In r-lib/vctrs: Vector Helpers

knitr::opts_chunk$set(collapse = TRUE, comment = "#> ")

Exploration of the performance of unique as a proxy for the performance of the underyling dictionary + hash code.

library(tidyverse)
library(vctrs)
library(bench)

make_vec <- function(type, n, mix) {
  x <- switch(as.character(mix),
    unique = 1:n,
    constant = rep(0L, n),
    "50-50" = c(1:(n/2), rep(0L, n / 2))
  )

  as.vector(x, as.character(type))
}

Proportion of uniqueness

df <- bench::press(
  n = c(1e1, 1e2, 1e3, 1e4, 1e5, 1e6),
  mix = c("unique", "constant", "50-50"),
  {
    x <- make_vec("integer", n, mix)
    bench::mark(
      base = unique(x),
      vctrs = vec_unique(x),
      min_time = 0.05
    )
  }
)

Relatively small difference. Unsurprisingly, unique is the worst case, so I'll just only use that going forward.

ggplot(df, aes(n, as.numeric(min))) + 
  geom_point() + 
  geom_line(aes(colour = mix)) + 
  scale_x_log10() + 
  scale_y_log10() + 
  facet_wrap(~ expression)

Performance by type

df <- bench::press(
  type = c("integer", "double", "character"),
  n = c(1e1, 1e2, 1e3, 1e4, 1e5, 1e6),
  {
    x <- make_vec(type, n, "unique")
    bench::mark(
      base = unique(x),
      vctrs = vec_unique(x),
      min_time = 0.05,
      max_iterations = 20
    )
  }
)

ggplot(df, aes(n, as.numeric(min))) + 
  geom_point() + 
  geom_line(aes(colour = type)) + 
  scale_x_log10() + 
  scale_y_log10() + 
  facet_wrap( ~ expression)

Let's compare head-to-head with base. Clearly there's some larger setup cost needed for vctrs dictionary, and the performance difference declines as size increases.

ggplot(df, aes(n, as.numeric(min))) + 
  geom_point() + 
  geom_line(aes(colour = expression)) + 
  scale_x_log10() + 
  scale_y_log10() + 
  facet_wrap( ~ type)

Raw C code

If we just call the C function that underlies vec_unique() (i.e. ignoring the code that makes these functions generic) we see very little performance difference:

df <- bench::press(
  type = c("integer", "double", "character"),
  n = c(1e1, 1e2, 1e3, 1e4, 1e5, 1e6),
  {
    x <- make_vec(type, n, "unique")
    bench::mark(
      base = unique(x),
      vctrs = .Call(vctrs:::vctrs_unique_loc, x),
      min_time = 0.05,
      max_iterations = 20,
      check = FALSE
    )
  }
)

ggplot(df, aes(n, as.numeric(min))) + 
  geom_point() + 
  geom_line(aes(colour = expression)) + 
  scale_x_log10() + 
  scale_y_log10() + 
  facet_wrap(~ type)

r-lib/vctrs documentation built on Oct. 30, 2024, 8:54 a.m.