add explanation about dispatch heuristic

jinsolp · jinsolp · commit df3b5775db63 · 2025-11-20T02:20:52.000Z
diff --git a/cpp/src/hdbscan/detail/condense.cuh b/cpp/src/hdbscan/detail/condense.cuh
@@ -47,7 +47,11 @@ value_t get_lambda_host(value_idx node, value_idx num_points, const value_t* del
 }
 
 /* Heuristic dispatching to bottom-up implementation. A high persistent_ratio means there are more
-chances to stop early as we climb up the tree, making it more efficient for bottom-up approach*/
+chances to stop early as we climb up the tree, making it more efficient for bottom-up approach
+Note: These thresholds are based on empirical observations from running the algorithm on
+datasets ranging from 30K to 5M points, with varying distributions and thread counts. For more
+details, please refer to the tables in this PR: https://github.com/rapidsai/cuml/pull/7459
+*/
 bool dispatch_to_bottom_up(int num_persistent, int n_leaves)
 {
   int n_nodes            = 2 * n_leaves - 1;

Original file line number	Diff line number	Diff line change
`@@ -47,7 +47,11 @@ value_t get_lambda_host(value_idx node, value_idx num_points, const value_t* del`
`47`	`47`	`}`
`48`	`48`
`49`	`49`	`/* Heuristic dispatching to bottom-up implementation. A high persistent_ratio means there are more`
`50`		`-chances to stop early as we climb up the tree, making it more efficient for bottom-up approach*/`
	`50`	`+chances to stop early as we climb up the tree, making it more efficient for bottom-up approach`
	`51`	`+Note: These thresholds are based on empirical observations from running the algorithm on`
	`52`	`+datasets ranging from 30K to 5M points, with varying distributions and thread counts. For more`
	`53`	`+details, please refer to the tables in this PR: https://github.com/rapidsai/cuml/pull/7459`
	`54`	`+*/`
`51`	`55`	`bool dispatch_to_bottom_up(int num_persistent, int n_leaves)`
`52`	`56`	`{`
`53`	`57`	`int n_nodes = 2 * n_leaves - 1;`