Subject: Re: [geometry] Adapters at the point and/or box abstraction level?
From: Barend Gehrels (barend_at_[hidden])
Date: 2014-05-28 04:50:38
Patrick J. LoPresti wrote On 27-5-2014 23:13:
> I have an existing library of code dealing with 2D points, boxes,
> etc., and I am looking for a good spatial index. I believe
> boost::geometry::rtree might fit the bill.
> My library uses SSE registers to represent 2D points/vectors. This
> makes the operations I care about very fast (e.g. vector arithmetic,
> box union/intersection), but accessing the components of a point is a
> bit painful.
> Is there any way to teach boost::geometry -- and especially
> boost::geometry::rtree -- to use my functions for these operations,
> avoiding get<0> and get<1> as much as possible?
get<0> and get<1> are *the* methods to teach Boost.Geometry to access
They are all templated and inlined so it should have no speed penalty
(on "normal" C++ applications). But I don't have experience with that in
combination with SSE.
If you want to optimize one operation further (say: box intersection),
then you can make a specialization (of dispatch) for your specific box
type. That is not often done, but it should be possible. The
box-intersection can then make use of SSE (while other things still work
the "Boost.Geometry" way).
This must be done per algorithm / geometry type.
Most functions are modelled like that (using dispatch). But vector
arithmetic is currently not modelled like that - if you continue with
this, and it works fine, we can consider adding a dispatch layer there too.
> (I have read the documentation but not yet the source code, so please
> forgive me if this is a stupid question.)
It is a very interesting question ;-)
Geometry list run by mateusz at loskot.net