News for hackers. Stuff that matters.

A comment on The End of Numeric Error: An interview with John L. Gustafson. Parent comment

Re:The End of Numeric Error: An interview with John L. Gusta... by eli_gottlieb on Tuesday April 26, 2016 @02:37PM (#11574440)
I dunno. I just started doing probabilistic programming in Haskell yesterday, and it's turned out to be faster at it than every other PPL I've tried. It didn't help that WebChurch, Anglican, and Venture are all interpreted languages. Treating a sampler as a lazy stream of samples is a nice abstraction, too.
[ Parent | on HN ]
Re:The End of Numeric Error: An interview with John L. Gusta... by neutronicus on Tuesday April 26, 2016 @02:39PM (#11574458)
As a counterpoint, my attempts to write fast Haskell PDE solvers have been underwhelming.
[ Parent | on HN ]
Re:The End of Numeric Error: An interview with John L. Gusta... by jfoutz on Tuesday April 26, 2016 @03:53PM (#11575043)
Once you wrap your head around the Haskell way, it seems easy to get something working. IMHO that something will be minimal and pretty as well. Once i try to make things fast, it gets weird. There are people who can write very fast Haskell code, but it seems like all of those people worked on ghc. I don't have a good mental model of when, say, a thunk will get created, so i liberally sprinkle strictness annotations, and fool around with list fusion operations.

I really like Haskell. Even with all the effort i've put into it, it seems like i've barely scratched the surface if i want performance. sigh

[ Parent | on HN ]
Re:The End of Numeric Error: An interview with John L. Gusta... by marmaduke on Tuesday April 26, 2016 @03:52PM (#11575040)
Did you try Stan? It compiles your model to SSE2 enabled c++
[ Parent | on HN ]
Re:The End of Numeric Error: An interview with John L. Gusta... by eli_gottlieb on Tuesday April 26, 2016 @07:27PM (#11576604)
For my current problem, I don't want the available inference methods to impose any restrictions on what sort of model I can write, so I'm using general PPL only.
[ Parent | on HN ]
Re:The End of Numeric Error: An interview with John L. Gusta... by marmaduke on Wednesday April 27, 2016 @03:49AM (#11578534)
You can write you model in Stan and sample without inference too.
[ Parent | on HN ]