|
|
Blitz Support : |
From: Xavier WARIN(Compte LOCAL) - I23 (xavier.warin_at_[hidden])
Date: 2004-08-19 04:37:50
Hi Xianglong
I use a P4 Xeon 2.6 Ghz with 2 GB memory (i don't know the cache)
I come from the fortran world where compilers were very effective for
optimizing. A friend of mine did some test one year ago with gcc 2.95.3
which was not able to optimize correctly the loops.
One way to optimize the code was to do what i proposed and to swap loops.
The most effective was to use ATLAS library (by far).
Xavier
Xianglong Yuan wrote:
> Hi Xavier,
>
> Xavier wrote:
>
>
>>plain array : 240000
>>plain array 2 : 70000 --> my version
>>blitz array : 780000
>>blitz array 2 : 70000 --> my version
>
>
> I believe these numbers are not real, due to the fact that 'plain
> array 2' repeats the first run. By simply swapping the positions of
> these two loops, 'plain array' runs faster than 'plain array 2'. See
> the results below.
>
> before swapping positions of the two loops:
>
>
>>plain array : 140000
>>plain array 2 : 110000
>
>
> after swapping:
>
>
>>plain array 2 : 140000
>>plain array : 110000
>
>
> (gcc3.3.1 with -O3 -funroll-loops -fstrict-aliasing)
>
> I don't know why the difference is so small in my case. Maybe it
> relates with cache and/or memory. More than 3 times difference in
> your case is very impressive. What kind of hardware do you use? Mine
> is P4 3.06 GHz (512K cache?) with 1GB memory.
>
> In my blitz_test.cpp, there are some multiplications in the inner
> loops that can be moved out. But I don't think we need to worry about
> them. Modern compilers are pretty smart in optimizing loops. I've
> tried before with other codes to get all the calculations out of inner
> loops, but the improvement are trivial. You may get significant
> improvement by doing these under the -O0 flag, but achieve little
> under -O3 flag. The compiler can do all these common optimizations
> and does better. Optimizing codes at higher levels, nonetheless, is
> indeed very important.
>
>
> Xianglong
>
>
> _______________________________________________
> Blitz-support mailing list
> Blitz-support_at_[hidden]
> http://www.oonumerics.org/mailman/listinfo.cgi/blitz-support
>
>