Serial Test notes:

Compiler versions & options:
  EKO2.0, PGI5.2-4, and gcc 3.4.3, 
  g95 version info: gcc version 3.5.0 20040824 
                    (experimental) (g95!) Dec 18 2004

f77 and f90 flags:
  pathf90 -O3 -WOPT:aggstr=off
  pgf90 -O3 -fastsse
  g77 -O4
f90 flags:
  pathf90 -O3 -WOPT:aggstr=off -OPT:unroll_size=256
  pgf90 -O3 -fastsse
  g95 -O4
C flags:
  pathcc -Ofast -LNO:prefetch=0 -OPT:unroll_size=256 -m32
  pgcc -O3 -Munroll=c:256 -tp k8-32
  gcc -O4 -funroll-loops -m32

System Under Test:  
  1 x Opteron 248 2.2Ghz CPU
  4x512, DDR400, PC3200, Corsair CL2 Memory
  SuSE Linux 9.1 64 bit Kernel 2.6.4-52-default

Source code:
  Serial F77, F90 and C versions are the 'M'edium or 'dynamic allocate' 
  versions from this page:
  http://w3cic.riken.go.jp/HPC/HimenoBMT/program1.htm


=================================================

OpenMP Parallel test notes:


Compiler versions & options:
  EKO 2.0, PGI5.2-4

F77/OpenMP Flags
  pathf90 -O3 -WOPT:aggstr=off -mp
  pgf90 -O2 -Munroll -Mnoframe -Mscalarsse -Mvect=sse -Mcache_align -Mflushz -mp

System Under Test:
    4 x 2.2 Ghz Opteron 848 CPUs
    16x1024, DDR400, PC3200  Memory
    Fedora Core 2 64 bit Kernel 2.6.8-52

Source Code:
Original F77 OpenMP version (M data size) is from this page:
http://w3cic.riken.go.jp/HPC/HimenoBMT/program2.htm

Original Himeno source    -- himenobmtxp_m_omp.f
PathScale modified source -- himenobmtxp_m_omp.1st.touch.f

PathScale's parallel first touch data initialization changes to improve scalability on NUMA machines:

$ diff himenobmtxp_m_omp.f himenobmtxp_m_omp.1st.touch.f
137a138,140
> C$OMP PARALLEL SHARED (kmax,jmax,imax,nn,a,p,b,c,bnd,wrk1,wrk2,
> C$OMP&  omega,gosa) PRIVATE (k,j,i)
> C$OMP DO
156a160
> C$OMP END DO NOWAIT
157a162
> C$OMP DO
176a182,183
> C$OMP END DO
> C$OMP END PARALLEL


Differences to change timers:
$ diff  himenobmtxp_m_omp.f.orig himenobmtxp_m_omp.f
77c77
<       cpu0=dtime(time0)
---
>       cpu0=etime(time0)
82c82
<       cpu1= dtime(time1)
---
>       cpu1= etime(time1)
96c96
<       cpu0=dtime(time0)
---
>       cpu0=etime(time0)
99c99
<       cpu1= dtime(time1)
---
>       cpu1= etime(time1)
112c112
<       pause
---
> c      pause