Performance Optimizations

[John-194]

Performance Optimizations

• Improved performance 1.7x - 2.9x
• Fixes high idle CPU usage

Tested using:

• 3K 3D — 3k-point 3D cloud
• 13K 2D — 13k-point 2D cloud
• 100K 2D — 100k-point 2D cloud

The baseline and these optimizations return the same DBSCAN results.
Memory leaks were not detected.

Results vs. baseline (dev branch)

Speed (ms/iter)

Commit	Description	3K 3D	13K 2D	100K 2D
5c7fdd5	baseline	0.86	1.29	9.38
39805c2	parfor heuristic	0.38	0.76	8.86
8da5e33	260% idle CPU usage fix	0.40	0.79	8.88
96c9cff	pre-compute cell keys	0.40	0.79	7.71
4a4025f	defer cell/mutex init	0.41	0.79	6.27
607417e	O(n) cluster renumbering	0.43	0.66	5.51
847e798	eliminate sqrt	0.32	0.64	5.41
909a621	dynamic thread count	0.30	0.65	5.45

Speedup per step (vs. previous commit)

Commit	Description	3K 3D	13K 2D	100K 2D
5c7fdd5	baseline	-	-	-
39805c2	parfor heuristic	2.26x	1.70x	1.06x
8da5e33	260% idle CPU usage fix	0.95x	0.96x	1.00x
96c9cff	pre-compute cell keys	1.00x	1.00x	1.15x
4a4025f	defer cell/mutex init	0.98x	1.00x	1.23x
607417e	O(n) cluster renumbering	0.95x	1.20x	1.14x
847e798	eliminate sqrt	1.34x	1.03x	1.02x
909a621	dynamic thread count	1.07x	0.98x	0.99x

Cumulative speedup (vs. baseline)

Commit	Description	3K 3D	13K 2D	100K 2D
5c7fdd5	baseline	-	-	-
39805c2	parfor heuristic	2.26x	1.70x	1.06x
8da5e33	260% idle CPU usage fix	2.15x	1.63x	1.06x
96c9cff	pre-compute cell keys	2.15x	1.63x	1.22x
4a4025f	defer cell/mutex init	2.10x	1.63x	1.50x
607417e	O(n) cluster renumbering	2.00x	1.95x	1.70x
847e798	eliminate sqrt	2.69x	2.02x	1.73x
909a621	dynamic thread count	2.87x	1.98x	1.72x

Summary of changes

1. 39805c2 — Replace timing-based granularity with simple heuristic in parfor
Removed get_granularity() which ran a timed benchmark to decide chunk sizes at runtime. Replaced with a simple n / (4 * num_threads) heuristic plus a sequential fast-path for small ranges. Biggest win on the small 3K cloud (2.26x).

2. 8da5e33 — Fix idle CPU usage via progressive backoff and CV-based sleep
Rewrote the work-stealing loop in scheduler.h with a 3-phase backoff: aggressive spinning, yielding, then sleeping on a condition variable. spawn() only signals the CV when threads are actually sleeping (zero-cost during active computation). Also changed wait() to opportunistically execute stolen work instead of re-entering start(), preventing potential deadlocks. Neutral on performance but eliminates CPU burn on idle threads.

3. 96c9cff — Pre-compute cell keys for grid sort, avoiding floor()
In grid.h, pre-computes integer cell coordinates (floor((P[i][d] - pMin[d]) / r)) into a flat array before sorting, so the sort comparator uses integer comparisons instead of repeated floating-point floor().

4. 4a4025f — Defer grid cell/mutex init to used cells only, reduce hash table size
Moves std::mutex construction and nbrCache/cells initialization from all cellCapacity slots to only the numCells actually used after insertion. Also shrinks the initial hash table from cellMax*2 to max(2048, cellMax/4) with a rebuild if the estimate is too small.

5. 607417e — O(n) cluster renumbering via prefix sum
Replaced an O(n log n) sampleSort + hash-table lookup for remapping cluster IDs with a simple O(n) approach: mark which IDs are used in a flag array, prefix-sum to get sequential IDs, then remap. Eliminates the myPair struct, hashSimplePair, and Table allocation entirely.

6. 847e798 — Eliminate sqrt from distance comparisons
Added nodeDistanceSqr() to kdNode.h and switched all distance comparisons in coreBccp.h, kdNode.h, and kdTree.h to use the existing distSqr() and the new nodeDistanceSqr() instead of dist()/nodeDistance(). The BCCP threshold r <= epsilon becomes r <= epsilon * epsilon. Uses ternary operators instead of std::max for branchless codegen hints. 1.34x on 3K.

7. 909a621 — Dynamic thread count via getWorkers()
Replaced hardcoded static const intT P = 36*8 with getWorkers() * 8 in boundingBoxParallel() and pMinParallel(). Also switched VLA-style stack arrays to heap-allocated (newA + free) since the count is now dynamic. Portable across machines with different core counts.