Make efficient C++ versions of window functions

[hadley]

Loosely grouped below

• lag and lead
• nth_value, first_value, last_value
• row_number, rank, dense_rank
• percent_rank, cume_dist, ntile
• cumsum, cummin, cummax

[romainfrancois]

I've put some code for first (last will follow). I don't have the full argument matching at my disposal, so I'm handling the arguments this way:

• the first is always considered to be the variable. I just check that the argument is either unnamed or called 'x'
• for the other arguments, they must be named. I don't do position matching if not, but partial matching is allowed.

That's a bit more code than I anticipated for the implementation of something conceptually simple. The advantage over calling the R version of first is that there is no materialisation of the data. i.e:

> df <- data.frame( x = 1:16, g = rep(1:4, each = 4), y = 16:1 )
> df
    x g  y
1   1 1 16
2   2 1 15
3   3 1 14
4   4 1 13
5   5 2 12
6   6 2 11
7   7 2 10
8   8 2  9
9   9 3  8
10 10 3  7
11 11 3  6
12 12 3  5
13 13 4  4
14 14 4  3
15 15 4  2
16 16 4  1
> df %.% group_by(g) %.% summarise( first_x = first(x) )
Source: local data frame [4 x 2]

  g first_x
1 1       1
2 2       5
3 3       9
4 4      13
> df %.% group_by(g) %.% summarise( first_x = first(x, order_by = y) )
Source: local data frame [4 x 2]

  g first_x
1 1       4
2 2       8
3 3      12
4 4      16

In each case, I don't have to materialise the 4 vectors to pass them to the R function first. n the easy case first(x) I just pick the first one from the data virtually indexed by the indices.

For the case first(x, order_by = y) I loop around the y variable to find the smallest, but at no point I am materialising either x or y.

last should be straightforward, and nth should not be too hard.

[hadley]

Nice, thanks Romain.

[romainfrancois]

The code for lead and lag was already there but not enabled. I've fixed it a bit and enabled. Although for now it does not support the full arguments of their R counterparts. They only handle the 2 args form, i.e. they don't handle default or order_by. If the call is anything more than 2 args, it just falls back to R eval.

I think I can handle default easily.

Then for order_by I can borrow some code from first ...

I'll handle the case order_by = symbol hybridly as it's the main use case I guess, and then anything else will fall back to R as it's more complicated to handle.

[hadley]

Is this finished?