uber / h3-go / 36494950939

100%

push

github

perf: value-receiver coordIJK ops and deferred normalize in x/h3go (#142)

Profile GridDistance and GridPath on the pure-Go path. The time was not in
the algorithm but in how the coordIJK helpers compiled: normalize, upAp7,
downAp7 and friends were pointer-receiver mutators, and inlining a
pointer-receiver call marks the receiver address-taken, which pins it to the
stack. Every iteration of the aperture-7 loops in toFaceIjkWithInitializedFijk,
localIjkToCell and faceIJK.toH3 was doing three loads and three stores instead
of working in registers. unitToDigit and the upAp7 pair also sat just above the
inlining budget.

- Convert the coordIJK mutators (normalize, upAp7, upAp7r, downAp7, downAp7r,
  downAp3, downAp3r, toCube, fromCube) to value receivers that return the new
  coordinate, matching add, sub, scale and the rotations. Call sites become
  ijk = ijk.upAp7() or chain directly.
- toFaceIjkWithInitializedFijk normalizes once after the loop instead of once
  per resolution. Both aperture maps send a uniform offset to a uniform offset
  (the all-ones vector is an eigenvector of each), so the result is identical;
  magnitudes stay below 2^32 after 15 resolutions.
- unitToDigit rejects out-of-range components with one unsigned comparison
  each, which brings it under the inlining budget and drops the LUT bounds
  checks.
- faceIJK.toH3 walks the coordinate as a value instead of through a pointer
  into a copy.

Output is unchanged; the cgo parity tests cover every path touched. Because
toFaceIjk sits under most of the API, the win reaches well beyond GridDistance.

benchstat, Go implementation only, ten interleaved runs of each side on an
Apple M3 Max under background load (medians agree with an earlier idle run):

                              │ old (ebee012) │                 new                  │
                              │    sec/op     │    sec/op     vs base                │
GridDistance-16                 130.45n ± 34%   40.02n ±  9%  -69.... (continued)

79 of 79 new or added lines in 3 files covered. (100.0%)

4309 of 4309 relevant lines covered (100.0%)

2903949.97 hits per line