Reduce panel parameter setup in facetted plots

[teunbrand]

This PR aims to fix #5427.

Briefly, it determines unique combinations of x/y scales and (1) only sets up panel parameters and guides and (2) renders axes for these unique combinations. These are then repeated for the duplicated combinations.

In theory, the panel's x parameters and y parameters could be setup independently to further reduce the amount of setup is performed, but the Guide$transform method should have access to both at the same time.

Below is a reprex from the issue that shows that the x-axis is only trained and drawn once:

devtools::load_all("~/packages/ggplot2")
#> ℹ Loading ggplot2

train <- 0
draw  <- 0

# Setting up a counting axis
x <- guide_axis()
x$train <- function(self, params, scale, aesthetic = NULL, ...) {
  train <<- train + 1
  GuideAxis$train(params, scale, aesthetic, ...)
}
x$draw <- function(self, theme, params) {
  draw <<- draw + 1
  GuideAxis$draw(theme, params)
}

ggplot(mpg, aes(displ, hwy)) +
  geom_point() +
  guides(x = x) +
  facet_grid(drv ~ cyl)

train
#> [1] 1
draw
#> [1] 1

^{Created on 2023-09-21 with reprex v2.0.2}

This mostly will affect plots that have many facets but no/few unique scales. Benchmark below is for an optimal case (100 panels all sharing all scales) using this PR. All the panel param setup and drawing happens after ggplot_build(), so I'm testing the ggplot_table() performance.

devtools::load_all("~/packages/ggplot2")
#> ℹ Loading ggplot2

df <- data.frame(x = rnorm(10000), y = rnorm(10000),
                 facet = sample(100, 10000, replace = TRUE))

p <- ggplot(df, aes(x, y)) +
  geom_point() +
  facet_wrap(~ facet)

build <- ggplot_build(p)

bench::mark(ggplot_gtable(build), min_iterations = 10)
#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 1 × 6
#>   expression                min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>           <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 ggplot_gtable(build)    948ms    993ms     0.964    15.5MB     4.82

^{Created on 2023-09-21 with reprex v2.0.2}

Compared to the same benchmark for the current main branch:

#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 1 × 6
#>   expression                min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>           <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 ggplot_gtable(build)    1.82s     1.9s     0.522    18.1MB     5.64

To me this is a clear performance benefit in this optimal case.

[teunbrand]

@clauswilke Do you happen to know if there is any case that coord_sf() will need to draw a different graticule per panel in a facetted plot with fixed scales? It's a tiny change to revert the behaviour in this PR for coord_sf().

[clauswilke]

If every panel uses the same projection and the same scales then the graticule should be identical. I'm only moderately confident in this statement though, haven't thought it through that carefully.

[teunbrand]

Thanks Claus! I'll take your 'moderately confident' over my 'I guess so' anytime.

[thomasp85]

How will this affect extension packages with new facets? Do they need updating?

[teunbrand]

Do they need updating?

I expect they don't need any updates. The compression is within each step but between each step the full, expanded panel parameters are available, so this shouldn't affect anything outside the methods affected.

[teunbrand]

I just realised that this would affect extensions where a RNG is used in any of the affected methods, but I expect this not to be a typical case as most plots are deterministic.

[thomasp85]

LGTM

[teunbrand]

I didn't see it at the time writing this PR, but it conflicts with #4467.
That conflict basically means we cannot do the optimisation at the facet level, as we need separately drawn axes for interior panels versus exterior panels.
At the coord level, it takes ggplot_build(p) from ~737ms to ~23.8ms, so I'd argue that is still a worthwhile improvement (note that guide setup has moved from the gtable step to the build step in 3.5.0 after this PR was written).

[teunbrand]

Asking @thomasp85 to see if his approval still holds with the coord-level optimisation only.