when comparing the Invel V7 compiler for Linux with GNU gcc2 and gcc3 we had some test cases where the Intel compiler was significantly slower than gcc. Even playing with the optimization options (-O3 -ipo -tpp7 -xW) did not close the performance gap *if* certain C++ features explained below where used. Example : FFT (attached)

gcc 2.95.3 : 30.8s
gcc 3.2.1 : 24,2s
icc 7 build 20021021Z : 45.8s

System : P4 1.7 GHz, 512 MB RAM

The performance gap increases, if templates are extensively used. Profiling showed that larger template classes are often not inlined, even if the maximum template size for inlining in increased via command line option.

Are there other possibilities to enhance performance of our code ? People where reporting amazing performance increases up to 30% we would like to share :).

Thank you in advance,

4 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.

>Example : FFT (attached)
I am not seeing the attachment.

Better still, could you please open this as a performance issue on the compiler support site at http://premier.intel.com?

Thanks for helping us give you better products!



>> Example : FFT (attached)
> I am not seeing the attachment.

Sorry. The test case should now be attached to this mail.

We are using the evaluation version of the compiler for performance tests and do not have access to premier.intel.com yet.


On a P4/2800, with gcc-3.3:
g++ -O2 -march=pentium4 -mfpmath=sse -funroll-loops -fbranch-probabilities performance.cpp

13.10 sec

With icc build 20021203
icc -xW -prof_use performance.cpp

11.56 sec.

Leave a Comment

Please sign in to add a comment. Not a member? Join today