[Rd] R-devel does not update the C++ returned variables

[Martin Maechler]

> On 2 March 2015 at 09:09, Duncan Murdoch wrote:
> | I generally recommend that people use Rcpp, which hides a lot of the 
> | details.  It will generate your .Call calls for you, and generate the 
> | C++ code that receives them; you just need to think about the real 
> | problem, not the interface.  It has its own learning curve, but I think 
> | it is easier than using the low-level code that you need to work with .Call.

> Thanks for that vote, and I second that.

> And these days the learning is a lot flatter than it was a decade ago:

> R> Rcpp::cppFunction("NumericVector doubleThis(NumericVector x) { return(2*x); }")
> R> doubleThis(c(1,2,3,21,-4))
> [1]  2  4  6 42 -8
> R>

> That defined, compiled, loaded and run/illustrated a simple function. 

> Dirk

Indeed impressive,  ... and it also works with integer vectors
something also not 100% trivial when working with compiled code.

When testing that, I've went a step further:

##---- now "test":
require(microbenchmark)
i <- 1:10
(mb <- microbenchmark(doubleThis(i), i*2, 2*i, i*2L, 2L*i, i+i, times=2^12))
## Lynne (i7; FC 20), R Under development ... (2015-03-02 r67924):
## Unit: nanoseconds
##           expr min  lq      mean median   uq   max neval cld
##  doubleThis(i) 762 985 1319.5974   1124 1338 17831  4096   b
##          i * 2 124 151  258.4419    164  221 22224  4096  a 
##          2 * i 127 154  266.4707    169  216 20213  4096  a 
##         i * 2L 143 164  250.6057    181  234 16863  4096  a 
##         2L * i 144 177  269.5015    193  237 16119  4096  a 
##          i + i 152 183  272.6179    199  243 10434  4096  a 

plot(mb, log="y", notch=TRUE)
## hmm, looks like even the simple arithm. differ slightly ...
##
## ==> zoom in:
plot(mb, log="y", notch=TRUE, ylim = c(150,300))

dev.copy(png, file="mbenchm-doubling.png")
dev.off() # [ <- why do I need this here for png ??? ]
##--> see the appended *png graphic

Those who've learnt EDA or otherwise about boxplot notches, will
know that they provide somewhat informal but robust pairwise tests on
approximate 5% level.
>From these, one *could* - possibly wrongly - conclude that
'i * 2' is significantly faster than both 'i * 2L' and also
'i + i' ---- which I find astonishing, given that  i is integer here...

Probably no reason for deep thoughts here, but if someone is
enticed, this maybe slightly interesting to read.

Martin Maechler, ETH Zurich

-------------- next part --------------
A non-text attachment was scrubbed...
Name: mbenchm-doubling.png
Type: image/png
Size: 7244 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-devel/attachments/20150302/0f61efab/attachment.png>