Message-ID: <38BFC3C4.BCB4CB5D@ife.ee.ethz.ch> Date: Fri, 03 Mar 2000 14:53:08 +0100 From: Alex Rhomberg <rhomberg@ife.ee.ethz.ch> Subject: Re: [suse-axp] Good news!
Mattias Hermansson wrote:
>
> I compiled lame 3.51 with ccc. The speedincrease was exceptional, more
> than 3 times the speed of gcc. That's how it's supposed to be. Look here:
Only three times means that the code was not well optimised ;-)
Check this out:
starting point 15 us / Element (on 500 MHz ev56)
End point after pipeline stalls caused by aliasing problems were
removed:
(on 666 MHz ev67)
g++ -O3
Timer 1 integrate & scatter: 7.462 us (pitiful 107.2 MFLOPS!)
ccc -O3
Timer 1 integrate & scatter: 1.537 us (mindboggling 520.5 MFLOPS!)
(on 500 MHz ev56)
g++ -O3
Timer 1 integrate & scatter: 12.226 us (pitiful 65.4 MFLOPS!)
ccc -O3
Timer 1 integrate & scatter: 3.083 us (mindboggling 259.5 MFLOPS!)
A slight difference of a factor 4 to 5 with well good code. The bad code
I had before had about a factor of two difference
- Alex
This archive was generated by hypermail 2.1.0 : Mon Jun 04 2001 - 04:18:04 PDT