In case anyone is interested, below is a table of the
run times of the various Polyhedron 2005 benchmarks compiled
with -ffast-math -funroll-loops -msse3 -O3 and the additional
flags noted for each column.
pb05_test gcc43 gcc44 gcc44 gcc44 gcc44 gcc44
graphite graphite graphite graphite
loop-block loop-block loop-block
loop-interchange
loop-interchange
loop-stripmine
ac 13.04 12.74 12.74 12.74 12.74 12.74
aermod 27.42 25.40 25.12 25.12 25.12 25.12
air 7.53 6.71 6.72 6.72 6.72 6.72
capacita 53.36 50.67 50.70 41.77 41.61 41.64
channel 3.39 1.81 3.07 3.07 3.06 3.07
doduc 35.10 34.70 34.58 34.61 34.59 34.59
fatigue 10.80 10.17 10.09 10.09 10.09 10.09
gas_dyn 7.71 7.62 13.67 13.69 13.67 13.66
induct 48.41 15.26 15.02 15.02 15.02 15.02
linpk 15.50 15.45 15.45 15.46 15.41 15.45
mdbx 12.89 12.19 12.19 12.19 12.19 12.19
nf 27.10 25.25 25.50 25.52 25.37 25.39
protein 41.41 39.63 39.74 39.74 39.75 39.73
rnflow 32.58 30.68 30.71 32.18 32.09 32.11
test_fpu 10.79 10.44 10.72 10.76 10.76 10.76
tfft 2.13 2.02 2.02 2.02 2.02 2.02
The most significant improvement in gcc 4.4 is in the induct benchmark
which is 3.17 fold faster. The first graphite column represents the use
of -fgraphite-identity which causes the GIMPLE loop representations to be
converted to graphite polyhedrals and then back again to GIMPLE. This is
supposed to provide additional detail on the loops but (as you can see)
curently causes some degrations because of lost information. However there
is at least one instance of a significant win with -fgraphite-identity
-floop-block for the capacita benchmark which improves by ~20%. The
current usage of graphite in gcc 4.4 is very conservative. We will should
see the performance regressions eliminated and many more instances of
performance improvements in gcc 4.5.
Jack
ps One other improvement in gcc 4.4 is the availability of the
-static-libgfortran
flag on darwin so that quasi-static fortran binaries can now be created.
------------------------------------------------------------------------------
_______________________________________________
Fink-devel mailing list
[email protected]
http://news.gmane.org/gmane.os.apple.fink.devel
