Hi,
On 10 August 2010 07:51, Romain Francois <rom...@r-enthusiasts.com> wrote:
> Le 10/08/10 07:35, baptiste auguie a écrit :
>>
>> I see –– thanks for the clarification (it is indeed what Dirk really
>> had suggested, the accidental shortcut was mine). This passing by
>> reference seems very interesting (I pass several large complex arma
>> matrices in other functions), but I need to work out how it works
>> inside the functions.
>
> It should make no difference in how you use the object between :
>
> void foo( arma:::cx_mat m ) ;
> void foo( const arma::cx_mat& m ) ;
OK, thanks. I have not been able to produce a minimal code that would
exhibit an improved performance using this passing-by-reference idea.
My naive attempt follows,
extern "C" {
arma::mat dum1(NumericMatrix A)
{
arma::mat B = Rcpp::as< arma::mat >( A ) ;
B=inv(B*38);
return (B);
}
arma::mat dum2(const NumericMatrix& A)
{
arma::mat B = Rcpp::as< arma::mat >( A ) ;
B=inv(B*38);
return (B);
}
RCPP_FUNCTION_1(NumericMatrix, dummy1, NumericMatrix A) {
return wrap(dum1(A));
}
RCPP_FUNCTION_1(NumericMatrix, dummy2, NumericMatrix A) {
return wrap(dum2(A));
}
}
Timing a 10^3 x 10^3 matrix yields similar results for both functions.
I wonder in what situations passing by reference may improve
efficiency; also it is not clear to me whether there can be downsides
to this approach.
Thanks again,
baptiste
>
> inside foo, you use m exactly the same way. Although the const here is there
> to protect you from yourself. Having the const here means that inside foo
> you are only allowed to call const methods of the arma::cx_mat class, so you
> can't modify the object.
>
> If you want to be able to modify the object, you could use
>
> void foo( arma::cx_mat& m ) ;
>
>> Thanks again for the precious comments,
>>
>> baptiste
>>
>> PS: my package is now on r-forge, hoping to see if it compiles there too:
>>
>> https://r-forge.r-project.org/scm/viewvc.php/pkg/cda/?root=photonics
>
> cool. I'll make a few targetted comments when I have time to read it more
> carefully, but so far it looks good.
>
>> On 10 August 2010 01:38, Davor Cubranic<cubra...@stat.ubc.ca> wrote:
>>>
>>> On 2010-08-07, at 7:03 PM, baptiste auguie wrote:
>>>
>>>> Thanks Davor, Douglas and Dirk –– all your comments have been very
>>>> helpful and I used them to improve my code (along with getting a
>>>> better understanding of C++). The only thing I haven't tried yet is
>>>> avoid duplicating the objects in memory using const (seems tricky from
>>>> what I read on the net).
>>>
>>> Using 'const' won't save you from duplicating objects: all it does is
>>> mark a variable/parameter as a constant, so that you can't accidentally
>>> change it. What does save you from duplication is passing arguments by
>>> reference. And passing arguments by reference is often combined with marking
>>> them 'const' for safety, because accidentally changing them would introduce
>>> wide-ranging bugs into your code that could be difficult to find.
>>>
>>> Also, note that passing by const reference only saves space for complex
>>> objects, not primitive types like doubles and consts. So for the 'euler'
>>> function you used in your example, it would be recommended to make its
>>> signature:
>>>
>>> arma::mat euler(const double phi, const double theta, const double psi);
>>> In other words, 'phi', 'theta', and 'psi' are doubles, which I'm passing
>>> as constants.
>>>
>>> While something like matrix multiply could be:
>>>
>>> arma::mat mul(const arma::mat& a, const arma::mat& b);
>>> Here, 'a' and 'b' are matrix objects, which I'm passing by constant
>>> reference.
>>>
>>> Hope this helps.
>>>
>>> Davor
>
> --
> Romain Francois
> Professional R Enthusiast
> +33(0) 6 28 91 30 30
> http://romainfrancois.blog.free.fr
> |- http://bit.ly/aAyra4 : highlight 0.2-2
> |- http://bit.ly/94EBKx : inline 0.3.6
> `- http://bit.ly/aryfrk : useR! 2010
>
>
--
____________________
Dr. Baptiste Auguié
Departamento de Química Física,
Universidade de Vigo,
Campus Universitario, 36310, Vigo, Spain
tel: +34 9868 18617
http://webs.uvigo.es/coloides
____________________
_______________________________________________
Rcpp-devel mailing list
Rcpp-devel@lists.r-forge.r-project.org
https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
