My next suggestion requires you to update your AFNI binaries, sometime on Friday (12 Feb 2016).
I added two options to 3dGroupInCorr:
-read makes the program read the data into memory, rather than memory map it
-ztest makes the program test the data (no matter how it was input) to see if it is all zero; this happens before the computations start, and a message is printed like
++ Running -ztest now
+ total of 0 datasets tested as all zero
So update your copy of AFNI, then run 3dGroupInCorr with the '-ztest' option to see if it detects any problems with the data.