Hi,
I don’t see any code going with the revised version, but
looking at the code coming with the original paper, see
http://economics.sas.upenn.edu/~jesusfv/RBC_codes.zip,
the following points come up:
1)There is no Rcpp code, so there is no way to verify the claims about Rcpp.
2)It is not at all obvious that the comparison is about computation – the
following chunk for diagnostic printing is inside a double for loop (this is
the .R version):
===
iteration = iteration+1;
if ((iteration %% 10)==0 | iteration ==1){
cat(" Iteration = ", iteration," Sup Diff = ", maxDifference,"\n");
}
===
Printing to the R console may well take significant proportion of the time.
3)The R version is a direct translation of the cpp version. This in itself is
not a fair comparison. Moreover, note that the snipped in the cpp version of
the above chunk uses the ‘||’ operator, while the R chunk above for some reason
uses the ‘|’ operator, so the second condition in the ‘if’ statement is
executed at every iteration.
4)The assignments of the data are done at compile time, while in R at execution
time. Moreover there are man redundant operations in the R code, eg
unnecessary conversions by as.matrix in the R code, which are absent in the cpp
code.
5)The cpp function is for fixed data. Hence, the compilation and linking time
need to be added to the overall time.
In summary, I would say that these results show, at best, that if a C++
function is converted to an interpreted language it will be slower, hardly
surprising. The exact timings do not seem relevant since they mix computation
and output. There is also evidence that the authors have been more careful with
their C++ code than the R code (I haven’t looked at the others). The claim
about Rcpp is completely unverifiable. Moreover, for a fixed task, the time
with the C++ executable should have included compilation, linking and invoking
the executable.
I would ignore these comparisons and look at more reliable sources, such as
examples by Dirk in the Rcpp package and other materials and papers by him.
Georgi Boshnakov
From: Rcpp-devel [mailto:rcpp-devel-boun...@lists.r-forge.r-project.org] On
Behalf Of Jordi Molins
Sent: 22 September 2018 08:03
To: rcpp-devel@lists.r-forge.r-project.org; Jordi Molins
Subject: [Rcpp-devel] Question on performance
Hello, I am new in this distribution list. I am using Rcpp for my project, and
I enjoy the Rcpp implementation quite a lot.
I have read the paper
https://www.sas.upenn.edu/~jesusfv/Update_March_23_2018.pdf and I see the
performance of Rcpp is about 4x times slower than C++ proper code, or 3x times
slower than Numba/Cython Python code.
Do you think that these figures are realistic? Of course, performance
calculations are always dependent on many factors (very often, not explicit).
But I am surprised that Rcpp is that slow in comparison to C++ code. I do not
understand that Rcpp may be so slow, unless there are many calls back and forth
between Rcpp and R. The same for compiled Python code vis a vis Rcpp.
Or maybe I am wrong?
Jordi Molins i Coronado
+34 69 38 000 59
_______________________________________________
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
