[Rd] inline C/C++ in R: question and suggestion

[Simon Urbanek]

Oleg,

On May 22, 2007, at 1:59 PM, Oleg Sklyar wrote:

> We've been discussing in the group that it would be nice to have a  
> mechanism for something like "inline" C/C++ function calls in R. I  
> do not want to reinvent the wheel, therefore, if something like  
> that already exists, please give me a hint -- I could not find  
> anything. If not, here is a working solution, please criticise so I  
> could improve it.
>
> Example: I work on images (Bioconductor:EBImage) and just came to a  
> point where I need to apply certain functions to image data, which  
> are grey scale intensities in the range [0,1]. Assume I want to  
> transform my image data from I(x,y,i) to exp(-(d/s)^2)*I(x,y,i),  
> where I is the original intensity in dependence on coordinates x,y  
> and frame i; s is a given value and d^2=(x-centre.x)^2+(y-centre.y) 
> ^2 for a given centre. Trying an R loop will run forever already on  
> moderate image sizes as I do not see how to vectorize it.
>

That is actually a (rare) case that can be completely vectorized:

d=(cx-rep(1:dim(I)[1],dim(I)[2]*dim(I)[3]))^2+(cy-rep(1:dim(I) 
[2],each=dim(I)[1],times=dim(I)[3]))^2
I=I*exp(-(d/s^2))

Clearly the drawback is the use of memory, but you could vectorize  
per frame if you wish. At any rate it's not that slow anyway:

 > I=array(runif(100*100*10),c(100,100,10))
 > system.time({d=(cx-rep(1:dim(I)[1],dim(I)[2]*dim(I)[3]))^2+(cy-rep 
(1:dim(I)[2],each=dim(I)[1],times=dim(I)[3]))^2;I=I*exp(-(d/s^2))})
    user  system elapsed
   0.022   0.010   0.032
 > system.time(funx(I,15,c(30,35)))
    user  system elapsed
   0.008   0.001   0.010

Of course C wins, no doubt about that :).


> Now, below is the solution using the "inline" C code, completely in  
> R and runs instantly. I created a small package "inline" that  
> simply encapsulates a quite simple function "cfunction". The  
> package source is available from http://www.ebi.ac.uk/~osklyar/ 
> inline -- please give it a try and I would be happy to hear your  
> comments, both on already existing implementations for "inline"  
> calls and on the current one.

I really like the idea! Except for the fact that it's forcing the use  
of C++ which adds unnecessary overhead :P I'd like a configurable  
extension [including .m] and the ability to prepend functions as  
code. What would be very useful now is a set of tools that would  
allow you to construct the source with R commands, so that you could  
compute on it, edit it etc. That would be really cool ... you could  
even imagine compiling a very restricted set of R into C code ...  
yes, I know I'm dreaming ;)

Thanks for the good idea!

Cheers,
Simon