Ditch Rcpp

[mpadge]

Ditch Rcpp

[mpadge]

And now the timing comparisons:

Compile and load times

Compiling and loading the C version is enormously faster than Rcpp. Total compile-and-load times depend on whether it's done within the same R session or within a new one.

f <- function () {
    pt <- proc.time ()
    devtools::load_all (".", export_all = FALSE, recompile = TRUE)
    print (proc.time () - pt)
}

Version	same R session?	time (s)
Rcpp	no	5.67
Rcpp	yes	5.29
C	no	0.83
C	yes	0.41

So compiling and loading package is around 10 times faster without Rcpp.

Installed package sizes

I used this script to get the sizes of the installed packages:

install.packages("../geodist_0.0.0.9000.tar.gz")
library(fs)
library(tidyverse)
library(rvest)
installed.packages() %>%
    as_data_frame() %>%
    filter (Package == "geodist") %>%
    mutate(pkg_dir =
           sprintf("%s/%s", LibPath,
                   Package)) %>%
    select(Package, pkg_dir) %>%
    mutate(pkg_dir_size = map_dbl(pkg_dir, ~{ fs::dir_info(.x, all=TRUE,
                                                           recursive=TRUE) %>%
                       summarise(tot_dir_size = sum(size)) %>% 
                       pull(tot_dir_size)
               })) %>%
    select (Package, pkg_dir_size)

Results:

Version	size (kB)
Rcpp	176.955
C	50.528

So Rcpp inflates the package size to over three times. Finally, the most important comparison

Calculation speed

Based on the following function:

f <- function ()
{
    n <- 1e3
    x <- cbind (-180 + 360 * runif (n), -90 + 180 * runif (n))
    y <- cbind (-180 + 360 * runif (2 * n), -90 + 180 * runif (2 * n))
    colnames (x) <- colnames (y) <- c ("x", "y")
    pt <- proc.time ()
    for (i in 1:10)
    {
        d <- geodist (x, y)
    }
    pt <- proc.time () - pt
    print (pt)
}

Version	same R session?	time (s)
Rcpp	no	6.91
Rcpp	yes	3.92
C	no	2.61
C	yes	2.32

And the C version is definitely faster, but depends very much on whether repeats are conducted within the same R session or not. For a one-off calculation (that is, within a fresh session), the C version is about 3 times faster, reducing to only about 40% faster if repeated within the same session. Either way, Rcpp successfully ditched with some very positive results.

cc @karpfen