On Wed, Mar 16, 2011 at 11:06 PM, nathan binkert <n...@binkert.org> wrote:
> > Does anyone actually use it? > In python, probably not. I used it in EXTRAS at one point though for sure. > > > Would it really matter if we had two different > > RNGs (one python and once C++) as long as we can seed both of them > > reproducibly from a single input value? > I don't know that the argument is overwhelming. The biggest concern is > that we have a good RNG. Bad RNGs are really bad to use. The biggest > problem is that we're already using several different RNGs. Ruby and > the testers don't use M5's RNG, but rather random() > > My reasoning for wanting to use only the RNG in src/base is that it's > a proven, *very* good RNG and it would be the single canonical source > for randomness. > The Python RNG is Mersenne Twister just like the one in src/base ( http://docs.python.org/library/random.html). My issue right now is that the current random.i prevents you from using the official python random module, which is just plain broken. I don't understand why it is broken; you'd think that having it in the m5.internal package would prevent them from interacting, but it doesn't. It could be fixed by renaming or (better) figuring out why the package isolation isn't working, but if all we have are hypotheses about possible scenarios where it might be useful in the future, and a recognition that even if the interference problem is fixed it might not be adequate, then I think we should just get rid of it now, and when someone actually has a need then we can put in something that addresses the actual problem at hand. Steve _______________________________________________ m5-dev mailing list m5-dev@m5sim.org http://m5sim.org/mailman/listinfo/m5-dev
