Issues linking with sequential MKL on native MIC code.

Issues linking with sequential MKL on native MIC code.

I am trying to link a sequential version of MKL with my code for some testing. I am using ICC 13.0.1 (20121010). When I link it I get an undefined reference error:

icc -g -Wall -Wextra -O3 -mmic -o test_v5_pth main.o -L/opt/apps/intel13_1/mkl/11/lib/mic -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lpthread -lm
/opt/apps/intel13_1/mkl/11/lib/mic/libmkl_core.so: undefined reference to `omp_get_max_threads'

Obviously if I and -liomp5 I can get over the omp_get_max_threads, but I am wondering why does mkl_core reference omp_get_max_threads when I am trying to compile a sequential version of it. Is this a known issue or am I required to link with libiomp5, this is not what MKL link advisor gives me either.

8 帖子 / 0 全新
最新文章
如需更全面地了解编译器优化,请参阅优化注意事项

This is probably an issue for older versions of ICC and MKL. I tried with a recent version (ICC 13.1.0 20130121) and it didn't have this problem. Please consider upgrade your Composer XE. Thanks!

Thanks for the information, I will look into upgrading to the newer version of Intel Composer suite.

Building with the compiler option "-mkl" may also solve some of the problems.  It knows what libraries to include and does so automatically.

-mkl:sequential seems to be the matching option according to the mention at the top of this thread.  It would avoid any problem with wrong order or spelling typos of the library options.  Note in http://software.intel.com/en-us/articles/intel-mkl-link-line-advisor that you must use the  --start-group --end-group options when specifying individual static libraries on linux; that shouldn't have been needed when dynamic libraries are used.

Upgrading to the newer version of Composer XE fixed the problem.

Just to be clear in the older version of Composer XE 13.0.1 (20121010)

$ nm libmkl_core.so | grep omp_get_max
U omp_get_max_threads

In the newer version 13.1.1 (20130313)

$ nm libmkl_core.so | grep omp_get_max

Probably there was an issue with the MKL in our older version which got fixed with upgrade.

Yes, one would expect mkl_sequential to stub out get_threads calls from inside MKL and simply set 1 thread.

发表评论

登录添加评论。还不是成员?立即加入