Fix Intel10_64lp STATIC case

[massich]

No description provided.

[massich]

Here's the output of mkl_link_tool for Intel10_64lp ( with --parallel=yes) and Intel10_64lp_seq (--parallel=no). They differ in the fact that parallel adds -liomp5 to be linked dynamically.

❯ ${MKLROOT}/tools/mkl_link_tool -libs --parallel=yes --compiler=gnu_c --os=lnx --arch=intel64 --linking=static --interface=lp64

       Intel(R) Math Kernel Library (Intel(R) MKL) Link Tool v4.5
       ==========================================================

Output
======

Linking line:
 -Wl,--start-group $(MKLROOT)/lib/intel64/libmkl_intel_lp64.a $(MKLROOT)/lib/intel64/libmkl_intel_thread.a $(MKLROOT)/lib/intel64/libmkl_core.a -Wl,--end-group -liomp5 -lpthread -lm -ldl

~/Documents/code/random_thoughts/cpp-mkl-test/build_dynamic_noseq/CMakeFiles/CMakeTmp Intel10_64lp_static*
❯ ${MKLROOT}/tools/mkl_link_tool -libs --parallel=no --compiler=gnu_c --os=lnx --arch=intel64 --linking=static --interface=lp64

       Intel(R) Math Kernel Library (Intel(R) MKL) Link Tool v4.5
       ==========================================================

Output
======

Linking line:
 -Wl,--start-group $(MKLROOT)/lib/intel64/libmkl_intel_lp64.a $(MKLROOT)/lib/intel64/libmkl_sequential.a $(MKLROOT)/lib/intel64/libmkl_core.a -Wl,--end-group -lpthread -lm -ldl

For the parallel case, my local system generates this:
CMAKE_REQUIRED_LIBRARIES : ${MKLROOT}/lib/intel64_lin/libmkl_intel_lp64.a;${MKLROOT}/lib/intel64_lin/libmkl_intel_thread.a;${MKLROOT}/lib/intel64_lin/libmkl_core.a;/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/intel64_lin/libiomp5.a;-lpthread;-lm -ldl. Notice that libiomp is set to link statically and that is not the same as the one proposed by mkl_link_tool.
Even more, whatever CMake is picking up is not there. It has found the .so in /opt/intel/clk/... and change its extesnsion to .a. But is not there.

❯ locate libiomp5                           
/opt/intel/clck/2018.0.002/provider/share/common/lib/intel64/libiomp5.so
/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/ia32_lin/libiomp5.a
/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/ia32_lin/libiomp5.dbg
/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/ia32_lin/libiomp5.so
/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/intel64_lin/libiomp5.a
/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/intel64_lin/libiomp5.dbg
/opt/intel/compilers_and_libraries_2018.0.128/linux/compiler/lib/intel64_lin/libiomp5.so

Other things that I don't understand are:
1 - why guide appears here if it is only set up for old versions (see here).
If that where not the case, guide should also appear here but is not the case.

2 - why both cmake test compilations pass (here for 64lp_seq and here for 64lp). they replicate on my system. Both compile with no problem. However 64lp case returns FALSE wheras 64lp_seq retruns TRUE.

3 - the cmake failure log shows that its looking for -lpthreads (ended with s instead of `lpthread) (see here). What I don't understand is why this appears on the log but not during the cmake run. This issue with cmake-3.7 + ubuntu-14 might bring something to the discussion

[massich]

ping @agramfort

[massich]

Based on this commend, I think we should do find_package(Threads) in CMakelists.txt and break execution is this is not done when doing FindBLA for Intel case.

But still, this only means that all those errors were normal. Even if we were ignoring them and hardcoding -lpthread which we should stop doing ! So not what we are looking for.
But then it makes no sense, 'cos the try_compile was already propely generated for sgemm_

[agramfort]

you think doing : set(THREADS_PREFER_PTHREAD_FLAG On) should help?

[massich]

It does not. I think it should be related shomehow (see https://gitlab.kitware.com/cmake/cmake/issues/16221)