forder3

[mattdowle]

Slow down in dev ordering all-unique input fixed; thanks to @arunsrinivasan's testing.

N = 1e8
set.seed(1)
x = sample(N)                       # v1.11.8  master     this PR
system.time(o <- forderv(x))        # 4.5s     5.9s-7.0s  1.2s
!base::is.unsorted(x[o])  # TRUE

x = sample(N/10, N, replace=TRUE)   
system.time(o <- forderv(x))        # 2.6s     1.3s       0.8s

Long-standing flip-flop group sizes removed; now recursive and writes directly to final group size stack via thread buffers.
Nested parallelism recently added, removed. skew still handled.
timetaken() in verbose mode now displays both elapsed and cpu time

Looks like AppVeyor timings on this PR are back to normal at about 5 mins (were over 10 min on master) so that's a good sign.

Needs to be in master to be picked up by db-bench as I've only tested on laptop size (N=1e8).

[codecov]

Codecov Report

Merging #3162 into master will increase coverage by 0.13%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #3162      +/-   ##
==========================================
+ Coverage   92.05%   92.18%   +0.13%     
==========================================
  Files          61       61              
  Lines       11450    11582     +132     
==========================================
+ Hits        10540    10677     +137     
+ Misses        910      905       -5

Impacted Files	Coverage Δ
R/data.table.R	92.74% <ø> (ø)	⬆️
src/forder.c	99.71% <100%> (+0.77%)	⬆️
R/timetaken.R	100% <100%> (+9.09%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 55168ed...de2855a. Read the comment docs.

[jangorecki]

those are incredible speed ups! 3-4 times reduced vs. 1.11.8

[mattdowle]

@jangorecki yep fingers crossed. needs battle testing. could you set off a db-bench rerun please (just data.table if possible) to see how it does on server as I missed the 11pm start time. I'm hoping this will fix the oddness of the 2nd run being slower for the first test, too.