Function suggestion: withFrozen

[Akhra Mellivora Gannon]

The pure System.Random interface is nice, but still considerably slower than working within the pcg-random idiom. Chasing an alternative, I've come up with this:

withFrozen :: (Variate a)
           => (forall s. Gen s -> ST s a) -> FrozenGen -> (a, FrozenGen)
withFrozen f g = runST $ flip runStateT g $ do
  gen <- get >>= \frozen -> restore frozen
  x <- lift $ f gen
  save gen >>= \frozen -> put frozen
  return x

By way of example, I've also got a function

roll :: (PrimMonad m) => Int -> Gen (PrimState m) -> m Int

and it works with either withFrozen (roll n) g or withSystemRandom $ roll n.

I'm sketchy on the order of arguments -- I also had the idea of working this into a StateT variant, but putting the function last would be more convenient in a lot of cases -- and make no guarantees as to my code being anywhere close to optimal, but it does work and it's MUCH faster than the System.Random interface.

[Christopher Chalmers]

Yeah, the System.Random interface is unfortunately very slow because of how the class is defined. There's a GSOC project that will hopefully fix this.

There's no need for StateT here, it's simply:

withFrozen :: (forall s. Gen s -> ST s a) -> FrozenGen -> (a, FrozenGen)
withFrozen f s = runST $ do
  g  <- restore s
  a  <- f g
  s' <- save g
  return (a, s')

or even

withFrozen f s = runST $ restore s >>= \g -> liftA2 (,) (f g) (save g)

which I'll add.

I've actually started a pure branch here which performs faster than using ST. I couldn't figure out a nice interface and since I only use the ST version I've just left it until the GSOC project comes.

[Christopher Chalmers]

Added in aa5b482. I decided to flip the arguments.

[Akhra Mellivora Gannon]

My predictions of suboptimality have yet to come up short. ;) Ed Kmett brought up GSOC yesterday on #haskell-game too, sounds like there's real momentum there.