------- Comment #7 from rguenth at gcc dot gnu dot org 2009-12-01 16:46 -------
Just reverting rev. 154688 and using the training set gets us from
464.h264ref -- 228 -- S
to
464.h264ref -- 170 -- S
at -O3 -ffast-math -funroll-loops -march-native (-march=k8-sse3 -msahf --param
l1-cache-size=64 --param l1-cache-line-size=64 --param l2-cache-size=512
-mtune=k8).
After the patch the oprofile looks like
CPU: AMD64 processors, speed 2000 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit mask
of 0x00 (No unit mask) count 100000
samples % symbol name
2241755 51.1739 SetupFastFullPelSearch
442008 10.0900 BlockMotionSearch
271429 6.1961 SubPelBlockMotionSearch
269337 6.1483 FastPelY_14
170961 3.9026 UMVLine16Y_11
159311 3.6367 SetupLargerBlocks
155556 3.5510 SATD
127230 2.9044 FastLine16Y_11
72328 1.6511 dct_luma
69728 1.5917 getNonAffNeighbour
All but the *Y_1[14] functions are inside mv-search.c, just re-compiling
that file is enough to reproduce the issue.
after reverting it
CPU: AMD64 processors, speed 2000 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Cycles outside of halt state) with a unit mask
of 0x00 (No unit mask) count 100000
samples % symbol name
1223269 36.8289 SetupFastFullPelSearch
450306 13.5573 BlockMotionSearch
320395 9.6461 SubPelBlockMotionSearch
245175 7.3815 FastPelY_14
160694 4.8380 SetupLargerBlocks
155353 4.6772 SATD
153292 4.6152 UMVLine16Y_11
79531 2.3944 FastLine16Y_11
72455 2.1814 dct_luma
70733 2.1296 getNonAffNeighbour
42828 1.2894 UMVPelY_14
34396 1.0356 UnifiedOneForthPix
--
rguenth at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
Summary|[4.5 Regression] rev |[4.5 Regression] rev 154688
|15458[78] regress |regress 464.h264ref peak 20%
|464.h264ref peak 20% |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42216
