[Rcpp-devel] performance profile from Rcpp or RcppArmadillo code? Matrix subviews, etc

[Dirk Eddelbuettel]

On 6 January 2013 at 16:17, Paul Johnson wrote:
| Hi Mario:
| 
| On Sat, Jan 5, 2013 at 9:09 PM, Mario Bourgoin <m.bourgoin at gmail.com> wrote:
| > Hadley's guide to high performance computing using Rcpp was recommended on
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

(To Mario) 

Isn't that a bit of stretch?  It is a very nice intro to Rcpp, and Rcpp helps
you to get results faster.  But that's not really HPC or an HPC intro ...

| I'm stuck at a more fine-grained situation, where I really want a
| profile of a large function using Armadillo and am as frustrated as
| hell trying to get one.
| 
| Following R Writing Extensions, I am trying sprof and oprofile.

(To Paul) 

You can look at my older "Intro to HPC with R" slides which contain worked
examples for profiling. I use the Google Perftools which generate results
which can visualized via graphviz and sliced and diced differently.

Also note that Romain added a class "Timer" directly in Rcpp (0.10.2 or newer)
which works at highest support (kernel-level) granularity.  With that you can
possibly use fairly fine-grained comparisons rather than the "whole program",
or "whole subroutine" which Perftools too uses.

Hth,  Dirk

PS Did you reply on-list to an off-list reply? I never saw Mario's post.

-- 
Dirk Eddelbuettel | edd at debian.org | http://dirk.eddelbuettel.com