build_libsmm generate small1

Alfio Lazzaro alaz... at cray.com
Wed Sep 18 21:35:53 CEST 2013


Hi Joe,
sorry I was busy with a workshop this week, don't have much time to reply 
to you.
First of all, it is a good idea to switch everything to Intel compiler. I 
found that it gives better performance than GNU when compiling libsmm.
Then, I understood correctly that with this new setup you are able to 
execute small1 up to

/output_linux.intel/small_find_9_32_32.x > 
output_linux.intel/small_find_9_32_32.out

? If so, than it means that my guess that the the problem was in MKL 
compatibility with GNU was correct.
Definitely now you are facing a different problem.

I downloaded your file and compiled with:

ifort -O2 -funroll-loops -vec-report2 -warn -xHost -fno-inline-functions 
-nogen-interfaces -openmp -nolib-inline -static-intel 
small_find_9_32_32.f90 -o small_find_9_32_32.x -mkl=sequential

and then I did:

export OMP_NUM_THREADS=1; ./small_find_9_32_32.x

It works for me (ifort version 13.1.0)...

Now, I can suggest to try to use:

ulimit -s unlimited

If it still doesn't work, then I can also suggest to try to add this flag 
during compilation: -heap-arrays 64
The problem is that Intel compiler puts temporaries on the stack, which 
might lead to segmentation faults is the stack limit is set to low. 

Alfio




Il giorno sabato 14 settembre 2013 09:33:45 UTC+2, Joe Greenstone ha 
scritto:
>
>  Hi Alfio,
>  
> Attached is small_find_9_32_32.f90 so you could compile/link and run it 
> locally.
> If amongst other variations you were to compile/link with the same options 
> I mentioned in my previous reply, it may go some way toward establishing 
> that there is a difference between different levels of the Intel compiler 
> that matter in this case.
>  
> Thanks again. 
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20130918/bc817c51/attachment.html>


More information about the CP2K-user mailing list