Log of /branches/dev-api-4/xvidcore/src/motion/sad.c
Parent Directory
Revision
1205 -
(
view)
(
download)
(
annotate)
-
[select for diffs]
Modified
Thu Nov 13 23:11:24 2003 UTC (20 years, 7 months ago) by
edgomez
File length: 6642 byte(s)
Diff to
previous 1161
MMXed the calculation of SSE for 8x8 16bit blocks. This helps quite
a lot VHQ=4 mode.
My tests show with trellis:chroma_me:
- ~20% speed improvement for vhq=4.
- at least 5% when using vhq=1.
Of course this speedup vanishes if more CPU intensive features are
used. CruNcher who used gmc/qpel, noticed "only" a ~5% speed
improvement.
NB: i'm of course talking about overall speed improvement. Such a
small patch for such a big improvement :-)
Revision
982 -
(
view)
(
download)
(
annotate)
-
[select for diffs]
Modified
Thu Apr 10 13:05:54 2003 UTC (21 years, 2 months ago) by
edgomez
File length: 6680 byte(s)
Diff to
previous 957
2003-04-10 GMT Edouard Gomez <ed.gomez@free.fr> patch-14
Summary:
Removed all ABS() macros.
Revision:
xvidcore--devapi4--1.0--patch-14
All ABS macros have been replace with their stdlib.h/math.h
equivalent. This gives a 33% overall speedup for the plain C encoder,
while the ia32 one seems to suffer a small speed loss. However this
speed loss is very small and it seems it depends on the CPU type as
the abs/fabs usage is impacting badly on sad functions but impacting
well on interpolate functions ... weird inst'it ?
modified files:
src/bitstream/mbcoding.c src/global.h src/image/image.c
src/motion/motion_comp.c src/motion/motion_est.c
src/motion/sad.c src/prediction/mbprediction.c
src/utils/mbtransquant.c src/xvid.c
This form allows you to request diffs between any two revisions of this file.
For each of the two "sides" of the diff,
enter a numeric revision.