Merge branch 'stafah/centroid' into 'main'

Use centroids for SAH partition (GH-1102)

See merge request omniverse/warp!1802

Author: shi-eric

CHANGELOG.md

@@ -19,6 +19,7 @@
 - Add more examples to `Tiles and SIMT code` documentation, demonstrating caveats when switching between the CPU and GPU and using `wp.tile()` ([GH-1042](https://github.com/NVIDIA/warp/issues/1042)).
 - Add support for `int64` and `uint64` key types to `wp.tile_sort()` ([GH-1089](https://github.com/NVIDIA/warp/issues/1089)).
 - Add `wp.tile_scan_max_inclusive()` and `wp.tile_scan_min_inclusive()` for cumulative maximum and minimum operations across tiles ([GH-1090](https://github.com/NVIDIA/warp/issues/1090)).
+- Changed BVH SAH constructor to use centroids instead of bounds to determine partition axis, improving build quality and traversal performance in some scenarios ([GH-1102](https://github.com/NVIDIA/warp/issues/1102)).
 
 ### Removed
 

warp/native/bvh.cpp

@@ -392,15 +392,20 @@ float TopDownBVHBuilder::partition_sah_indices(const vec3* lowers, const vec3* u
     assert(end - start >= 2);
 
     int n = end - start;
-    vec3 edges = range_bounds.edges();
-
-    bounds3 b = calc_bounds(lowers, uppers, indices, start, end);
 
+    bounds3 centroid_bounds;
+    for (int i = start; i < end; ++i)
+    {
+        vec3 item_center = 0.5f * (lowers[indices[i]] + uppers[indices[i]]);
+        centroid_bounds.add_point(item_center);
+    }
+    vec3 edges = centroid_bounds.edges();
+    
     split_axis = longest_axis(edges);
 
     // compute each bucket
-    float range_start = b.lower[split_axis];
-    float range_end = b.upper[split_axis];
+    float range_start = centroid_bounds.lower[split_axis];
+    float range_end = centroid_bounds.upper[split_axis];
 
     // Guard against zero extent along the split axis to avoid division-by-zero
     if (range_end <= range_start)