ICC performance gain
 

Hello, just out of interest i compared msieve performance when compiled with ICC and gcc 4.3.

Hardware :
model name : Intel(R) Core(TM)2 Duo CPU E8500 @ 3.16GHz
ram : DD2 8500

Software:
gcc (Debian 4.3.2-1) 4.3.2
icc (ICC) 11.0 20080930

Flags :
gcc : -O3 -fomit-frame-pointer -march=core2 -D_FILE_OFFSET_BITS=64
icc : -fast -fp-model precise -D_FILE_OFFSET_BITS=64

GMP was used, compiled with gcc (I could'nt get it to compile with ICC before giving up)


Number to be factored (random 100 digits number generated with openssl):
1802716097522165018257858828415111497060066282677325501816640492782221110851604465066510547671104729

I tried to be as fair as possible and copied the factor base from the first run (gcc) for the icc run and fixed the random number in the source.

Results :
GCC :
[CODE]Mon Nov 17 03:29:59 2008 prp50 factor: 38589340584901213653958931179714585367490014795273
Mon Nov 17 03:29:59 2008 prp50 factor: 46715390058453362424711065024780497160328505582673
Mon Nov 17 03:29:59 2008 elapsed time 31:46:11[/CODE]

ICC:
[CODE]
Tue Nov 18 13:39:01 2008 prp50 factor: 38589340584901213653958931179714585367490014795273
Tue Nov 18 13:39:01 2008 prp50 factor: 46715390058453362424711065024780497160328505582673
Tue Nov 18 13:39:01 2008 elapsed time 27:00:43[/CODE]

Full logs are attached

Thanks, that's good to know. Brian Gladman has kindly performed a complete overhaul of the preprocessor defines for compiler-specific inline asm, so that the next version will allow icc to reuse all the gcc-specific assembly language. I suspect that it currently does not.

Was this a 64-bit or 32-bit binary?

64 bits.
Do you mean msieve will use gcc intrinsics ?

I have seen with other programs that ICC register allocator and optimisations are dramatically more efficient than gcc (for SSE2)

I meant using gcc-format inline asm statements, which according to Brian can be parsed by icc. He also reports favorable speedups on windows.

maybe it would be worth trying intrinsics, i'll try to check if I have time

Actually, Intel's compiler could probably benefit from a prefetch intrinsic. The code has gcc and MSVC versions but no icc version in include/util.h