[patch] Make prefetching parameters controlable with --param

Zdenek Dvorak rakdver@atrey.karlin.mff.cuni.cz
Sun Nov 12 19:08:00 GMT 2006


Hello,

> Presumably the new k8 value of 100 simultaneous prefetches (a significant
> increase from the previous value of 6) has demonstrable advantages for
> some benchmark?

here are the results of stream benchmark (-O2 -march=k8
-fprefetch-loop-arrays):

before:

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:        2290.1313       0.0140       0.0140       0.0140
Scale:       2244.8565       0.0143       0.0143       0.0143
Add:         2482.1733       0.0194       0.0193       0.0194
Triad:       2432.7146       0.0197       0.0197       0.0198

after:

Function      Rate (MB/s)   Avg time     Min time     Max time
Copy:        2375.4509       0.0135       0.0135       0.0135
Scale:       2323.3923       0.0139       0.0138       0.0148
Add:         2642.0119       0.0182       0.0182       0.0182
Triad:       2597.6954       0.0186       0.0185       0.0191

(the improvements are consistent over several runs).

Zdenek



More information about the Gcc-patches mailing list