[Qgis-developer] Multiple labeling layers ?

[Hugo Mercier]

Le 12/02/2015 15:19, Martin Dobias a écrit :
> 
> On Thu, Feb 12, 2015 at 8:50 PM, Hugo Mercier <hugo.mercier at oslandia.com
> <mailto:hugo.mercier at oslandia.com>> wrote:
> 
>     >
>     >
>     > Sounds good. Do I understand correctly you want to avoid having to
>     > register features on every map rendering?
>     >
>     > What exactly would be cached - features / label candidates? for one view
>     > / for the whole map?
> 
>     If I am correct, during labeling the most demanding task is the global
>     optimization to compute label positions. So these label positions could
>     be cached for one set of labels (in a labeling layer), provided the
>     underlying vector layers are also cached.
> 
> 
> Actually not - the global optimization is quite fast (after all that was
> the intent of the PAL library), usually it is extraction of label
> candidates and calculation of their costs (for polygons) that make up
> most of the time.

I see. I thought this part was also done by PAL. I understand that
computing position of candidate bounding boxes (evaluating expressions,
include font metrics, centroids, etc.) may be the hardest part.

> 
> In some cases finding of the solution can still be a bottleneck - e.g.
> zoomed out map with huge amount of overlapping labels.
> 
> I have done measuring of the labeling performance quite a long time ago,
> so it may be a good idea to check again where the speed issues are.
> Various placement options have their own issues.

Yes. So caching results from registering and PAL may be a good option.

> 
>  
> 
>     I am also investigating another optimization idea: the good old global
>     labeling on top of every other layers, but with a way to "partially"
>     solve the placement problem. Could it be possible to inject precomputed
>     label positions in PAL ?
> 
> 
> Yes, this is already being done with data-defined positions of labels.

Ah good to know.
I guess it is the fixedPos and fixedAngle parameters of
pal::Layer::registerFeature ?

>  
> 
>     So that a part of the label positions would be
>     already known from a previous frame and some other labels would have to
>     be recomputed ?
> 
> 
> It is hard to guess what would help the most. Probably the best would be
> to first measure the times of few scenarios and decide what/where/how to
> cache based on the hard data.

Sure.