[Rcpp-devel] RcppArmadillo

[Simon Zehnder]

Dear Rcpp-Devels,

this list was suggested to me by Dirk Eddelbüttel in regard to a question using C++ Extensions in relation with the Armadillo library. 

At first I have to make compliments to the developers of Rcpp/RcppArmadillo. Dirk, Francois, this is a marvelous work! As someone programming a lot in C++ and using R Extensions regularly, it is cleaning away all this cumbersome programming dirt connected to SEXPs. 

There are still remaining questions for me, which can be surely answered by the subscribers to this list: 

1. I saw the Rcpp.package.skeleton function and I ask myself, if a similar offer is made for RcppArmadillo for including automatically the Armadillo library in a package? 

2. If I want to use Armadillo solely in the C++ files, not via inline functions as shown in the RcppArmadillo paper, should I use solely the Rcpp package? No, right? The RcppArmadillo package provides objects wrappers for Armadillo objects to be passed to R? 

3. My package is based on S4 classes and I saw the S4 class wrapping in the Rcpp package. I miss an example on this. Can you refer to any document or website for this issue? 

4. Further: What is your experience regarding performance with S4 classes and OOP in C++: Does it make a difference mapping the S4 class to a struct in C++ or using directly the attributes of the S4 class (like vectors, etc.) as Armadillo vectors etc. in C++?
As I work a lot on the HPC in Aachen/Germany together with some of the contributors to the OpenMP API, I am highly influenced by the parsimonious approach, i.e. use only basic objects in C++ to get high performance (although I know, that one of the main work now in the OpenMP API is the extension to complex/user-defined objects inside the #pragmas). 

5. Using OpenMP with RcppArmadillo: Up to now I used almost exclusively the Scythe Statistical Library (http://scythe.wustl.edu), which is pretty fast. I encountered lately problems with it, using parallel computing (OpenMP). Also important in this regard is the possibility to apply a 'first-touch-principle' where all my approaches failed in Scythe, due to the object structure. 
Now, I would like to use RccpArmadillo with OpenMP and 
		a) I want to get best performance: how would I proceed?
	
		b) I want to apply the 'first-touch-principle': where do I apply it? What is the internal data structure in Rcpp-/RcppArmadillo-Objects that allocates the memory?

I am very excited now to start work with the Rcpp/RcppArmadillo package in my own one, which is at least planned to be pushed to CRAN one day. 

I am looking forward to your answers

Best

Simon