faster internal version of update_tibble_attrs() by romainfrancois · Pull Request #717 · tidyverse/tibble

romainfrancois · 2020-02-13T14:05:28Z

This has come up as part of performance benchmarks for dplyr 1.0.0:

library(tibble)
library(rlang)

update_tibble_attrs_old <- function(x, ...) {
  # Compat for https://github.com/r-spatial/sf/pull/1232
  attribs <- list(...)
  if (has_length(attribs) && is_named(attribs)) {
    attributes(x)[names(attribs)] <- attribs
  }
  
  x
}
update_tibble_attrs <- tibble:::update_tibble_attrs


df <- tibble(x = rnorm(1e6), g = sample(rep(1:1e4, 100)))
groups <- tibble(g = unique(df$g))

bench::mark(
  new = update_tibble_attrs(df, groups = groups), 
  old = update_tibble_attrs_old(df, groups = groups) 
)
#> # A tibble: 2 x 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 new          2.09µs   2.72µs   344885.        0B      0  
#> 2 old          3.98ms   4.96ms      201.    3.81MB     18.7

^{Created on 2020-02-13 by the reprex package (v0.3.0.9000)}

I believe this is because the form attributes(x)[names(attribs)] <- pays for setting all the attributes, including row.names which is expensive to get/set.

closes tidyverse/dplyr#4855

lionel- · 2020-02-13T15:39:50Z

-  }
-
-  x
+  .Call(`tibble_update_attrs`, x, pairlist2(...))


Probably safer to check if any of the special attributes is supplied, and fall back to the R code in that case?

I'm not sure how safer that would be, setAttrib() does some internal dispatch to deal with special attributes, i.e. names, row.names, class ...

oh yes, I was confused!

There's still some special handling at R level in structure() though, but that doesn't seem important:

specials <- c(".Dim", ".Dimnames", ".Names", ".Tsp", ".Label")

krlmlr · 2020-02-14T06:24:50Z

Thanks! I have removed compat code that is no longer necessary with the sf update. It's still very slow, though.

I have also added a test regarding overwriting of attributes. I think our implementation will need to account for this. Do we need to adapt?

romainfrancois · 2020-02-14T07:40:12Z

structure() seems to be as slow:

library(tibble)
library(rlang)

update_tibble_attrs_old <- function(x, ...) {
  # Compat for https://github.com/r-spatial/sf/pull/1232
  attribs <- list(...)
  if (has_length(attribs) && is_named(attribs)) {
    attributes(x)[names(attribs)] <- attribs
  }
  x
}
update_tibble_attrs <- tibble:::update_tibble_attrs

df <- tibble(x = rnorm(1e6), g = sample(rep(1:1e4, 100)))
groups <- tibble(g = unique(df$g))

bench::mark(
  new = update_tibble_attrs(df, groups = groups),
  old = update_tibble_attrs_old(df, groups = groups), 
  str = structure(df, groups = groups)
)
#> # A tibble: 3 x 6
#>   expression      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr> <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 new          2.02µs   2.71µs   335090.        0B      0  
#> 2 old          3.97ms   4.77ms      211.    3.82MB     22.4
#> 3 str          3.96ms    4.3ms      226.    3.81MB     15.8

^{Created on 2020-02-14 by the reprex package (v0.3.0.9000)}

romainfrancois · 2020-02-14T07:46:59Z

which test @krlmlr ?

romainfrancois · 2020-02-14T10:37:26Z

I believe vectbl_restore() has the same issue, i.e. using attributes()[<-

vectbl_restore <- function(xo, x) {
  n <- fast_nrow(xo)

  # FIXME: Use new_tibble()?
  forbidden <- attrs_names_only
  attrs <- attributes(x)
  attrs <- attrs[!(names(attrs) %in% forbidden)]

  attributes(xo)[names(attrs)] <- attrs

  # Need to patch manually
  attr(xo, "row.names") <- .set_row_names(n)

  xo
}

krlmlr · 2020-02-14T11:56:08Z

Good catch, pushed the tests.

I still wonder why vectbl_restore() seems to be much faster:

library(tibble)

x <- tibble(a = 1:3)
y <- x

vectbl_restore <- tibble:::vectbl_restore

bench::mark(vectbl_restore(x, y), y)
#> # A tibble: 2 x 6
#>   expression                min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>           <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 vectbl_restore(x, y)    5.8µs   6.93µs   123302.    3.97KB     24.7
#> 2 y                      23.1ns  28.17ns 23492926.        0B      0

^{Created on 2020-02-14 by the reprex package (v0.3.0)}

romainfrancois · 2020-02-14T11:57:02Z

For similar reasons, I've updated vec_tbl_restore() :

library(tibble)
library(rlang)

df1 <- tibble(x = rnorm(1e6), g = sample(rep(1:1e4, 100)))
df2 <- tibble(x = rnorm(1e6))

vectbl_restore <- tibble:::vectbl_restore
fast_nrow <- function(x) {
  .row_names_info(x, 2L)
}

attrs_names_only <- c("names", "row.names")
vectbl_restore_old <- function(xo, x) {
  n <- fast_nrow(xo)
  
  # FIXME: Use new_tibble()?
  forbidden <- attrs_names_only
  attrs <- attributes(x)
  attrs <- attrs[!(names(attrs) %in% forbidden)]
  
  attributes(xo)[names(attrs)] <- attrs
  
  # Need to patch manually
  attr(xo, "row.names") <- .set_row_names(n)
  
  xo
}

bench::mark(
  vectbl_restore(df1, df2), 
  vectbl_restore_old(df1, df2)
)
#> # A tibble: 2 x 6
#>   expression                        min median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                   <bch:tm> <bch:>     <dbl> <bch:byt>    <dbl>
#> 1 vectbl_restore(df1, df2)        929ns 1.27µs   753454.        0B      0  
#> 2 vectbl_restore_old(df1, df2)   3.97ms 4.81ms      209.    3.81MB     22.7

^{Created on 2020-02-14 by the reprex package (v0.3.0.9000)}

romainfrancois · 2020-02-14T11:57:52Z

closes tidyverse/dplyr#4855

krlmlr

Thanks for looking into it. If you revert 752bfa2 as part of this PR, the CI tests should pass.

krlmlr · 2020-02-14T12:08:04Z

Oh, it's only a problem with long data frames:

library(tibble)
library(rlang)

df1 <- tibble(x = rnorm(1e6), g = sample(rep(1:1e4, 100)))
df1a <- df1[1:10000, ]
df1b <- df1[1:1000, ]
df1c <- df1[1:100, ]

vectbl_restore <- tibble:::vectbl_restore

bench::mark(
  vectbl_restore(df1, df1),
  vectbl_restore(df1a, df1a),
  vectbl_restore(df1b, df1b),
  vectbl_restore(df1c, df1c),
  check = FALSE
)
#> # A tibble: 4 x 6
#>   expression                      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                 <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 vectbl_restore(df1, df1)     3.57ms   3.97ms      241.    3.81MB     23.4
#> 2 vectbl_restore(df1a, df1a)  40.44µs  43.49µs    20181.   39.11KB     16.4
#> 3 vectbl_restore(df1b, df1b)   9.54µs  11.04µs    75970.    3.95KB     15.2
#> 4 vectbl_restore(df1c, df1c)   6.44µs   7.39µs   118001.      448B     23.6

^{Created on 2020-02-14 by the reprex package (v0.3.0)}

Interesting threshold: going from 10k to 100k slows down by a factor of 100.

romainfrancois · 2020-02-14T12:11:10Z

yes @krlmlr I suppose this is row.names being inconveninentnly materialized when using attributes[]<-

krlmlr · 2020-02-14T12:15:45Z

If you revert 752bfa2 the CI errors should go away.

Applying [ to a tibble speeds up subsequent execution of vectbl_restore():

library(tibble)
library(rlang)

df1 <- tibble(x = rnorm(1e6), g = sample(rep(1:1e4, 100)))
df1a <- df1[1:100000, ]
df1b <- df1[1:10000, ]
df1c <- df1[1:1000, ]

vectbl_restore <- tibble:::vectbl_restore

bench::mark(
  vectbl_restore(df1, df1),
  vectbl_restore(df1a, df1a),
  vectbl_restore(df1b, df1b),
  vectbl_restore(df1c, df1c),
  check = FALSE
)
#> # A tibble: 4 x 6
#>   expression                      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                 <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 vectbl_restore(df1, df1)     3.54ms   4.05ms      232.    3.81MB     23.4
#> 2 vectbl_restore(df1a, df1a) 345.57µs 401.14µs     2235.  390.67KB     18.3
#> 3 vectbl_restore(df1b, df1b)  40.69µs  46.54µs    19021.   39.11KB     13.2
#> 4 vectbl_restore(df1c, df1c)   9.79µs  11.77µs    71793.    3.95KB     14.4

^{Created on 2020-02-14 by the reprex package (v0.3.0)}

I guess this no longer holds with the C implementation?

romainfrancois · 2020-02-14T12:25:38Z

library(tibble)
library(rlang)

df1 <- tibble(x = rnorm(1e6), g = sample(rep(1:1e4, 100)))
df1a <- df1[1:100000, ]
df1b <- df1[1:10000, ]
df1c <- df1[1:1000, ]

vectbl_restore <- tibble:::vectbl_restore

bench::mark(
  vectbl_restore(df1, df1),
  vectbl_restore(df1a, df1a),
  vectbl_restore(df1b, df1b),
  vectbl_restore(df1c, df1c),
  check = FALSE
)
#> # A tibble: 4 x 6
#>   expression                      min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                 <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 vectbl_restore(df1, df1)      947ns   1.09µs   629424.        0B      0  
#> 2 vectbl_restore(df1a, df1a)    932ns   1.06µs   635387.        0B      0  
#> 3 vectbl_restore(df1b, df1b)      1µs   1.23µs   516068.        0B      0  
#> 4 vectbl_restore(df1c, df1c)    949ns    1.1µs   782567.        0B     78.3

^{Created on 2020-02-14 by the reprex package (v0.3.0.9000)}

Co-Authored-By: Kirill Müller <krlmlr@users.noreply.github.com>

krlmlr · 2020-02-14T12:48:33Z

Thanks!

* 🧞tidyverse/tibble#717 * prototype vec_sprinkle() * require fallback for more types * handle names when sprinkling * need to vec_cast() all the chunks to common type * make all sprinkle steps internal with short_vec_init() and vec_cast() callables * internal recycling by repeatedly calling vec_assign_impl() with the sprinkling * rows_i_j needs PROTECT() * swing back to master tibble * globalVariables(o_rows) because it is defined lazily * rename to vec_unchop() and move arguments around * Apply suggestions from code review Co-Authored-By: Lionel Henry <lionel.hry@gmail.com> * Apply suggestions from code review Co-Authored-By: Lionel Henry <lionel.hry@gmail.com> * PROTECT(Rf_getAttrib()) * skip again for a different reason * intercept mutate() recycling problems earlier * using vctrs::vec_unchop() * using the callable vctrs::short_vec_recycle() so that chunks are recycled to fit * No need to hang on to rows_lengths * - simplify DataMask constructor - provide DataMask$get_rows() * Move size handling until after all the expressions. use vec_unchop() then * recycling and then call vec_c() instead of using vec_unchop() * use master vctrs * vec_assign_impl() no longer used * shirt_vec_init() no longer used * vec_cast() not used intrenally Co-authored-by: Lionel Henry <lionel.hry@gmail.com>

romainfrancois requested review from krlmlr and lionel- February 13, 2020 15:34

lionel- reviewed Feb 13, 2020

View reviewed changes

romainfrancois added a commit to tidyverse/dplyr that referenced this pull request Feb 13, 2020

🧞 ♂️ tidyverse/tibble#717

71de6eb

romainfrancois mentioned this pull request Feb 13, 2020

mutate() performance improvements tidyverse/dplyr#4858

Merged

romainfrancois mentioned this pull request Feb 14, 2020

performance of dplyr_col_modify() tidyverse/dplyr#4855

Closed

romainfrancois added 5 commits February 14, 2020 12:59

faster internal version of update_tibble_attrs()

2b0bc78

needs to shallow duplicate first

a9f452a

unrelated changes in tests outputs

ea1d28f

using x != R_NilValue

11840e7

internal version of vectbl_restore()

8ca9312

romainfrancois force-pushed the update_tibble_attrs branch from 5aec45a to 8ca9312 Compare February 14, 2020 12:00

krlmlr reviewed Feb 14, 2020

View reviewed changes

Comment thread src/attributes.c Outdated

Comment thread src/attributes.c Outdated

This comment has been minimized.

Sign in to view

depend on dev vctrs and rlang

de0efc9

romainfrancois and others added 2 commits February 14, 2020 13:26

Update src/attributes.c

3a54c4f

Co-Authored-By: Kirill Müller <krlmlr@users.noreply.github.com>

remove unnecessary handling of row.names attribute

82eccba

krlmlr merged commit cba28e9 into tidyverse:master Feb 14, 2020

romainfrancois deleted the update_tibble_attrs branch February 14, 2020 13:37

romainfrancois added a commit to tidyverse/dplyr that referenced this pull request Feb 21, 2020

🧞 ♂️ tidyverse/tibble#717

91acb09

github-actions Bot locked as resolved and limited conversation to collaborators Feb 14, 2021

Conversation

romainfrancois commented Feb 13, 2020 • edited by krlmlr Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lionel- Feb 13, 2020

Choose a reason for hiding this comment

Uh oh!

romainfrancois Feb 13, 2020

Choose a reason for hiding this comment

Uh oh!

lionel- Feb 13, 2020

Choose a reason for hiding this comment

Uh oh!

Uh oh!

krlmlr commented Feb 14, 2020

Uh oh!

romainfrancois commented Feb 14, 2020

Uh oh!

romainfrancois commented Feb 14, 2020

Uh oh!

romainfrancois commented Feb 14, 2020

Uh oh!

krlmlr commented Feb 14, 2020

Uh oh!

romainfrancois commented Feb 14, 2020

Uh oh!

romainfrancois commented Feb 14, 2020

Uh oh!

krlmlr left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

krlmlr commented Feb 14, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

romainfrancois commented Feb 14, 2020

Uh oh!

This comment has been minimized.

krlmlr commented Feb 14, 2020

Uh oh!

romainfrancois commented Feb 14, 2020 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

krlmlr commented Feb 14, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

romainfrancois commented Feb 13, 2020 •

edited by krlmlr

Loading

krlmlr commented Feb 14, 2020 •

edited

Loading

romainfrancois commented Feb 14, 2020 •

edited

Loading