- vectorized calculations or<\/strong><\/li>
- if vectorization is not possible, use sapply() et al.<\/strong><\/li><\/ol>\n\n\n\n
Since then, the R core team and the community has invested tons of time to improve R and also to make it faster. There are things like RCPP and parallel computing to speed up loops. <\/p>\n\n\n\n
But what still relatively few R users know: loops are not that slow anymore<\/strong>. We want to demonstrate this using two examples. <\/p>\n\n\n\n
Example 1: sqrt()<\/h2>\n\n\n\n
We use three ways to calculate the square root of a vector of random numbers:<\/p>\n\n\n\n
- Vectorized calculation. This will be the way to go because it is internally optimized in C.<\/li>
- A loop. This must be super slow for large vectors.<\/li>
- vapply() (as safe alternative to sapply).<\/li><\/ol>\n\n\n\n
The three approaches are then compared via bench::mark() regarding their speed for different numbers n<\/em> of vector lengths. The results are then compared first regarding absolute median times, and secondly (using an independent run), on a relative scale (1 is the vectorized approach). <\/p>\n\n\n
\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\nlibrary(tidyverse)\nlibrary(bench)\n\n# Calculate square root for each element in loop\nsqrt_loop <- function(x) {\n out <- numeric(length(x))\n for (i in seq_along(x)) {\n out[i] <- sqrt(x[i])\n }\n out\n}\n\n# Example\nsqrt_loop(1:4) # 1.000000 1.414214 1.732051 2.000000\n\n# Compare its performance with two alternatives\nsqrt_benchmark <- function(n) {\n x <- rexp(n)\n mark(\n vectorized = sqrt(x),\n loop = sqrt_loop(x),\n vapply = vapply(x, sqrt, FUN.VALUE = 0.0),\n # relative = TRUE\n )\n}\n\n# Combine results of multiple benchmarks and plot results\nmultiple_benchmarks <- function(one_bench, N) {\n res <- vector(\"list\", length(N))\n for (i in seq_along(N)) {\n res[[i]] <- one_bench(N[i]) %>% \n mutate(n = N[i], expression = names(expression))\n }\n \n ggplot(bind_rows(res), aes(n, median, color = expression)) +\n geom_point(size = 3) +\n geom_line(size = 1) +\n scale_x_log10() +\n ggtitle(deparse1(substitute(one_bench))) +\n theme(legend.position = c(0.8, 0.15))\n}\n\n# Apply simulation\nmultiple_benchmarks(sqrt_benchmark, N = 10^seq(3, 6, 0.25))<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\nAbsolute timings<\/h3>\n\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/02\/Rplot02-1.jpeg\")
Absolute median times on the “sqrt()” task<\/figcaption><\/figure>\n\n\n\n Relative timings (using a second run)<\/h3>\n\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/02\/Rplot05.jpeg\")
Relative median times of a separate run on the “sqrt()” task <\/figcaption><\/figure>\n\n\n\n We see:<\/p>\n\n\n\n
- Run times increase quite linearly with vector size.<\/li>
- Vectorization is more than ten times faster than the naive loop.<\/li>
- Most strikingly, vapply() is much slower than the naive loop. Would you have thought this?<\/li><\/ul>\n\n\n\n
Example 2: paste()<\/h2>\n\n\n\n
For the second example, we use a less simple function, namely<\/p>\n\n\n\n
paste(“Number”, prettyNum(x, digits = 5))<\/p>\n\n\n\n
What will our three approaches (vectorized, naive loop, vapply) show on this task?<\/p>\n\n\n
\n\t\t\t\n\t\t\t\t\n\t\t\t\tR<\/div>\n\t\t\t<\/div><\/div>\n\t\t\t<\/div>\n\t\t\t\n\npretty_paste <- function(x) {\n paste(\"Number\", prettyNum(x, digits = 5))\n}\n\n# Example\npretty_paste(pi) # \"Number 3.1416\"\n\n# Again, call pretty_paste() for each element in a loop\npaste_loop <- function(x) {\n out <- character(length(x))\n for (i in seq_along(x)) {\n out[i] <- pretty_paste(x[i])\n }\n out\n}\n\n# Compare its performance with two alternatives\npaste_benchmark <- function(n) {\n x <- rexp(n)\n mark(\n vectorized = pretty_paste(x),\n loop = paste_loop(x),\n vapply = vapply(x, pretty_paste, FUN.VALUE = \"\"),\n # relative = TRUE\n )\n}\n\nmultiple_benchmarks(paste_benchmark, N = 10^seq(3, 5, 0.25))<\/pre><\/div>\n\n<\/div><\/div>\n\t\t<\/div>\n\n\nAbsolute timings<\/h3>\n\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/02\/Rplot03.jpeg\")
Absolute median times on the “paste()” task <\/figcaption><\/figure>\n\n\n\n Relative timings (using a second run)<\/h3>\n\n\n\n
![\"\"](\"https:\/\/lorentzen.ch\/wp-content\/uploads\/2022\/02\/Rplot04.jpeg\")
Relative median times of a separate run on the “paste()” task<\/figcaption><\/figure>\n\n\n\n - In contrast to the first example, vapply() is now as fast as the naive loop.<\/li>
- The time advantage of the vectorized approach is much less impressive. The loop takes in median only 50% longer.<\/li><\/ul>\n\n\n\n
Conclusion<\/h2>\n\n\n\n
- Vectorization is fast and easy to read. If available, use this. No surprise.<\/li>
- If you use vapply\/sapply\/lapply, do it for the style, not for the speed. In some cases, the loop will be faster. And, depending on the situation and the audience, a loop might actually be even easier to read. <\/li><\/ol>\n\n\n\n
- if vectorization is not possible, use sapply() et al.<\/strong><\/li><\/ol>\n\n\n\n