boxblur: replace the per-pixel division with an exact reciprocal

[georgmartius]

Both boxblur passes end in acc/size. size is a runtime value, so that compiled to a real integer division — ~20-30 cycles inside boxblur_hori's serial running sum, and in boxblur_vert it kept the output loop from vectorizing at all, since no SIMD unit has an integer divide.

This was listed under "Known limits" in docs/simd.md, with the claim that a magic-number reciprocal is "provably exact for size <= 4096 with a 32-bit multiply". That bound is wrong. acc <= 255*size, so the product only fits in 32 bits if shift <= 24; exactness then needs 255*size*(size-1) < 2^24, i.e. it holds for odd sizes up to 265 and fails first at 267.

So the constants are checked rather than assumed. vs_reciprocal() searches downward from shift=31 for a shift satisfying both the overflow and the exactness condition, and returns valid=0 if none exists — in which case both passes keep the division. boxblurPlanar only ever passes stepSize-derived sizes, so the fast path is always taken in practice.

Applied to boxblur_hori too, where the division sat in the serial chain; that turned out to be the larger absolute win.

Correctness

Output is bit-identical.

• Exhaustive off-line check: every size in 1..8192 × every reachable acc, (acc*mul)>>shift == acc/size. No mismatch.
• boxblur.c's vertical output loop now reports loop vectorized using 16 byte vectors under -fopt-info-vec; it did not before.
• tests/test_boxblur.c now also checks boxblur_hori_C against a verbatim copy of the original — it was previously untested — and the size list straddles 265/267 so both the reciprocal and the division fallback are exercised. 500 → ~1200 checks.
• Full suite: 38/38 units pass. Mutation-checked that the new assertions do fail on a deliberately wrong magic number.

Measurements

Ryzen 9 9900X, 1080p, size=15, best of three (bench/bench_boxblur.c):

	1 thread	24 threads
boxblur_hori	2.23 → 1.15 ms/frame	0.22 → 0.11 ms/frame
boxblur_vert	2.35 → 0.44 ms/frame	0.88 → 0.27 ms/frame

2.9x on the blur in both configurations.

🤖 Generated with Claude Code