Closed solomonik closed 8 years ago
Try configuring with FC=ifort.
I tried rebuilding with FC=ftn, but there is still -lgomp in FCLIBS in the resulting Makefile. It would be nice to ensure -lgomp does not get into FCLIBS when compiling with Intel CXX. It results in a runtime memory corruption when running with threads, which is a frustrating bug to encounter/debug since it looks like a logical error in some threaded code region in CTF or AQ. I've spent a bunch of time tracking it down twice now, did the commit that fixed it last time (if there was one, I am not 100% sure) get undone when you reverted recently?
I'll check the commit logs, but could you also send the config.log file?
On 7/4/16 1:50 AM, Edgar Solomonik wrote:
I tried rebuilding with FC=ftn, but there is still -lgomp in FCLIBS in the resulting Makefile. It would be nice to ensure -lgomp does not get into FCLIBS when compiling with Intel CXX. It results in a runtime memory corruption when running with threads, which is a frustrating bug to encounter/debug since it looks like a logical error in some threaded code region in CTF or AQ. I've spent a bunch of time tracking it down twice now, did the commit that fixed it last time (if there was one, I am not 100% sure) get undone when you reverted recently?
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/devinamatthews/aquarius/issues/14#issuecomment-230217545, or mute the thread https://github.com/notifications/unsubscribe/AFAMoZBGr6DUaIZC4gIlnQ-1eavUE_WQks5qSK1LgaJpZM4JDCMj.
When configuring on Edison with an Intel environment, AQ includes the -lgomp link flag, yielding to both Intel and GNU OpenMP symbols being linked in, and causing errors when running with threads. We dealt with a similar problem before, I don't remember how/whether it was resolved.