Since the fact that gfortran performance has improved
over the major releases, I decided to benchmark the current
releases on a MacPro with the Polyhedron 2005 benchmarks
using -ffast-math -funroll-loops -msse3 -O3. The results
are...
gcc release
gcc 4.2.4 gcc 4.3.3 gcc 4.4-pre gcc 4.3.3/ gcc 4.4-pre/
gcc 4.2.4 gcc 4.2.4
ac 15.32 13.04 12.65 0.851 0.826
aermod 22.96 26.65 25.52 1.161 1.111
air 8.15 7.58 7.10 0.930 0.871
capacita 52.16 52.45 49.56 1.006 0.950
channel 5.17 3.39 1.81 0.656 0.350
doduc 34.43 34.89 35.28 1.013 1.025
fatigue 12.57 10.92 10.13 0.869 0.806
gas_dyn 13.59 7.68 7.63 0.565 0.561
induct 28.61 48.46 14.88 1.694 0.520
linpk 15.46 15.48 15.50 1.001 1.003
mdbx 12.41 12.94 12.36 1.043 0.996
nf 27.08 27.25 25.50 1.006 0.942
protein 41.61 41.18 39.12 0.990 0.940
rnflow 35.37 32.58 30.68 0.921 0.867
test_fpu 11.94 10.79 10.50 0.904 0.879
tfft 2.13 2.11 2.08 0.991 0.977
Certainly it would be nice eventually resync llvm-gcc against something
newer.
Jack
On Jan 24, 2009, at 7:35 PM, Jack Howarth wrote:> Since the fact that gfortran performance has improved > over the major releases, I decided to benchmark the current > releases on a MacPro with the Polyhedron 2005 benchmarks > using -ffast-math -funroll-loops -msse3 -O3. The results > are... > > gcc release > gcc 4.2.4 gcc 4.3.3 gcc 4.4-pre gcc 4.3.3/ > gcc 4.4-pre/ > gcc 4.2.4 > gcc 4.2.4 > ac 15.32 13.04 12.65 0.851 > 0.826 > aermod 22.96 26.65 25.52 1.161 > 1.111 > air 8.15 7.58 7.10 0.930 > 0.871 > capacita 52.16 52.45 49.56 1.006 > 0.950 > channel 5.17 3.39 1.81 0.656 > 0.350 > doduc 34.43 34.89 35.28 1.013 > 1.025 > fatigue 12.57 10.92 10.13 0.869 > 0.806 > gas_dyn 13.59 7.68 7.63 0.565 > 0.561 > induct 28.61 48.46 14.88 1.694 > 0.520 > linpk 15.46 15.48 15.50 1.001 > 1.003 > mdbx 12.41 12.94 12.36 1.043 > 0.996 > nf 27.08 27.25 25.50 1.006 > 0.942 > protein 41.61 41.18 39.12 0.990 > 0.940 > rnflow 35.37 32.58 30.68 0.921 > 0.867 > test_fpu 11.94 10.79 10.50 0.904 > 0.879 > tfft 2.13 2.11 2.08 0.991 > 0.977 > > Certainly it would be nice eventually resync llvm-gcc against > something > newer.There are some heavy licensing issues involved in this (which I won't go into here) that make this difficult. I don't think that the problem is the version of GCC we're using as a front-end, but rather how our optimizations perform compared with GCC's optimizations. We don't really use any of GCC's optimizations for the code we generate. -bw