openmp

[jboulanger]

Hi,

I had hard time to cross compile code using openmp.

I found discussion (2009) about the support of openmp on the mailing list but the current status was unclear.

By default the gcc.mk has openmp disabled (--disable-libgomp) and the default thread model is win32 (which i read is incompatible with gcc implementation of libgomp). I found using
--enable-threads=posix
--enable-libgomp \

• adding pthreads as a dependency working (although a cyclic dependency warning is triggered).

An enhancement would be to have the possibility to enable openmp as an global option... if this makes sense. At least document these steps.

Otherwise many thanks for this great project.

[tonytheodore]

Thanks for the feedback. If I recall correctly, libgomp was disabled and moved into it's own package as it never really worked correctly - pthreads-w32 has since been updated so it's worth revisiting. There are plans to allow for user-defined variants, but if this is now working I'd rather make it the default.

Do you know if the standalone libgomp package works and/or if there's a test suite that we could use to verify either that or enabling it in the gcc build?

[jboulanger]

Finally, the previous trick was not enough to use the x64_86-w64-mingw toolchain so i went for another try with

• compile gcc with default .mk (make gcc)
• compile libgomp (make liggomp)
• modifiy gcc.mk to use posix thread and enable libgomp
• compile gcc (make gcc)
• continue with other libs
• compile my code (./configure detects openmp as enabled 👍 )
• enjoy running the program with huge allocation and many threads

The src directory contains already a test from openmp website (src/libgomp-test.c)

[tonytheodore]

enjoy running the program with huge allocation and many threads

Good to hear.

I tidied up the standalone libgomp and found an easily accessible validation suite that is now included in MXE.

This isn't intended to be an absolute test, just a relative one to see the impact of different configurations. Testing with i686-w64-mingw32 on Win7, both the standalone libgomp and the method above return the same results:

Total tests: 218  Total Failed: 20 Total Passed: 198

I'll try an inline build with gcc and see if there's any difference.

[tonytheodore]

I had hard time to cross compile code using openmp.

Were you initially trying to build fortran code?

[jboulanger]

the code is in c++ (sorry for the delay)

[TimothyGu]

@jboulanger are you having any further problems with the existing libgomp infrastructure? There are multiple packages using OpenMP successfully in MXE already, like flann. If not, I'll close the issue.

[jboulanger]

No more problem with openmp and MXE. You can close the issue. Thanks a lot for the support!
Edit:
1/ You still need to modify the gcc.mk as openmp is disabled by default (line 396 gcc.mk). Applying the earlier (Feb 21) described step allow to use openmp (re tested today).
2/ disabling openmp will probably no induce a fail in a compilation benchmark as it can be safely ignored by gcc.
3/ Again this is not a bug but more about documentation i think.
4/ Sorry for commenting on a closed issue