Polyhedron 2005 Benchmarks

[Dominique Dhumieres]

Paul,

your results look like:

	     g95       gfc      ifc

ac           43.44    33.45    22.83
aermod      115.12   105.47   108.72 
air          66.91    49.94    48.30 
capacitor   270.45   245.17   266.55 
channel     119.58    65.69    89.81
doduc       106.78    88.73    92.39 
fatigue      83.26    29.03    22.40
gas_dyn      79.33    45.58    63.35
induct      153.96    71.66    96.89
linpk       129.91   114.28   124.71
mdbx         64.94    64.17    65.61
nf          193.69   201.18   130.06
protein     167.25   123.41   119.37
rnflow      136.43   120.68   122.26
test_fpu    101.43    87.77    87.34
tfft         29.30    28.14    28.27


mine are:

		  1.8Ghz G5, OSX 10.3.9            2.0Ghz AMD64, Linux
		     512k L2 cache                   1024k L2 cache

		g95     gfc    xlf-3   xlf-5      g95     gfc     pgf     ifc
			    
ac              40.2    42.9    32.4    17.8*     24.1    18.0    29.5    21.2   
aermod         115.7   107.8    76.2    76.6      65.0    err.    57.3*   71.6   
air             49.0    48.4    42.6    25.3      25.6    21.1    19.0    17.1*  
capacita       206.5   193.6   174.1   171.9     127.8   112.5    76.6*  132.2    
channel         65.6    16.5*   40.4    42.0      30.9    17.6    20.6    20.4   
doduc           76.1    72.5    73.9    58.7      65.4    52.5    45.3*   57.3  
fatigue         86.1    34.5    18.4    11.3'*    44.6    25.5    13.1    20.1  
gas_dyn         66.8    55.7    39.9    30.4      27.3    21.0    10.7*   11.3  
induct         140.6   295.2    50.8    42.8'*    83.7   225.6    47.9    95.5  
kepler          88.9    95.0    46.9    41.5     106.4    93.9    28.8*   34.2  
linpk           35.8    35.2    34.7    34.2*     43.9    39.5    41.4    39.5
mdbx            34.9    33.7    29.2    28.2      32.4    30.0    25.9*   27.5
nf             103.5   116.9    40.3    39.5      60.1    68.7    39.4*   42.5 
protein        107.5    69.9    82.4    85.1      93.3    66.3*   79.4    72.8
rnflow          85.3    81.7    63.5    63.1      57.5    61.4    51.4    48.4*
test_fpu        52.6    49.1    32.5    40.0      34.1    28.8    26.2    24.0*  
tfft            17.2    17.9    16.4    16.2      10.0*   10.2    10.6    10.1

* fastest timing.
' module free_input compiled with -O3 and the remaining with -O5

It seems that your AMD is slower than mine and have a smaller L2 cache.
If so, it confirms that some of the Polyhedron tests are sensitive to the 
L2 cache size (I have proof of that for tfft).

I also confirm that mod() is slow in g95 and gfc (g77 as well), that there
is a large benefit to inline dot_product (at least for small vectors).  

In addition replacing 3 dot_products by a matmul of a 3x3 matrix times a
length 3 vector gives catastrophic results (with all compilers) and
g95 and gfc are unable to reorder loops such as

do i = 1, ni
  do j = 1, nj
     a(i,j) = b(i,j)
   end do
end do

Such constructs are used almost everywhere in air.f90 (on purpose to test
compilers or because of a C programmer?). A manual reordering of the most 
obvious loops gives:

air.v1          31.4    25.9    24.9    25.3      24.1    19.2    17.3    17.2*  

(almost a factor 2 for gfc on G5 -> cache effect again?)

Dominique

PS1. The latest available build of x86_64-unknown-linux-gn seems to be 
2005/10/24. Is there an other source than quatramaran.ens.fr? If no, where 
can I find instructions on how to build it from sources as a non-root user?

PS2. In my test the flags are:

-O3 -ffast-math -funroll-loops
    for g95 and gfortran (gfc) (plus -i4 for tfft with g95 on AMD), 
-O3 or -O5 for xlf (plus -qsave for test_fpu.f90: 
    on OSX 10.3.9 stacksize has a 64Mbytes hard limid, if someone knows 
    how to change it, I'll be glad to know!-), 
-Bstatic -V -fastsse -Munroll=n:4 -Mipa=fast,inline -tp k8-64 
    for the Portland Group fortran (pgf), and
-O3 -xW -ipo -static -V 
    for the Intel fortran (ifc, no -ipo for aermod).

The compiler versions are:

on G5

G95 Fortran 95 version 4.0.2 (g95!) Nov 14 2005 (powerpc-apple-darwin7.9.0)
GNU F95 version 4.1.0 20051112 (experimental) (powerpc-apple-darwin7)
xlf beta

on AMD

G95 Fortran 95 version 4.0.1 (g95!) Nov 11 2005 (x86_64-unknown-linux-gnu)
GNU F95 version 4.1.0 20051024 (experimental) (x86_64-unknown-linux-gnu)
Intel(R) Fortran Compiler for 32-bit applications, Version 8.1    Build 20041019Z Package ID: l_fc_pu_8.1.021
PGF90/x86 Linux/x86-64 6.0-2

I recently realized that some bus errors I got with xlf (air, linpk, nf,
rnflow, and test_fpu) and ifc (capacita and nf) were due toa too small
stack size.  This was solved with

limit stacksize unlimited       (csh)
ulimit -s unlimited             (sh)