvmc_decode_mb-profile.tbz

-- Daniel
-------------- next part --------------
A non-text attachment was scrubbed...
Name: xvmc_decode_mb-v1.patch.bz2
Type: application/octet-stream
Size: 5439 bytes
Desc: 
Url : http://mythtv.org/pipermail/mythtv-dev/attachments/20040824/39f4167c/xvmc_decode_mb-v1.patch.obj


More information about the mythtv-dev mailing list