fix: sub-day precision Date should be floored when treated as integer

[eitsupi]

Same as apache/arrow-nanoarrow#674, duckdb/duckdb-r#981

Because the Date type in R allows sub-days, it must be rounded to -Inf before being forced to integerish.
Conventional rounding in the direction of zero would change the direction of rounding before and after Unix origin.

[gaborcsardi]

Thanks! I wonder if there is a faster way to do this, because it makes the conversion ~50% slower:

days <- as.Date(rep(c(-1.1, -0.1, 0, 0.1, 1.1), 10000), origin = "1970-01-01")
f1 <- function(x) mode(x) <- "integer"
f2 <- function(x) .Date(as.integer(floor(as.numeric(x))))
bench::mark(f1(days), f2(days), check = FALSE)

# A tibble: 2 × 13
  expression      min median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time
  <bch:expr> <bch:tm> <bch:>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm>
1 f1(days)     95.5µs  105µs     9371.     195KB     31.3  4193    14      447ms
2 f2(days)    105.5µs  147µs     6681.     781KB     96.6  2421    35      362ms
# ℹ 4 more variables: result <list>, memory <list>, time <list>, gc <list>

[gaborcsardi]

I don't really understand why, but this is even faster than the original:

days <- as.Date(rep(c(-1.1, -0.1, 0, 0.1, 1.1), 10000), origin = "1970-01-01")
f1 <- function(x) mode(x) <- "integer"
f2 <- function(x) .Date(as.integer(floor(as.numeric(x))))
f3 <- function(x) { cls <- class(x); x <- floor(as.numeric(x)); class(x) <- cls; x }

stopifnot(f2(days) == f3(days))

bench::mark(f1(days), f2(days), f3(days), check = FALSE)

# A tibble: 3 × 13
  expression     min  median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time
  <bch:expr> <bch:t> <bch:t>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm>
1 f1(days)    95.5µs 104.5µs     9340.     195KB     34.0  4121    15      441ms
2 f2(days)     106µs 150.5µs     6840.     792KB    103.   2649    40      387ms
3 f3(days)    54.3µs  79.5µs    12617.     409KB     92.2  4652    34      369ms
# ℹ 4 more variables: result <list>, memory <list>, time <list>, gc <list>

[gaborcsardi]

Very slightly faster, still:

f4 <- function(x) { cls <- class(x); x <- floor(unclass(x)); class(x) <- cls; x }

[gaborcsardi]

Except that I forgot to convert to integer. /o\ Never mind then.

Proper benchmark:

days <- as.Date(rep(c(-1.1, -0.1, 0, 0.1, 1.1), 10000), origin = "1970-01-01")
f2 <- function(x) .Date(as.integer(floor(as.numeric(x))))
f3 <- function(x) { cls <- class(x); x <- as.integer(floor(as.numeric(x))); class(x) <- cls; x }
f4 <- function(x) { cls <- class(x); x <- as.integer(floor(unclass(x))); class(x) <- cls; x }

waldo::compare(f2(days), f3(days))
#> ✔ No differences

waldo::compare(f2(days), f4(days))
#> ✔ No differences

bench::mark(f2(days), f3(days), f4(days))

# A tibble: 3 × 13
  expression      min median `itr/sec` mem_alloc `gc/sec` n_itr  n_gc total_time
  <bch:expr> <bch:tm> <bch:>     <dbl> <bch:byt>    <dbl> <int> <dbl>   <bch:tm>
1 f2(days)    106.1µs  143µs     6805.     781KB     89.0  2371    31      348ms
2 f3(days)     99.9µs  129µs     7783.     605KB     78.2  2986    30      384ms
3 f4(days)    145.1µs  173µs     5796.     605KB     60.2  1347    14      232ms
# ℹ 4 more variables: result <list>, memory <list>, time <list>, gc <list>

[gaborcsardi]

Thank you!

[eitsupi]

Thanks for taking a look at this. I did a little digging and it seems that the time consuming part is mainly the floor() and the process of converting double to integer.
So to make it any faster, I think it would have to be implemented in C++, but I don't have the C++ skills to do it here.
(Also, I don't know if it's worth it considering the percentage of execution time to the overall process.)

In the polars package I did the processing on the Rust side.
https://github.com/pola-rs/r-polars/blob/b900d6391b82a9245d8529e93f7caeb3ccfede8e/src/rust/src/series/construction.rs#L141-L160