Le 13 d?c. 2016 ? 12:38, Johannes Kruisselbrink <johannes at kruisselbrink.eu> a ?crit :
Here is a further reduced example (see below). Now it is a function to count NaNs in a vector, and it shows the same behaviour. Code is also available at https://github.com/jwkruisselbrink/rcpp-timings/tree/master/minimal.
Regarding your question:
| Why not drop data and codes and use sData1(i,k) - sData2(j,k)
Well, I wanted to rule out that Rcpp sugar was causing the slowdown. If it isn't we'll put it back in after having a fix for this issue.
/=============== call.c ===============/
#include <R.h>
#include <Rinternals.h>
SEXP CountNans(SEXP sX, SEXP sLength) {
int i, n = *INTEGER(sLength);
int *nanCount;
double *x = REAL(sX);
SEXP output = PROTECT(allocVector(INTSXP, 1));
nanCount = INTEGER(output);
*nanCount = 0;
for (i = 0; i < n; i++) {
if (ISNAN(x[i])) {
*nanCount += 1;
}
}
UNPROTECT(1);
return output;
}
/=============== rcpp.cpp ===============/
#include <R.h>
#include <Rcpp.h>
using namespace Rcpp;
// [[Rcpp::export]]
int CountNans(NumericVector x) {
int n = x.length();
int nanCount = 0;
for (int i = 0; i < n; i++) {
if (ISNAN(x[i])) {
nanCount++;
}
}
return nanCount;
}
/=============== R code ===============/
library(Rcpp)
library(microbenchmark)
library(ggplot2)
sourceCpp('rcpp.cpp')
if (is.loaded(paste("call", .Platform$dynlib.ext, sep=""))) {
dyn.unload(paste("call", .Platform$dynlib.ext, sep=""))
}
system(paste("R CMD SHLIB call.c", sep=""))
dyn.load(paste("call", .Platform$dynlib.ext, sep=""))
n <- as.integer(100000)
x <- rnorm(n)
mb <- microbenchmark(
rcpp <- CountNans(x),
call <- .Call("CountNans", x, n),
times = 10000
)
autoplot(mb)