Here are some results for cstream.c running on a generic intel motherboard
with a 90mhz pentium and 256k of secondary cache. N was changed to 512*512
(2mb per array) because otherwise the VM was thrashing.
This was compiled with Watcom C/C++ V10.0, and run under os/2
warp. The system was idle, other than network daemons, a clock, and
a cpu-usage bar.
The assignment being slower than the scale is explained by the
fact that (a) the pentium is write-back, but not write-allocate in secondary
and primary cache. This prevenets the writes from being bursted, and (b)
the compiler "optimized" the assignment into two non-floating point
operations, thus causing 32 bit accesses rather than 64.
compiled with WCL386 /fp5 /5r /oneatx /l=os2v2 cstream.c
Timing calibration ; time = 600.00 usec.
Increase the size of the arrays if this is < 300
and your clock precision is =< 1/100 second.
---------------------------------------------------
Function Rate (MB/s) RMS time Min time Max time
Assignment: 46.603 118.364 90.000 150.000
Scaling : 69.905 74.766 60.000 100.000
Summing : 69.905 120.499 90.000 130.000
SAXPYing : 69.905 117.004 90.000 160.000
I had to change the mtime() function slightly in order to
use only ANSI functions (clock() combined with CLOCKS_PER_SEC, which
is 100 under watcom).
This archive was generated by hypermail 2b29 : Tue Apr 18 2000 - 05:23:04 CDT