Kondo timings
         From: original table sent K. Haule 99/02/02


Summary

    (1) KCC compiles this code better than gcc or CC.
    (2) This contradicts the HINT benchmark which favors gcc against KCC 
        on dune (Sun 4500 336 MHz). 
	(see http://Tink.ijs.si/~mark/dune-gcc-kcc.ps, also
	referred to at http://www-f1.ijs.si/~krivec/benchmarks.html).
    (3) The relative performance of kondo, using KCC, is correlated with 
        the CLOCK FREQUENCY (i.e., saturn (75 MHz R8000) behaves as if it 
	had only a two-way pipeline). Using gcc, it is correlated with the
	MFLOPS ratings. 



saturn:    SGI Power Challenge, 75-90 MHz R8000
dune:      Sun Enterprise 4500, 336 MHz
iapetus:   SGI O2 180MHz



Table I.  kondo: N=8, Nf=4, S=0 (K. Haule).  RC is 
10**5 divided by the CPU time and clock frequency. 
RF is 10**5 divided by the CPU time and theoretical 
MFLOPS rating (which is clock * pipeline).
----------------------------------------------------
                        time    rsize       RC    RF
                         [s]     [MB]  
----------------------------------------------------
dune    KCC               12                25    12  *1
        gcc               24                12     6

saturn  KCC               48                25     6  *2
        gcc               62                19     5
        CC (-n32)         79                15     4
	
iapetus CC                56                 9     5
----------------------------------------------------
*1  Two-way pipeline.
*2  Four-way pipeline.


R. Krivec