[gmx-users] Double precision bug

[Andrew Horsfield]

Hi,

Does anyone know why the double precision version of mdrun can fail for 
conjugate gradient and steepest descent runs when I use PME, but not 
cut-off, with the error:

Fatal error: ci = -2147483648 should be in 0 .. 1099 [FILE nsgrid.c, LINE 214] 

The failure occurs at the first step. Using cut-off gives nonsense results
(the energy diverges to -infinity). The unit cell has a net charge of
-3,and contains 16118 atoms. Single precision does not seem to be adequate
to relax the cell.

I am using a PIII running Linux (RH7.1), and am using gcc to compile. I
have tried changing the compile options to reduce the level of
optimization, and used efence to trap any memory leaks, but without
success.

Cheers,

Andrew