Search universes with many cells with a bounding box tree

[GuySten]

Description

This PR is stacked on the region tree PR (#4167) and on #4163. Only the last two commits are new.

Finding the cell containing a point checks the cells of a universe in order. With #4163, cells whose bounding boxes do not contain the point are skipped. In a universe with many cells, such as the layers of many TRISO particles placed directly in a matrix, testing every bounding box is itself the cost. Neighbor lists don't help either: the matrix's list grows to hold most of the cells.

With this PR, a universe whose cells have finite bounding boxes that rarely overlap is searched with a bounding volume hierarchy over the boxes. It uses the same tree, and the same rule, as the region tree PR: at least 32 boxes, and a point in at most an eighth of them on average.

• Same cell found: of the cells whose boxes contain the point, the first in the universe that contains the point is returned, as in a search of all of the cells.
• Neighbor lists: in such a universe they hold at most 8 cells. A cell with few neighbors is still found from its neighbor list. A cell next to many others, such as the matrix, fills its list, and searches from it then go to the tree directly.

The rule for using a tree moves to box_tree so that regions and universes share it, and the geometry methods documentation describes the search.

Results

Transport time, single thread, median of interleaved runs (3 to 6 per build), compared with the region tree PR (#4167). Results are bit-identical on every model.

Model	Region tree PR	This PR	Ratio
10 cm slice of a flat TRISO compact	295 s	7.0 s	0.024
Flat TRISO	88.5 s	13.3 s	0.15
ATR (5,968 cells), volume calculation	3.92 s	0.98 s	0.25
ATR, photon transport	12.5 s	7.1 s	0.57
ATR	7.27 s	5.75 s	0.79
17 other models, including a ring of 360 sectors whose boxes overlap (kept linear by the rule)			0.92–1.03

Flat models no longer need lattices to be fast. On the same machine:

Model	Flat	With a TRISO lattice
Flat TRISO on develop	359 s	28.3 s
Flat TRISO with this PR	13.3 s	18.2 s
10 cm compact slice with this PR	7.0 s	9.2 s

With delta tracking (#3971 merged with this PR):

Model	#3971	#3971 + this PR
10 cm slice of a flat TRISO compact	87.6 s	1.47 s
Same slice built with a lattice	1.76 s	1.60 s
ATR	33.9 s	6.9 s

Adversarial model checked against develop:

• Flat 10 × 10 grid of box cells with long histories: the neighbor lists of develop find each new cell in one or two checks. 0.90 times the time of develop. Before neighbor lists were kept in tree universes, this model was 1.2 times slower.

Testing

• test_cell_search.py (from Skip cells whose bounding boxes do not contain the point in cell searches #4163) now runs through the tree: cells found for random points agree with Python's evaluation of the regions.
• A new test puts three overlapping cells, one of them with an infinite box, in a tree universe in four orders. In each order, the first of them in the universe is found.
• The full test suite passes locally. The only errors are tests that need ENDF data or dlopen, which fail the same way on develop.

Checklist

• I have performed a self-review of my own code
• I have run clang-format (version 18) on any C++ source files (if applicable)
• I have followed the style guidelines for Python source files (if applicable)
• I have made corresponding changes to the documentation (if applicable)
• I have added tests that prove my fix is effective or that my feature works (if applicable)

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}