Re-introduce #10444: Don't load unnecessary tiles on terrain

[astojilj]

The commit includes original #10444 and #10446 with fix for issue reported in #10462.

The issue in #10462 is about assuming center elevation for parent nodes in tile pyramid: the approach needs to be applied only to final tile cover tiles.
This is because parent tiles don't have corresponding DEM tiles cover and could get clipped out "too early" from cover.

Fixes: #10463

The case #10462 is good for illustrating the issue with original PR and improvement compared to main. Following patch is used to log tiles loaded:

--- a/src/source/source_cache.js
+++ b/src/source/source_cache.js
@@ -738,6 +738,7 @@ class SourceCache extends Evented {
         const cached = Boolean(tile);
         if (!cached) {
             tile = new Tile(tileID, this._source.tileSize * tileID.overscaleFactor(), this.transform.tileZoom);
+            console.log(`${this.id} : ${tileID.toString()}`);
             this._loadTile(tile, this._tileLoaded.bind(this, tile, tileID.key, tile.state));
         }

This view uses 37 vector tiles and only those are loaded while opening screen.
Code in main branch loads 71 vector tiles, uses 37 of those once view is loaded.

TODO:

• test for Revert 10444 + 10446 #10462,

Considered refactoring getMinElevationBelowMSL() to MinMax and cache it for e.g. getBounds3D could also reuse it. However, for this case, only getMinElevationBelowMSL() is needed and performance impact is not such that it would require caching. getBounds3D is called only on getBounds() or query API calls. Caching would be orthogonal to the work here so leaving the PR as it is.

[astojilj]

Considered refactoring getMinElevationBelowMSL() to MinMax and cache it for e.g. getBounds3D could also reuse it. However, for this case, only getMinElevationBelowMSL() is needed and performance impact is not such that it would require caching. getBounds3D is called only on getBounds() or query API calls. Caching would be orthogonal to the work here so leaving the PR as it is.

PTAL

[karimnaaji]

LGTM

[karimnaaji]

Something to note is when terrain._visibleDemTiles == 0, we may hit a corner case: when there are tiles fully within negative elevation we might miss tiles, because they don't intersect with the frustum (so they are not accounted for in getMinElevationBelowMSL). One approach would be to use with an overly larger frustum for tile cover, to ensure they are marked as visible at least once. It's not directly related to this PR (same happens in main).

EDIT: Captured in #10493