Specialized templates for faster code?

Fixing dplyr’s behavior for summarise() on empty data frames (#3274) led to a review of all hybrid handlers for this corner case. Indeed, some hybrid handlers didn’t account for this in some cases.

• bugs hidden in specialized templates
• replaced with more generic code
  • conditioning on template parameter, targeting dead branch elimination
  • downside: code slightly more difficult to read (but fewer repetitions)
• important change: specialized code used the fact that is.na(x + NA) and avoided an early return, at least for the REALSXP version
• performance impact?

Microbenchmark to the rescue, using flights dataset with integer and numeric columns that have some NA values or don’t have NA values:

library(dplyr)

## 
## Attaching package: 'dplyr'

## The following objects are masked from 'package:stats':
## 
##     filter, lag

## The following objects are masked from 'package:base':
## 
##     intersect, setdiff, setequal, union

library(nycflights13)
summarise_at(flights, vars(arr_time, day, arr_delay, distance), funs(sum(is.na(.))))

## # A tibble: 1 x 4
##   arr_time   day arr_delay distance
##      <int> <int>     <int>    <int>
## 1     8713     0      9430        0

Benchmarking code, run for both master and the pull request:

library(nycflights13)
library(dplyr)

res <- microbenchmark::microbenchmark(
  integer_mean = summarise(flights, mean(arr_time)),
  integer_var = summarise(flights, var(arr_time)),
  integer_sum = summarise(flights, sum(arr_time)),

  integer2_mean = summarise(flights, mean(day)),
  integer2_var = summarise(flights, var(day)),
  integer2_sum = summarise(flights, sum(day)),

  real_mean = summarise(flights, mean(arr_delay)),
  real_var = summarise(flights, var(arr_delay)),
  real_sum = summarise(flights, sum(arr_delay)),

  real2_mean = summarise(flights, mean(distance)),
  real2_var = summarise(flights, var(distance)),
  real2_sum = summarise(flights, sum(distance)),

  integer_mean_na_rm = summarise(flights, mean(arr_time, na.rm = TRUE)),
  integer_var_na_rm = summarise(flights, var(arr_time, na.rm = TRUE)),
  integer_sum_na_rm = summarise(flights, sum(arr_time, na.rm = TRUE)),

  integer2_mean_na_rm = summarise(flights, mean(day, na.rm = TRUE)),
  integer2_var_na_rm = summarise(flights, var(day, na.rm = TRUE)),
  integer_2sum_na_rm = summarise(flights, sum(day, na.rm = TRUE)),

  real_mean_na_rm = summarise(flights, mean(arr_delay, na.rm = TRUE)),
  real_var_na_rm = summarise(flights, var(arr_delay, na.rm = TRUE)),
  real_sum_na_rm = summarise(flights, sum(arr_delay, na.rm = TRUE)),

  real2_mean_na_rm = summarise(flights, mean(distance, na.rm = TRUE)),
  real2_var_na_rm = summarise(flights, var(distance, na.rm = TRUE)),
  real2_sum_na_rm = summarise(flights, sum(distance, na.rm = TRUE)),

  unit = "ms",
  times = 10
) %>%
  summary() %>%
  as_tibble()

file <- paste0(git2r::revparse_single(git2r::repository("."), "HEAD")@sha, ".rds")
saveRDS(res, file)

res

The 2 suffix means that no NA values are present.

Comparing results:

master <-
  readRDS("cf542082bcaf408c81d2f4b8c60fe9714989514e.rds") %>% 
  mutate(code = "master")
pr <-
  readRDS("0d22c5653592da9e86d6fb24617a3545a45287e0.rds") %>% 
  mutate(code = "pr")

both <-
  bind_rows(master, pr) %>% 
  mutate(
    type = gsub("2?_.*$", "", expr),
    has_na = !grepl("2", expr),
    fun = gsub("^[^_]+_([^_]+)(_na_rm)?$", "\\1", expr),
    na_rm = grepl("na_rm", expr)
  ) %>%
  select(-expr) %>%
  mutate_at(vars(code, type, fun), funs(factor(., levels = unique(.))))

library(ggplot2)
both %>%
  mutate(time_ms = time / 1e6) %>%
  ggplot(aes(x = fun, y = time_ms, fill = code)) +
  geom_col(position = "dodge") +
  facet_wrap(type + na_rm ~ has_na, labeller = label_both, scales = "free_y", ncol = 2)

both %>%
  tidyr::spread(code, time) %>% 
  lm((pr / master) ~ ., .) %>% 
  summary()

## 
## Call:
## lm(formula = (pr/master) ~ ., data = .)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.053676 -0.030997 -0.009526  0.028278  0.075004 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.10211    0.02141  51.470   <2e-16 ***
## typereal    -0.02679    0.01748  -1.532    0.143    
## has_naTRUE  -0.01588    0.01748  -0.908    0.376    
## funvar       0.01275    0.02141   0.595    0.559    
## funsum       0.00549    0.02141   0.256    0.801    
## na_rmTRUE   -0.04818    0.01748  -2.756    0.013 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.04283 on 18 degrees of freedom
## Multiple R-squared:  0.382,  Adjusted R-squared:  0.2103 
## F-statistic: 2.225 on 5 and 18 DF,  p-value: 0.0965