We were capturing the reference to an object that could go away when the request was canceled. However, the lambda could still be invoked by NSURLSession, creating a race condition with a resulting crash when the response to a tile request was 404 Not Found.
ProjectedCollisionBox encapsulates geometry of the projected
collision box, it is using union and thus provides memory save
- 12 bytes per collision box instance.
Before this change, the `RenderTileSetSource` implementation
ignored update calls for the sources whose description was not yet
loaded and it lead to missing of relayout requests.
Patch partly fixes#15163 in a way that it doesn't allow loading tens of thousands of tiles and attempt to show area above horizon:
Limit pitch based on edge insets. It is not too bad - current limit of 60 degrees stays active until center of perspective is moved towards the bottom, to 84% of screen height. The plan is to split removal of 60 degrees limit to follow up patch.
Fix max Z calculation in getProjMatrix. TransformState::getProjMatrix calculation of farZ was complex with possibility to lead to negative z values. Replacing it with simpler, precise calculation:
furthestDistance = cameraToCenterDistance / (1 - tanFovAboveCenter * std::tan(getPitch()));
TransformState::getProjMatrix calculation of farZ was an aproximation. Replacing it with simpler, but precise calculation.
Related to: #15163
Camera/Render mode updates can be streamed from various different stateless sources and we should make sure to skip an update to a mode that's already set. This keeps the callbacks tidy and does not perform unnecessary work.
When all location layer animations were reset, due to the mode change, we were recalculating the time of the animation, while we should be taking the remaining duration. Frequent recalculation requests resulted in a really short animation duration.
Before this change, `RenderOrchestrator::filteredLayersForSource` shrank inside
`createRenderTree()`, which potentially might have affected rendering performance.
This PR filters out invisible layers before quering features
(rather than after, like it happened before).
With this change `API_queryRenderedFeaturesAll` performance test
passes almost two times faster.
