Issue with random number generator.

[mcpherson]

Posting this issue on behalf of Guillaume:

I looked at the random generator used in CoMD with great interest since I did the exact same thing (in F90) for one of our code.

In my code, I bumped into the problem that with very large simulations, my seed could wrapup through 0. And if seed == 0 then my rng and lcg61 will stay at 0 forever. And I wasn't exiting gasdev :-(

It means also that the actual interval for lcg61 is ]0., 1.].

A guard against that situation in mkSeed would be useful. And I wonder if a protection against zero would not be a good idea too in lcg61.

As an improvement I would suggest to move lcg61 in the header as "inline" so that it can be vectorized easily if used outside gasdev (at least on SNB, something's wrong on KNC).

[HydroBench]

Running lcg61 for a (very) long time with a seed != 0 doesn't seem to wrap up via 0. Could it be possible to formally demonstrate that if seek != then the serie never goes through 0 ? I'll probe around me.
Guillaume Colin de Verdière