listed on r-bloggers.com

Mittwoch, 13. April 2011

compiler and runiregGibbs (bayesm)

So everyone's excited about the new R 2.13 release because of the compiler package.
Apparently it is easy to get a 3x speed increase by simply compiling a function.
Doing a lot of the MCMC stuff, I am particularly excited about speed in R. I just compiled a 2000-line file with code from my latest project, but none of the functions would run faster. Apparently I need to break down things a little more and use more subfunctions.

Well, so I tried a much easier example. I took runiregGibbs from the well known bayesm packages (which is a function completely written in R) and compiled it. There's a visible change, but it's quite small:

library(compiler)
library(bayesm)
#compile runiregGibbs function
runiregGibbsc=cmpfun(runiregGibbs)
set.seed(66); n=1000; R=10000
X=cbind(rep(1,n),runif(n)); beta=c(1,2); sigsq=.25
y=X%*%beta+rnorm(n,sd=sqrt(sigsq))
Data1=list(y=y,X=X); Mcmc1=list(R=R)
#run code
t.original = system.time(runiregGibbs(Data=Data1,Mcmc=Mcmc1))
t.compiled = system.time(runiregGibbsc(Data=Data1,Mcmc=Mcmc1))
#result
> t.original
User System verstrichen
2.20 0.11 2.30
> t.compiled
User System verstrichen
1.90 0.20 2.11
view raw gistfile1.r hosted with ❤ by GitHub