perf: diagnose the four open benchmark findings (lu_solve, the new matmath() sites, gaussian kernel)

[kalwalt]

Summary

The #86 benchmark suite produced several findings that survived investigation of the others. Four remain open on merged dev, measured over four full-suite runs on an idle machine (see the Current status table in bench/README.md).

This issue tracks diagnosing them. Unlike #86 — which is measurement infrastructure and explicitly rules out micro-optimising — work here may touch src/, with the parity suite green at every step.

The open findings

case	4 runs	notes
linalg.lu_solve (6x6)	1.43 / 1.49 / 1.46 / 1.47	The tightest signal in the suite — ±0.03, tighter than YAPE ever was. Never profiled.
homography2d.check_subset	1.35 / 1.27 / 1.28 / 1.63	One of three new matmath() sites
affine2d.run — 3 points	1.34 / 1.31 / 1.30 / 1.41	Same cause; worst when the real work is smallest
math.get_gaussian_kernel — size 7	1.23 / 1.21 / 1.16 / 1.15	4/4 jsfeat, at or just above the floor

lu_solve — the one to start with

#159 predicted in advance that its fix could not explain this one, because lu_solve constructs no internal matrix_t. That prediction held: every other linalg case dropped below the floor after #159, and lu_solve did not move.

One misattribution has already been ruled out: lu_solve reads JSFEAT_CONSTANTS.EPSILON inside a loop, which looks like the bug fixed in #165 — but original jsfeat reads jsfeat.EPSILON, also a property load. Both sides pay it, so it explains nothing here.

The three new matmath() sites

motion_model.ts lines 212 (affine2d.run), 352 (homography2d.run) and 541 (homography2d.check_subset) each construct a matmath per call; original jsfeat calls module-scope functions instead. #159 made each construction much cheaper — affine2d.run at 3 points fell from ~3.5x to ~1.35x — but did not remove them.

The shape of the remaining gap still fits a fixed per-call cost: affine2d.run is worst at 3 points, where the real arithmetic is smallest, and much better at 40.

get_gaussian_kernel size 7

Newly consistent rather than newly appeared: earlier series had it flipping sign, the merged-dev series has it 4/4 at 1.15–1.23. Lowest priority of the four, and possibly still floor noise — confirm it is real before investigating.

Also worth recording: an unexplained improvement

yape06 measured 1.14–1.47 on #166's branch, where that PR correctly reported the alias had not helped it. On merged dev it measures ~1.04 across four runs. The improvement is real but not attributable to any single one of #159/#165/#166. Not a defect, but if someone later needs to explain why yape06 is fast now, this is the open question.

Approach that worked before

#159 and #165 both came from the same loop: profile the workload the bench exercises, aggregate self-time by call frame, then isolate the suspected cost with a throwaway probe that changes exactly one thing. #165's probe compared five variants of one function; #166's replicated a whole function body under four structural shapes, asserting identical outputs before timing.

Two lessons from those, worth repeating:

• Verify the hypothesis before fixing. perf(yape): alias imported helpers to module-scope consts #166's documented theory (cross-module imports blocking inlining) was disproven by the profile.
• A control case is worth more than a big number. In motion_model, error() — the only method that allocated nothing — staying at noise while everything else moved was the strongest evidence in that investigation.

Acceptance criteria

• Each of the four either has a documented cause, or a recorded "profiled, cause not found".
• Any src/ change keeps npm test green and is measured before/after on an idle machine.
• bench/README.md's Current status table updated — it is the single place that tracks this.
• Disproven hypotheses recorded as disproven, not silently dropped.

Related

• Parent: Add a performance-benchmarking suite (per module/submodule) with result storage & regression comparison #86
• Resolved by the same effort: matrix_t re-allocates a data_type lookup table on every construction #159, perf(matmath): hoist EPSILON out of invert_3x3's per-call path #165, perf(yape): alias imported helpers to module-scope consts #166
• Coverage gap tracked separately: bench: finish imgproc coverage (12 unbenched methods) and pyramid_t.build #168