Turns out there are sometimes errors that will not prevent the
shader from functioning normally. As such, it is not possible
to simply throw when there are errors/warnings.
Since we are not logging to the console, these errors will have to stay hidden
unless the user requests them explicitly.
The hit detection render is now done against a canvas with half the
width/height of the main render. This still provides sufficient precision
while requiring a much smaller memory allocation (especially for
retina devices).
Various references were kept, preventing the layer and underlying
renderer and webgl context to be garbage collected.
Also, the Helper was simplified because it turns out deleting manually
all Webgl objects is useless: these objects will be released when
the context is garbage collected anyway.
Note: this touches the Layer and BaseLayer classes, as the following were
preventing the layer from being garbage collected:
* layer reference in the `state_` object in BaseLayer
* dangling listener for source change in Layer
Due to the fact that the points renderer does not know for sure
which attributes will be used for rendering, it is now mandatory
to provide both vertex/fragment shaders for rendering AND hit
detection.
The hit detection shaders should expect having an `a_hitColor`
that they should return to allow matching the feature uid.
This will be all one eventually by shader builders under the hood.
Now most attributes will be custom ones defined by the user of the renderer.
The hit detection is broken for now and still has to be fixed.
Also it is not possible anymore to give a texture to the renderer,
this will have to be fixed as well.
For each feature its opacity value index is encoded on 4 bytes
in the color values, and the uid is stored in the opacity
value, allowing for a much higher range of uids to be read.
For now only `forEachFeatureAtCoordinate` is implemented.
Each time the viewport is rendered, another similar render pass is
done using the specific hit detection instructions. Feature uid's are
encoded in the r,g,b,a channels and can then be decoded on the fly.
Note: the `readPixels` operation is taking a lot of time,
around 10-20ms each frame.
Hit detection is done by rendering features with their id encoded in the
color attribute. A parallel set of render instructions and a second
vertex buffer is used specifically for that.
The worker currently works by receiving GENERATE_BUFFERS messages and
will send back the same kind of message, with the generated buffers
attached. All properties of the original message are kept, so that
when a GENERATE_BUFFERS message comes back to the main thread it
is possible to know what and how the buffers where generated.
This is typically used for the `projectionTransform` matrix, and
will also be necessary when working with tiles.
The worker receives a transferable array of instructions
and sends back two transferable arrays (vertex and index buffer).
The projection transform is also sent so that when the main thread
receives the buffers from the worker it also knows which projection to
apply when rendering the geometries.
The base webgl renderer module now has two types of utilities:
* `writeXFeatureInstructions` will write a series of values in a given
typed array, which represent how a given feature will be rendered; for points,
this means position, size, color, etc.
* `writeXFeatureToBuffers` will fill up the given index & vertex buffers
with values based on the provided render instructions
As such, the logic for rendering features is:
user-input style > instructions array >(*) index/vertex buffers > draw
(*) this transformation is intended to be done on a worker.
