[Openmp-dev] omp_is_initial_device commit small issue

Peyton, Jonathan L jonathan.l.peyton at intel.com
Thu Feb 12 14:59:53 PST 2015


Are you getting error messages from the linker?

When I compile both the offload and runtime libraries then run a little offloading program with all four interface functions, I do not get conflicting symbols at link or load time.  The omp_*_default_device(), omp_get_num_devices() functions are only defined for MIC and Mac inside libiomp5.  For Linux, they are only defined in liboffload.  And omp_is_initial_device() is only defined in libiomp5 for all platforms.  It is a little confusing...

There is an issue with Windows buildsL: omp_[s|g]et_default_device() and omp_get_num_devices() aren't defined anywhere (liboffload.so.5 is only buildable on linux).

-- Johnny

-----Original Message-----
From: openmp-dev-bounces at cs.uiuc.edu [mailto:openmp-dev-bounces at cs.uiuc.edu] On Behalf Of C Bergström
Sent: Wednesday, February 11, 2015 5:22 AM
To: openmp-dev at dcs-maillist2.engr.illinois.edu
Subject: [Openmp-dev] omp_is_initial_device commit small issue

With the commit "OpenMP 4.0 standard function omp_is_initial_device() implemented." we now have conflicting symbols compared to our own device offloading support

This will probably be true for the following:
omp_set_default_device
omp_get_default_device
omp_get_num_devices
omp_is_initial_device
--------
What's the best way to tackle this? Should we rename stuff on our side
(_Offload_get_device_number) or #if or what's the best design?

Thanks
_______________________________________________
Openmp-dev mailing list
Openmp-dev at dcs-maillist2.engr.illinois.edu
http://lists.cs.uiuc.edu/mailman/listinfo/openmp-dev




More information about the Openmp-dev mailing list