Reduced `fsd()` performance in >=2.0.16

[andrewGhazi]

Hi there, I noticed that the performance of collapse::fsd() seems to have gone down on a little benchmark I've been running between versions 2.0.15 and 2.0.16. I tried a few combinations of R and and collapse going back to 4.2.1 and 1.8.9 respectively and I think the issue stems from collapse. This is the script I've been testing with:

install.packages(c("bench", "waldo", "remotes"))

remotes::install_version("data.table", "1.15.4")
#remotes::install_version("collapse",   "2.0.15")
remotes::install_version("collapse",   "2.0.16")

library(data.table); setDTthreads(1)
library(collapse); set_collapse(nthreads = 1)

n = 3e5

set.seed(123)

val_dt = data.table(g = rep(1:n, each = 6),
                    x = rt(6 * n, 3))

val_dt

dt_f1 = \(val_dt) val_dt[,.(x = sd(x)), by = g]

cl_f1 = \(val_dt) val_dt |> gby(g) |> fsd()

cl_f2 = \(val_dt) val_dt |> gby(g) |> smr(x = fsd(x))

check_fun = \(x,y) length(waldo::compare(x,y, tolerance = 1e-8)) == 0

res = bench::mark(data.table = dt_f1(val_dt),
                  collapse   = cl_f1(val_dt),
                  collapse2  = cl_f2(val_dt),
                  check = check_fun)

sessionInfo()

res |> 
  slt(expression:mem_alloc, n_itr)

On the rocker/r-ver:4.4.2 image with [email protected] that script produces:

# A tibble: 3 × 6
  expression      min   median `itr/sec` mem_alloc n_itr
  <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt> <int>
1 data.table  34.84ms   37.3ms      22.6    53.9MB    12
2 collapse     9.01ms   12.4ms      81.3    24.3MB    41
3 collapse2     8.9ms   12.1ms      75.0    24.2MB    38

But the same script on the same image with [email protected] yields:

# A tibble: 3 × 6
  expression      min   median `itr/sec` mem_alloc n_itr
  <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt> <int>
1 data.table   35.1ms   36.6ms      22.9    53.9MB    12
2 collapse     23.8ms   27.4ms      38.0    24.3MB    19
3 collapse2    44.4ms   48.2ms      21.0    24.2MB    11

The plain fsd() call went from 12ms to 27ms, and the smr(x = fsd(x)) went from 12ms to 48ms. A different test with fmean() showed no differences there, but I didn't try any of the other special fast statistical functions outside of those two.

Just thought I'd flag the issue since it seems it hasn't been noticed yet and I didn't see anything the seemed like it could be related in the NEWS for 2.0.16. I wish I could offer more to identify the root issue but it's beyond my capabilities I think.

[SebKrantz]

Thanks @andrewGhazi, I'll look into this as I find time! In general though if you want speedy standard deviations, do use the option stable.algo = FALSE to fsd(). You can also set it globally using set_collapse().