一ifelse
开始就省钱可以节省 57%...
almostEqual2 <- function(x, y, tolerance=1e-8) {
diff <- abs(x - y)
mag <- pmax( abs(x), abs(y) )
out <- logical(length(y))
out[ mag > tolerance ] <- (diff/mag <= tolerance)[ mag > tolerance]
out[ ! mag > tolerance ] <- (diff <= tolerance)[! mag > tolerance]
return( out )
}
require(microbenchmark)
set.seed(1)
x <- 1
y <- rnorm(1e6)
bm <- microbenchmark( almostEqual(x,y,tol=0.5) , almostEqual2(x,y,tol=0.5) , times = 25 )
print( bm , digits = 3 , unit = "relative" , order = "median" )
#Unit: relative
# expr min lq median uq max neval
# almostEqual2(x, y, tol = 0.5) 1.00 1.00 1.00 1.00 1.00 25
# almostEqual(x, y, tol = 0.5) 2.09 1.76 1.73 1.86 1.82 25
使用 Rcpp
我不明白为什么你不会在 CRAN 之外使用最依赖的包base
,但如果你愿意的话,你可以实现比我之前的努力提高 5 倍(在 OP 上 10 倍)并且它还可以处理NA 优雅地...
#include <Rcpp.h>
using namespace Rcpp;
//[[Rcpp::export]]
LogicalVector all_equalC( double x , NumericVector y , double tolerance ){
NumericVector diff = abs( x - y );
NumericVector mag = pmax( abs(x) , abs(y) );
LogicalVector res = ifelse( mag > tolerance , diff/mag <= tolerance , diff <= tolerance );
return( res );
}
使用Rcpp::sourceCpp('path/to/file.cpp')
. 结果...
bm <- microbenchmark( almostEqual(x,y,tol=0.5) , almostEqual2(x,y,tol=0.5) , all_equalC(x,y,tolerance=0.5) , times = 25 )
print( bm , digits = 3 , unit = "relative" , order = "median" )
#Unit: relative
# expr min lq median uq max neval
# all_equalC(x, y, tolerance = 0.5) 1.00 1.00 1.00 1.00 1.00 25
# almostEqual2(x, y, tol = 0.5) 4.50 4.39 5.39 5.24 7.32 25
# almostEqual(x, y, tol = 0.5) 8.69 9.34 9.24 9.96 10.91 25