Subject: Re: [geometry] partial_sort vs nth_element in the rtree implementation
From: Adam Wulkiewicz (adam.wulkiewicz_at_[hidden])
Date: 2014-02-28 08:08:26
Hi,
IMPORTANT!
Boost was moved to GitHub some time ago. SVN trunk has become GIT
develop branch. You can browse it here:
https://github.com/boostorg/geometry/tree/develop.
If you'd like to switch to GIT, here's some tutorial:
https://svn.boost.org/trac/boost/wiki/TryModBoost
So to use Boost.Geometry develop branch with Boost you need to replace
the last commends mentioned on the page above to something like this:
cdmodular-boost/libs/geometry
git checkout develop
git pull
and probably again run
./b2 headers
but, back to the subject.
Lu Wang wrote:
> Hi all,
>
> I'm now reading the source code of the rtree implementation in
> boost::geometry, especially in pack_create.hpp
> <http://svn.boost.org/svn/boost/trunk/boost/geometry/index/detail/rtree/pack_create.hpp>,
> I found that std::partial_sort is used to find the median along the
> longest dimension, which is then used to split the array and perform
> the splitting recursively:
>
> // pack_create.hpp, around line 281
> partial_sort(first, median, last, ...);
>
> per_level_packets(first, median, ...);
> per_level_packets(median, last, ...);
>
>
>
> It seems to me that the sortedness of the first half is never used,
> in this case, I guess that std::partial_sort may be replaced by
> std::nth_element, which should be faster.
>
Thanks for catching this! Tests passed, code pushed.
> Is there any particular reason that partial_sort is used?
No, I must have been under the mental influence of the R* ;).
The speedup is nice, ~4x faster for 1M random boxes.
I remember that someone on the list raised the issue of the small
preformance difference between linear split and packing algorithm some
time ago. Please recheck.
Regards,
Adam