# Invocation command line:
# /cpu2006/bin/runspec -a validate -c pathscale-amd64.cfg -T all fp
# output_root was not used for this run
############################################################################
# AMD64 (64-bit) Linux Pathscale v3.0 config file 
#  for CPU2006
#
# If you are building and running on a newer Linux distro that 
# uses the GCC v4.x compilers by default, then you need to either
# use "--define gnu4_fe" on the runspec command line or uncomment 
# the "%define gnu4_fe" line below:
#
# %define gnu4_fe
# %define gnu3_fe
#
#  gnu3_fe is the default, so really does not need to be defined.
#####################################################################

flagsurl000=http://www.spec.org/cpu2006/flags/CPU2006_flags.20090714.13.xml
tune          = base
reportable    = yes
ext           = amd64
output_format = asc,html
teeout        = yes
teerunout     = yes
mean_anyway   = yes

# Include machine-specific section
#include: SUT.inc
#  ----- Begin inclusion of 'SUT.inc'
############################################################################
################################################################
# Machine-specific section
################################################################
hw_cpu_name  = AMD Opteron 8220
hw_cpu_mhz   = 2800
hw_disk      = 1 x 36 GB SAS, 10000 RPM
hw_fpu       = Integrated
hw_memory    = 32 GB (16 x 2GB DDR2-5300 ECC)
hw_vendor    = IBM Corporation
hw_model     = IBM BladeCenter LS41 (AMD Opteron 8220)
hw_avail     = Feb-2007
hw_nchips    = 4
hw_ncores    = 8
hw_ncoresperchip= 2
hw_nthreadspercore = 1
hw_ncpuorder = 1, 2, 3, 4 chips
hw_pcache    = 64 KB I + 64 KB D on chip per core
hw_scache    = 1 MB I+D on chip per core
hw_tcache    = None
hw_ocache    = None
hw_other     = None
sw_file      = ext3
sw_os        = SLES 10 (x86_64), 2.6.16.21-0.8-smp
sw_state     = Multi-user, run level 3
makeflags     = -j2
EXTRA_LDFLAGS =  -IPA:max_jobs=2

use_submit_for_speed = 1
# For rate runs, comment out previous 2 lines and uncomment the following line
submit= MYMASK=`printf '0x%x' \$((1<<\$SPECCOPYNUM))`; /usr/bin/taskset \$MYMASK $command
notes_000= taskset utility used to bind CPU(s) to processes
notes_005 =DSPEC_CPU_TABLE_WORKAROUND was used for portability when compiling 447.dealII
notes_010 =due to compilation being performed on SLES 9 SP3
# ---- End inclusion of '/cpu2006/config/SUT.inc'

#####################################################
license_num    = 11
sw_compiler000 = QLogic PathScale
sw_compiler001 =  Compiler Suite, Release 3.0
sw_other       = MicroQuill SmartHeap 8.1
sw_auto_parallel    = No
sw_avail       = Mar-2007
test_date      = Jul-2007
prepared_by    = IBM Corporation
test_sponsor   = IBM Corporation
tester         = IBM Corporation


default=default=default:
#####################################################################
# Compiler selection
#
#
CC           = pathcc
CXX          = pathCC
FC           = pathf95

#  If compiling on a machine NOT having the target machine's
#  AMD64 architecture, then uncomment the following
#  compiler define variables.
# CC           = pathcc -march=opteron
# CXX          = pathCC -march=opteron
# FC           = pathf95 -march=opteron

SMARTHEAP_DIR = /tools/SmartHeap_8.1/lib


#####################################################################
# Portability for default 64-bit code generation
#####################################################################

default=default=default:
PORTABILITY = -DSPEC_CPU_LP64

#####################################################################
# INT Portability
#####################################################################

400.perlbench=default=default:
CPORTABILITY = -DSPEC_CPU_LINUX_X64

462.libquantum=default=default:
CPORTABILITY=  -DSPEC_CPU_LINUX

471.omnetpp=default=default:
# Needed to avoid -DSPEC_CPU_LP64 on -m32 C++ base codes
PORTABILITY =

473.astar=default=default:
# Needed to avoid -DSPEC_CPU_LP64 on -m32 C++ base codes
PORTABILITY =

483.xalancbmk=default=default:
CXXPORTABILITY= -DSPEC_CPU_LINUX
# Needed to avoid -DSPEC_CPU_LP64 on -m32 C++ base codes
PORTABILITY =

#################################################################
#                                                               #
#          SPECint Tuning & Notes                               #
#                                                               #
#################################################################

int=base:
COPTIMIZE    = -Ofast -OPT:malloc_alg=1
CXXOPTIMIZE  = -Ofast -m32
EXTRA_CXXLIBS= -L$(SMARTHEAP_DIR) -lsmartheap
sw_base_ptrsize  = 32/64-bit
sw_peak_ptrsize  = 32/64-bit

#####################################################################
# INT Peak Tuning
#####################################################################

int=peak=default:
COPTIMIZE    = -Ofast -OPT:malloc_alg=1
CXXOPTIMIZE  = -Ofast -m32
EXTRA_CXXLIBS= -L$(SMARTHEAP_DIR) -lsmartheap
PASS1_CFLAGS  = -fb_create fbdata 
PASS1_CXXFLAGS  = -fb_create fbdata 
PASS1_LDFLAGS = -fb_create fbdata
PASS2_CFLAGS  = -fb_opt fbdata
PASS2_CXXFLAGS  = -fb_opt fbdata
PASS2_LDFLAGS = -fb_opt fbdata

400.perlbench=peak=default:
COPTIMIZE=-Ofast -LNO:opt=0

401.bzip2=peak=default:
COPTIMIZE=-O3 -LNO:ou_prod_max=10 -OPT:Ofast:alias=disjoint
feedback=0

403.gcc=peak=default:
PORTABILITY =
COPTIMIZE=-m32 -O3 -OPT:Ofast

429.mcf=peak=default:
PORTABILITY =
COPTIMIZE=-m32 -O3 -ipa
feedback=0
EXTRA_CLIBS = -L$(SMARTHEAP_DIR) -lsmartheap

445.gobmk=peak=default:
COPTIMIZE=-O3 -OPT:alias=disjoint -LNO:simd=0:minvariant=off -WOPT:retype_expr=on

456.hmmer=peak=default:
COPTIMIZE=-O2 -OPT:alias=disjoint:malloc_alg=1 -CG:cflow=0
feedback=0

458.sjeng=peak=default:
COPTIMIZE=-O3 -IPA:plimit=50000 -IPA:pu_reorder=2

462.libquantum=peak=default:
COPTIMIZE=-O3 -ipa -CG:local_fwd_sched=on -IPA:space=1000
feedback=0

464.h264ref=peak=default:
COPTIMIZE=-O3 -IPA:plimit=20000 -OPT:alias=disjoint -LNO:prefetch=0

471.omnetpp=peak=default:
CXXOPTIMIZE= -Ofast -CG:gcm=off -m32
feedback=0

473.astar=peak=default:
basepeak=true

483.xalancbmk=peak=default:
CXXOPTIMIZE=-Ofast -m32 -OPT:unroll_times_max=8
feedback=0

#####################################################################
# FP Portability
#####################################################################

436.cactusADM=default=default:
FPORTABILITY=  -fno-second-underscore		   

%ifdef %{gnu4_fe}
447.dealII=default=default:
CXXPORTABILITY =
%else
447.dealII=default=default:
CXXPORTABILITY = -DSPEC_CPU_TABLE_WORKAROUND
%endif

481.wrf=default=default:
FPORTABILITY= -fno-second-underscore
CPORTABILITY= -DSPEC_CPU_LINUX

#################################################################
#                                                               #
#                        SPECfp Tuning                          #
#                                                               #
#################################################################

fp=base:
COPTIMIZE    = -Ofast
CXXOPTIMIZE  = -Ofast
FOPTIMIZE    = -Ofast -OPT:malloc_alg=1

sw_base_ptrsize  = 64-bit
sw_peak_ptrsize  = 32/64-bit

#####################################################################
# FP Peak Tuning
#####################################################################

fp=peak=default:
COPTIMIZE    = -Ofast
CXXOPTIMIZE  = -Ofast
FOPTIMIZE    = -Ofast -OPT:malloc_alg=1
PASS1_CFLAGS   = -fb_create fbdata
PASS2_CFLAGS   = -fb_opt fbdata
PASS1_CXXFLAGS = -fb_create fbdata
PASS2_CXXFLAGS = -fb_opt fbdata
PASS1_FFLAGS   = -fb_create fbdata
PASS2_FFLAGS   = -fb_opt fbdata
PASS1_LDFLAGS  = -fb_create fbdata
PASS2_LDFLAGS  = -fb_opt fbdata

410.bwaves=peak=default:
FOPTIMIZE=-O3 -OPT:Ofast:IEEE_arith=3 -LNO:blocking=off:ignore_feedback=off

416.gamess=peak=default:
FOPTIMIZE=-O2 -OPT:Ofast:ro=3:unroll_size=256

433.milc=peak=default:
COPTIMIZE=-Ofast -CG:cflow=off -LNO:prefetch=1 -OPT:malloc_alg=1
feedback=0

434.zeusmp=peak=default:
FOPTIMIZE=-Ofast -CG:local_fwd_sched=on -LNO:blocking=off:interchange=off:fu=10:full_unroll_outer=on
feedback=0

435.gromacs=peak=default:
FOPTIMIZE=-O3 -OPT:rsqrt=2:ro=3
COPTIMIZE=-O3 -OPT:rsqrt=2:ro=3
feedback=0

436.cactusADM=peak=default:
COPTIMIZE=-O3 -LNO:prefetch=3:prefetch_ahead=5:ou_prod_max=10:full_unroll=5 -ipa
FOPTIMIZE=-O3 -LNO:prefetch=3:prefetch_ahead=5:ou_prod_max=10:full_unroll=5 -ipa

437.leslie3d=peak=default:
basepeak=true

444.namd=peak=default:
CXXOPTIMIZE=-Ofast -fno-exceptions

447.dealII=peak=default:
# Needed to avoid -DSPEC_CPU_LP64
PORTABILITY =
%ifdef %{gnu4_fe}
CXXOPTIMIZE=-Ofast -static -INLINE:aggressive=on -OPT:malloc_alg=1 -m32 -fno-exceptions
feedback=0
%else
CXXOPTIMIZE=-Ofast -INLINE:aggressive=on -LNO:opt=0 -OPT:alias=disjoint -m32 -fno-exceptions
feedback=0
%endif

450.soplex=peak=default:
CXXOPTIMIZE=-m32 -O3 -OPT:IEEE_arith=3 -CG:load_exe=0:movnti=1 -LNO:minvariant=off:prefetch=1 -fno-exceptions
# Needed to avoid -DSPEC_CPU_LP64 
PORTABILITY =

453.povray=peak=default:
CXXOPTIMIZE=-Ofast -fno-fast-math

454.calculix=peak=default:
FOPTIMIZE=-Ofast -LNO:simd=0 -WOPT:mem_opnds=on
COPTIMIZE=$(FOPTIMIZE)
feedback=0

459.GemsFDTD=peak=default:
FOPTIMIZE=-Ofast -LNO:fission=2:prefetch=0
feedback=0

465.tonto=peak=default:
basepeak=true

470.lbm=peak=default:
basepeak=true

481.wrf=peak=default:
basepeak=true

482.sphinx3=peak=default:
%ifdef %{gnu4_fe}
COPTIMIZE=-O3 -OPT:Ofast -WOPT:aggstr=0 -m32
%else
basepeak=true
%endif