Compilation problems

Starting in 2005, there was an upsurge of reports on remlf90, airemlf90, gibbf90 etc giving incorrect answers or crashing. Typical messages could be:

"Matrix not semipositive-definite"

"DATAFILE not found".

Otherwise, one finds that estimates from reml and gibbs programs are different on different computers or with different compilers. These problems seem to be related to compiler bugs. F95 compilers are not equal, and with more challenging codes, some of them fail.

Below is an edited message from Jarmila Bohmanova on her attempt to get one analysis going…

I tested it on both Opteron and Xenon at UGA compiled with Absoft compilers. Everything looked fine and I got identical results after 10 rounds.

Then I tried to run it on one of CGIL Opetrons. It crashed after 5 rounds because of  "Matrix not semipositive-definite".  So, I downloaded binaries from UGA, run it and everything was OK.

I thought I have probably different version of gibbs2. So I downloaded both libraries and source code for gibbs2 from UGA Opteron. I compiled it with g95, run it but it crashed again.

Then I compiled it with Pathf90, run it and it crashed at the very beginning "DATAFILE not found".

I reformatted the parameter file to unix format 'dos2unix parfile', run it again, and voile, it worked just like on UGA Opetron (of course slightly faster, it is a brand-new machine).
I was not able to 'fix' g95 binaries, do not know what is going on there.