Question

I have an R/ggplot2 use case that seems to call for geom_raster: a regular Cartesian grid with z-values at x, y locations. I’ve been using geom_tile, and I expected a performance improvement from switching to geom_raster. But I don’t seem to be seeing one…

Here’s a toy example (but about the right size), using base graphics:

n <- m <- 200
x <- 1:n
y <- 1:m
f <- function(x, y) 10 * sin(x / n) * cos(y / m)
z <- outer(x, y, f)
system.time(image(z))

   user  system elapsed 
  0.998   0.007   1.023 

Here it is with ggplot2:

obs <- expand.grid(x=x, y=y)
obs$z <- as.numeric(as.list(z))
require(ggplot2)
p <- ggplot(obs, aes(x=x, y=y, fill=z))
system.time(show(p + geom_tile()))

   user  system elapsed 
  7.328   0.891   8.187 

require(ggExtra)
system.time(show(p + geom_raster()))

   user  system elapsed 
  7.000   0.637   7.799

So, a modest gain, but nowhere near what I was expecting. Am I doing it wrong? Many thanks in advance!

Answer 1

You should use geom_raster from the latest ggplot2 (dev version, currently), not the buggy prototype in ggExtra (this package is now deprecated, btw).

Doing so, I get better results, 4.705 vs. 1.416 (elapsed). Quite an improvement.

Edit: it turns out that ?geom_raster in ggplot2 already offers a better benchmark, on my system

benchplot(base + geom_raster())
       step user.self sys.self elapsed
1 construct     0.006    0.004   0.010
2     build     0.887    0.212   1.109
3    render     0.449    0.119   0.584
4      draw     0.108    0.005   0.141
5     TOTAL     1.450    0.340   1.844
> benchplot(base + geom_tile())
       step user.self sys.self elapsed
1 construct     0.016    0.005   0.026
2     build     1.031    0.329   1.365
3    render     1.021    0.297   1.318
4      draw     0.987    0.041   1.040
5     TOTAL     3.055    0.672   3.749