Showing that random-fu is x4 faster than mwc and thus monad-bayes

[idontgetoutmuch]

ghc samplePerformance.hs 
[1 of 1] Compiling Main             ( samplePerformance.hs, samplePerformance.o )
Linking samplePerformance ...

./samplePerformance 
benchmarking dists/StdGen/stdNormal/single sample
time                 45.20 ns   (45.16 ns .. 45.25 ns)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 45.28 ns   (45.22 ns .. 45.41 ns)
std dev              304.5 ps   (173.5 ps .. 545.5 ps)

benchmarking dists/StdGen/stdNormal/single sample MWC
time                 157.0 ns   (156.9 ns .. 157.2 ns)
                     1.000 R²   (1.000 R² .. 1.000 R²)
mean                 157.0 ns   (156.9 ns .. 157.3 ns)
std dev              600.7 ps   (338.0 ps .. 990.2 ps)

benchmarking dists/StdGen/stdNormal/single sample monad bayes
time                 157.2 ns   (156.9 ns .. 157.6 ns)
                     0.999 R²   (0.998 R² .. 1.000 R²)
mean                 158.5 ns   (157.1 ns .. 163.6 ns)
std dev              8.239 ns   (1.448 ns .. 17.33 ns)
variance introduced by outliers: 71% (severely inflated)

[idontgetoutmuch]

We may be comparing apples and oranges here and one thing I plan to try is replacing mwc-random by random-fu.

[idontgetoutmuch]

I'd also like to benchmark sampling in Haskell against other languages. But is a PR here the best place to do these experiments?

[turion]

I'd also like to benchmark sampling in Haskell against other languages. But is a PR here the best place to do these experiments?

Yes, I guess it is, but in a separate PR (and a separate benchmarking suite).

[turion]

Is that change intentional?

[turion]

I guess this was also not intentional?

[idontgetoutmuch]

We may be comparing apples and oranges here and one thing I plan to try is replacing mwc-random by random-fu.

@turion I just tried

newtype SamplerU g m a = SamplerU {runSamplerU :: ReaderT g m a} deriving (Functor, Applicative, Monad, MonadIO)

instance (StatefulGen g m, MonadReader g m) => MonadDistribution (SamplerU g m) where
  random = SamplerU undefined

  uniform a b = SamplerU undefined
  normal m s = undefined -- SamplerU (ReaderT $ (RF.sample (RF.normal m s)))
  gamma shape scale = SamplerU undefined
  beta a b = SamplerU undefined

  bernoulli p = undefined
  categorical ps = undefined
  geometric p = undefined

but have type errors (if I uncomment the definition for normal). I have to give up for the weekend as we have visitors but you had some time, maybe you could take a butcher's.

EDIT: But this seems to work (at least it type checks):

  normal m s = SamplerU (ReaderT $ RF.runRVar $ RF.normal m s)

[idontgetoutmuch]

Replaced by #323