[prev in list] [next in list] [prev in thread] [next in thread] 

List:       gcc-patches
Subject:    Re: PATCH: Update default arch for x86
From:       Jack Howarth <howarth () bromo ! med ! uc ! edu>
Date:       2010-02-28 22:26:28
Message-ID: 20100228222628.GA15251 () bromo ! med ! uc ! edu
[Download RAW message or body]

On Sun, Feb 28, 2010 at 09:57:09AM -0800, H.J. Lu wrote:
> 
> -mtune=core2 was added by Apple. I found it was slower than
> -mtune=generic on Linux/Core 2. I don't know how it performs
> against -mtune=generic on Darwin.
> 
> -- 
> H.J.

HJ,
   According to the polyhedron 2005 benchmark runtimes, it
is slower with -mtune=core2 on darwin10. These are on a late 2008
MacPro with dual 2.8GHz quad-core Xeons.

x86_64-apple-darwin10 target
Compile Command : gfortran -mtune=core2 -ffast-math -funroll-loops -O3 %n.f90 -o %n

benchmark      -mtune=generic   -mtune=core2
ac                 9.71             9.37
aermod            20.95            20.88
air                5.67             5.70
capacita          33.13            34.35
channel            1.84             1.84
doduc             27.71            28.60
fatigue            8.12             8.28
gas_dyn            4.42             4.36
induct            12.92            13.21
linpk             15.43            15.46
mdbx              11.36            14.17
nf                28.04            28.23
protein           36.15            36.18
rnflow            23.65            23.28
test_fpu           8.69             8.75
tfft               1.88             1.89

geo. mean         11.21            11.41

i686-apple-darwin10 target
Compile Command : gfortran -mtune=core2 -ffast-math -funroll-loops -O3 %n.f90 -o %n

benchmark      -mtune=generic   -mtune=core2
ac                 12.45           12.84 
aermod             24.97           25.44
air                 6.69            6.65
capacita           46.24           46.42
channel             2.17            2.16
doduc              34.30           33.89
fatigue            10.55           10.50
gas_dyn             5.26            5.29
induct             14.81           14.81
linpk              15.66           15.68
mdbx               12.04           14.17
nf                 25.86           25.74
protein            41.23           41.53
rnflow             26.81           26.18
test_fpu           10.77           10.79
tfft                2.07            2.07
 
geo. mean          13.05           13.19

  The mdbx benchmark seems to be particularly
degraded with -mtune=core2.
               Jack
ps I looked at the core2 cost settings in Apple's
gcc 4.2.1 source code and they appear to be identical
to those in gcc trunk with the exception of the new
entries for...

+  32,                                   /* size of l1 cache.  */
+  2048,                                 /* size of l2 cache.  */

in gcc trunk.
[prev in list] [next in list] [prev in thread] [next in thread]