even faster

[plinius]

Hi,
I made a very calculation-intensive program, which I want to optimize using the intel c++ 8.1 compiler. The code is allready as optimal as I can make it...

I used the following settings: (will only run on P IV for the moment...)

O3
QaxN
Qipo_S
Qsfalign

Are there any other settings I should do ?
(supp info: target platform win XP/2003
P IV's
w32-cmdline program )

And, PROFORDER, how does this works ?
(the help:

Usage: proforder
[-nologo]
[-omit_static]
[-prof_dir directory]
[-prof_file dpi_spi_file_basename]
[-prof_dpi dpi_file]
[-prof_spi spi_file]
[-i ipo_dump_file]
[-m method]
[-n number of modules to output to IPO selection (0 is default)
Only supported with -m 1]
[-o output_file]
[-ipo ipo_output_file]
didn't helped me a lot...)

can someone please give a easy to understand example ??


Thanks a lot.

Message Edited by plinius on 03-09-2005 05:22 PM

[Intel_C_Intel]

Dear Plinius,

The switch combination

/QxN /O3 /Qipo /Qprec-div-

seems a good first attempt in getting high performance. No need to use auto-cpu-dispatch (viz. /QaxN instead of /QxN) if you only plan to run the generated binary on a Pentium 4 processor. The switch /Qprec-div- may help if your application performs a lot of FP divisions and is numerically stable enough to allow additional optimizations on them. In some cases, profile-guided optimizations may help, where the compiler learns a few important properties of your code by applying it to a representative data set. A typical command line session for this would be:

(1) icl /Qprof_gen program.c
(2) program.exe < train_dataset.txt
(3) icl /Qprof_use /QxN /O3 /Qipo /Qprec-div- program.c

Which yields an executable that is optimized according to characteristics observed while running the training data set.
I hope this gives you a few initial ideas on how to start.

Aart Bik
http://www.aartbik.com/

[plinius]

OK.

Thanks a lot.