Advanced Micro Devices Inc (AMD)

[chipguy]

As you see the Intel compiler properly picks the options automatically, thus making base score equal to peak.

Wrong, look at the submission file for I2. The tuning section says that they didn't bother tweak compiler
flags for individual SPEC programs.



+FDO: PASS1=-prof_gen PASS2=-prof_use
Base tuning:
C programs: -ipo -O3 +FDO -ansi_alias
Fortran programs: -ipo -O3 +FDO
Portability:
178.galgel: -FI

Peak flags same as baseline (basepeak=true set globally)

Note:
EFI shell utility 'cpuconfig' used
to deconfigure other processors



Now compare that to the Opteron submission tuning:


+FDO: PASS1=-Qprof_gen PASS2=-Qprof_use
icl and ifl are the Intel C/C++ and Fortran compilers
f90 is the Compaq Fortran compiler
shlW32M6.lib is the SmartHeap library V6.0 from MicroQuill www.microquill.com
Portability:
178.galgel: -FI -Fe$@ -link -stack:32000000
Baseline: C icl +FD0 -O3 -QxW -Qipo
Baseline: Fortran ifl +FDO -O3 -QxW -Qipo
Peak tuning:
168.wupwise: ifl +FD0 -QxK -Qipo -Ow
171.swim: f90 -Optimize:5 -alignment:dcommons -alignment:records
-alignment:sequence -architecture:k7
-assume:noaccuracy_sensitive -math_library:fast -tune:k7
172.mgrid: ifl +FD0 -O3 -QaxW -Qipo -Oa -Qprefetch-
173.applu: ifl +FD0 -O3 -QxK -Qipo -Qscalar_rep- -Zp8
177.mesa: icl +FD0 -O3 -QxW -Qipo -Oa -Qscalar_rep-
178.galgel: f90 -Optimize:5 -fast
179.art: icl -Qipo -Oa -Qunroll4 -Zp4
183.equake: icl -O3 -QxK -Qipo -Oa shlW32M6.lib -Zp4
187.facerec: ifl +FD0 -O3 -QaxW -Qipo -Qscalar_rep- -Qunroll1
188.ammp: icl -QxW -Oa
189.lucas: ifl +FD0 -O3 -QxW -Qipo -Qprefetch-
191.fma3d: ifl basepeak=1
200.sixtrack: ifl -Qipo -Oa -Zp4
301.apsi: f90 -Optimize:5 -fast
ONESTEP is used for all base and peak runs

In other words Intel doesn't see fit to play stupid compiler flag tuning whack-a-mole games to squeeze
out higher numbers for peak scores.