Do we get the other 15% performance back too?

Do we get the other 15% performance back too?

[Burkhard Militzer]

Hello,

As a fan of the C++ gnu compiler project, I would like to congratulate the 
gurus on the recent performance increase probably coming from corrections 
in the prefetch routines. My benchmark shows

Processor/OS         Compiler/options             Time(sec)
Pentium 4 1.7Ghz     pgCC -O3 -fast -Minline=3/4   8.27  
Pentium 4 1.7Ghz     gnu g++ -O3 (2.95.2)          8.62  <-- fastest
Pentium 4 1.7Ghz     Intel icpc 5.0 -O3            8.65  
Pentium 4 1.7Ghz     gnu g++ -O3 (2.95.3)         10.94   
Pentium 4 1.7Ghz     gnu g++ -O3 (3.2)            11.30  <-- new
Pentium 4 1.7Ghz     gnu g++ -O3 (3.01)           15.09  <-- previous

The improvement from 3.01 to 3.2 are significant but it their a way of 
somehow getting the performance of 2.95.2 back?

Anyway, great work! Thanks a lot,

Burkhard Militzer

For comparison:
      						    
DEC Alpha 1GHz       KCC kcc3.4d -O3               5.2   
DEC Alpha 667MHz     KCC kcc3.4g2 -O3              8.4   
DEC Alpha 667MHz     g++ -O3 (2.96?)              10.3   
DEC Al-Linux 667Mhz  g++ -O3 (2.96?)              14.66  
O2K R10000           KCC -O3 +K3 --no_exceptions  18.26
IMB SP2 332MHz       KCC kcc3.4d -O3 +K3          27.56  

Re: Do we get the other 15% performance back too?

[Andreas Jaeger]

Burkhard Militzer <militzer@llnl.gov> writes:

> Hello,
>
> As a fan of the C++ gnu compiler project, I would like to congratulate the 
> gurus on the recent performance increase probably coming from corrections 
> in the prefetch routines. My benchmark shows

-O3 does not use prefetching it all.

Can you destill some small testcase that shows the problems?

Andreas

> Processor/OS         Compiler/options             Time(sec)
> Pentium 4 1.7Ghz     pgCC -O3 -fast -Minline=3/4   8.27  
> Pentium 4 1.7Ghz     gnu g++ -O3 (2.95.2)          8.62  <-- fastest
> Pentium 4 1.7Ghz     Intel icpc 5.0 -O3            8.65  
> Pentium 4 1.7Ghz     gnu g++ -O3 (2.95.3)         10.94   
> Pentium 4 1.7Ghz     gnu g++ -O3 (3.2)            11.30  <-- new
> Pentium 4 1.7Ghz     gnu g++ -O3 (3.01)           15.09  <-- previous
>
> The improvement from 3.01 to 3.2 are significant but it their a way of 
> somehow getting the performance of 2.95.2 back?
>
> Anyway, great work! Thanks a lot,
>
> Burkhard Militzer
>
> For comparison:
>       						    
> DEC Alpha 1GHz       KCC kcc3.4d -O3               5.2   
> DEC Alpha 667MHz     KCC kcc3.4g2 -O3              8.4   
> DEC Alpha 667MHz     g++ -O3 (2.96?)              10.3   
> DEC Al-Linux 667Mhz  g++ -O3 (2.96?)              14.66  
> O2K R10000           KCC -O3 +K3 --no_exceptions  18.26
> IMB SP2 332MHz       KCC kcc3.4d -O3 +K3          27.56  
>
>

-- 
 Andreas Jaeger
  SuSE Labs aj@suse.de
   private aj@arthur.inka.de
    http://www.suse.de/~aj

Re: Do we get the other 15% performance back too?

[Joe Buck]

> As a fan of the C++ gnu compiler project, I would like to congratulate the 
> gurus on the recent performance increase probably coming from corrections 
> in the prefetch routines. My benchmark shows
> 
> Processor/OS         Compiler/options             Time(sec)
> Pentium 4 1.7Ghz     pgCC -O3 -fast -Minline=3/4   8.27  
> Pentium 4 1.7Ghz     gnu g++ -O3 (2.95.2)          8.62  <-- fastest
> Pentium 4 1.7Ghz     Intel icpc 5.0 -O3            8.65  
> Pentium 4 1.7Ghz     gnu g++ -O3 (2.95.3)         10.94   
> Pentium 4 1.7Ghz     gnu g++ -O3 (3.2)            11.30  <-- new
> Pentium 4 1.7Ghz     gnu g++ -O3 (3.01)           15.09  <-- previous

You don't explain just what you are timing here.

I don't understand why you are getting such different results for 2.95.2
and 2.95.3, considering that there aren't really any differences to speak
of; the only job of the 2.95.2 -> 2.95.3 patch was to fix some breakage
having to do with compatibility with glibc 2.2, as well as to fix a couple
of minor bugs in a way that should not have had any effect on compiler
performance.

How did you produce these compilers?  Did you build them yourself, or
download them from somewhere?  If they are from .rpm's or .deb's, were
they built for 686 or 386?