This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
RE: [patch] k8 and amdfam10 cost model tuning
- From: "Sjodin, Jan" <Jan dot Sjodin at amd dot com>
- To: "Uros Bizjak" <ubizjak at gmail dot com>, "GCC Patches" <gcc-patches at gcc dot gnu dot org>
- Date: Thu, 20 Dec 2007 12:43:28 -0600
- Subject: RE: [patch] k8 and amdfam10 cost model tuning
- References: <476A80A1.7030509@gmail.com>
>
> Please note that -ftree-vectorize -fvect-cost-model is now
> implicit in
> -O3 and -msse3 is implicit in "-march=k8-sse3" as well as
> "-march=amdfam10". This can save you some typing...
>
Yes, I already had this in the configuration file and I
didn't want to modify it. ;-) Same thing below, but I wanted to keep
the flags identical with the bug report, although I added -static.
> BTW: Is it possible for you to confirm 30% performance regression in
> rnflow polyhedron benchmark for -m32 -mfpmath=387 on k8
> target as seen
> in [1]. This problem seems specific to k8 target, everything works
> normal on Core2Duo. This performance regression is covered by PR
> target/31897 [2].
>
> [1]: http://www.physik.fu-berlin.de/~tburnus/gcc-trunk/benchmark/
> [2]: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31897
>
> Thanks,
> Uros.
I ran rnflow with an unmodified compiler:
(Rev: 130934 Last Changed Date: 2007-12-14)
I could not duplicate the performance regression using the flags
from pr31897. Below is a run on a 2.4 GHz Athlon 64 x2:
20 Dec 2007 11:28:05 standardgfortransingle - Compile rnflow
command=gfortran -static -m32 -march=opteron -ffast-math -funroll-loops
-ftree-l
oop-linear -ftree-vectorize -mfpmath=sse -msse3 -O3 -w rnflow.f90 -o
rnflow
20 Dec 2007 11:28:20 standardgfortransingle - Execute rnflow
rnflow Run # 10 37.26570 37.23942 - Error= 0.0253%
Geometric Mean Execution Time = 37.24 seconds
command=gfortran -static -m64 -march=opteron -ffast-math -funroll-loops
-ftree-l
oop-linear -ftree-vectorize -mfpmath=sse -msse3 -O3 -w rnflow.f90 -o
rnflow
20 Dec 2007 11:37:09 standardgfortransingle - Execute rnflow
rnflow Run # 10 35.94570 35.99346 - Error= 0.0253%
Geometric Mean Execution Time = 35.99 seconds
Is anyone else able to reproduce the problem?
Thanks,
Jan