Re: Amazing Performance Difference between vec[] and iterators

[Bernd Strieder]

Hello,

Paulo Matos wrote:

> I got these times:
> $ time ./vec
> 5030 elements in array.
> sum: 27665
> 
> real    0m7.123s
> user    0m7.108s
> sys     0m0.008s
> 
> $ time ./vec-iter
> 5030 elements in array.
> sum: 27665
> 
> real    0m4.699s
> user    0m4.672s
> sys     0m0.020s
> 
> Any comments on these results? They don't seem intuitive to me. Why
> the difference?

> LANG=C g++ -v
Using built-in specs.
Target: i686-pc-linux-gnu
Configured with: ../gcc-4.1.1/configure --prefix=/net/opt/gcc-4.1.1
--enable-languages=c,c++ --with-system-zlib --enable-shared
--enable-__cxa_atexit --enable-threads=posix
Thread model: posix
gcc version 4.1.1

> g++ -o vec2 -O3 vec2.cc
> g++ -DITER -o veciter2 -O3 vec2.cc
> time ./vec2
5030 elements in array.
sum: 27665

real    0m5.666s
user    0m5.604s
sys     0m0.033s
> time ./veciter2
5030 elements in array.
sum: 27665

real    0m6.588s
user    0m6.537s
sys     0m0.037s
> g++ -o vec2 -O3 -funroll-loops vec2.cc
> g++ -DITER -o veciter2 -O3 -funroll-loops vec2.cc
> time ./vec2
5030 elements in array.
sum: 27665

real    0m3.209s
user    0m3.162s
sys     0m0.017s
> time ./veciter2
5030 elements in array.
sum: 27665

real    0m2.846s
user    0m2.818s
sys     0m0.015s

I can't reproduce the extreme difference, but adding -funroll-loops to
-O3 changes the picture for me, the iterator version gets the fastest,
and both remarkably better. I think the optimization in 4.1.1 is far
from finished, and will never be because of 4.2. In older releases the
iterator version seems consistenly worse.

Bernd Strieder