Composing tiles from the GeoJSON features is an expensive
operation that might block UI thread on updating the `GeoJsonSource`
with the new data.
This change moves tile composing to the background thread and thus
unblocks the UI thread.
This moves the LineAtlas from a shared texture that contained SDF dash patterns to use individual textures.
Previously, the texture space was limited to a texture of 512 pixels height. Dash patterns were never removed (and are still never removed as of this patch), which means that this texture could fill up for styles that use a lot of different dash patterns. In particular, dash patterns for lines with a round line cap take up 15 pixels of texture height, limiting the amount of unique dash patterns to 34. While this was probably enough for rendering a single style, we quickly exhausted this number when reusing the Map object to render different styles.
Instead of a global shared texture, we're now creating individual textures for every dash pattern. These textures are still cached so that we don't need to re-upload the texture on every frame.
This fixes rendering by account for the 1px texture padding around icons that were stretched with icon-text-fit. We've added the 1px padding before, but didn't scale it accordingly when we are resizing the icon when it is stretched to fit the text. Adjusts the code to match the logic in GL JS.
Unlike icons, pattern changes are not caught with
style-diff meaning that the existing request could
be from the previous style and we cannot coalesce
requests for them.
This commit refactors `utils::ThreadPool` into a template
`ThreadedScheduler` class and provides aux type aliases.
So that it is possible to obtain a sequenced schedule,where
all the scheduled tasks are guarantied to be executed
consequently.
The sequenced lightweight scheduler is required by both the
orchestration thread and the refactored `FileSource` implementation.
RenderOrchestrator holds a vector of ordered rendered layers,
in order to minimize calling of `getRenderLayer()`, which
could be quite expensive as it calculates hash of the layer Id.
This change improves orchestration performance in particular
for the styles that have multiple sources.
The `API_renderStill_multiple_sources` benchmark runs
approximately 70% faster with the changes applied.
