OK, seems to be resolved now. Forgot to apply the patches for the ival=5 error. The ival=12 error is still mysterious, but may be linked either to the math library handling, or something on our cluster (the failures were occurring only on Xeon Phi nodes, and only on some of those nodes).