How to improve performance of this numerical computation in Haskell?

Question

I\'m in the middle of porting David Blei\'s original C implementation of Latent Dirichlet Allocation to Haskell, and I\'m trying to decide whether to leave some of the low-l

Answer 1

Before the optimization work, I wouldn't say that your original translation is the most idiomatic way to express in Haskell what the C code is doing.

How would the optimization process have proceeded if we started with the following instead:

trigamma :: Double -> Double
trigamma x = foldl' (+) p' . map invSq . take 6 . iterate (+ 1) $ x
where
  invSq y = 1 / (y * y)
  x' = x + 6
  p  = invSq x'
  p' =(((((0.075757575757576*p-0.033333333333333)*p+0.0238095238095238)
              *p-0.033333333333333)*p+0.166666666666667)*p+1)/x'+0.5*p