uber / h3-go
100%
master: 100%

LAST BUILD BRANCH: justinhwang/forward-gnomonic-opt
DEFAULT BRANCH: master
Repo Added
Files 18
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

LAST BUILD ON BRANCH justinhwang/forward-gnomonic-opt
branch: SELECT
CHANGE BRANCH
x

coverage: 100.0%. Remained the same
27738291857

Pull #134

github

justinhwang 
perf: fold forward gnomonic projection (toHex2d) into algebra

toHex2d is the forward gnomonic projection behind LatLngToCell, PolygonToCells,
and local-IJ -> cell. Like the earlier toVec3 work it computed the angle and
radius with transcendentals: acos for the angular distance, atan2 (inside
azimuthRads) for the azimuth, tan for the gnomonic scaling, and a final sin/cos
pair, plus two vec3d.normalize() sqrts (the tangent basis and the projection).

Replace them with algebra:
  - the azimuth's cos/sin come straight from the tangent-plane components
    (cosAz = N/hyp, sinAz = E/hyp) instead of atan2 -> cos/sin; because only the
    N/E ratio is used, neither the tangent basis nor the projection needs
    normalizing -- both scale by the same factor, which cancels (drops 2 sqrt);
  - the per-face azimuth offset and the Class III rotation fold in via
    angle-subtraction against the precomputed faceAxis0Cos/Sin and ap7Rot tables;
  - the gnomonic tan(acos x) reduces to sqrt(1-x^2)/x with x = 1 - sqd/2, written
    as sqrt((sqd/2)(2-sqd/2))/x = (1-x)(1+x) to avoid cancellation near center.

Five transcendentals + 2 normalize sqrts become 2 sqrt. The transform is
algebraically exact; output differs from C only at the ULP level, within the
parity suite's 1e-5 tolerance. azimuthRads and posAngleRads are now unused and
removed. All white-box and parity tests pass; package stays at 100% coverage.

NOTE on the face-center guard: the old `acos(1-sqd/2) < epsilon` short-circuit
(the azimuth is singular at the face center, where the tangent projection
vanishes) is re-expressed on sqd as `sqd < epsilon*epsilon`. The exact form
2(1-cos epsilon) underflows to 0 at epsilon=1e-16 (cos(1e-16) rounds to 1), so
its leading-order equivalent epsilon^2 is used; without a working guard a point
exactly on a face center divides 0/0 and yields NaN coordinates.

benchstat, pure-Go, 4953a75 -> this commit (Apple M3 Max, count=10):

                              │   sec/op    │  ... (continued)
Pull Request #134: perf: fold forward gnomonic projection (toHex2d) into algebra

27 of 27 new or added lines in 1 file covered. (100.0%)

4268 of 4268 relevant lines covered (100.0%)

3071535.36 hits per line

Relevant lines Covered
4268 RELEVANT LINES 4268 COVERED LINES
3071535.36 HITS PER LINE
Source Files on master

Recent builds

Builds Branch Commit Type Ran Committer Via Coverage
27738291857 justinhwang/forward-gnomonic-opt perf: fold forward gnomonic projection (toHex2d) into algebra toHex2d is the forward gnomonic projection behind LatLngToCell, PolygonToCells, and local-IJ -> cell. Like the earlier toVec3 work it computed the angle and radius with transcendentals... Pull #134 justinhwang github
100.0
27736586680 justinhwang/tovec3-projection-opt perf: fold projection trig into precomputed tables and fused IJK descent Two hot transforms behind nearly every geometry call are reworked. toVec3 (the inverse gnomonic projection behind CellToLatLng, CellToBoundary, CellArea*, EdgeLength*, vert... Pull #133 justinhwang github
100.0
27736352283 justinhwang/tovec3-projection-opt perf: fold projection trig into precomputed tables and fused IJK descent Two hot transforms behind nearly every geometry call are reworked. toVec3 (the inverse gnomonic projection behind CellToLatLng, CellToBoundary, CellArea*, EdgeLength*, vert... push justinhwang github
100.0
27735272019 justinhwang/tovec3-projection-opt perf: fold toVec3 inverse projection into precomputed trig + one sqrt toVec3 (the inverse gnomonic projection behind CellToLatLng, CellToBoundary, CellArea*, EdgeLength*, vertex/edge boundaries, and the polygon paths) computed atan2 + atan + two ... push justinhwang github
99.77
27734628439 justinhwang/paritytest-benchmarks test+perf: parity benchmarks and allocation cuts for x/h3go Add a cgo-vs-pure-Go benchmark suite under x/h3go/paritytest covering every publicly exposed function and method, each run as "impl=cgo" and "impl=go" sub-benchmarks so `benchstat -col /... Pull #132 justinhwang github
99.76
27731679822 justinhwang/remove-h3core refactor: remove internal/h3core and reorganize x/h3go Remove the internal/h3core package and inline the shared H3 index-format constants and pentagon base-cell table into both consumers, so neither depends on the shared package: - h3: bit-layout... Pull #131 justinhwang github
100.0
27731605010 justinhwang/remove-h3core refactor(h3go): reorganize declarations and genericize shared index ops File organization: - Group const/var/type declaration blocks at the top of each source file. - Split error vars into errors.go; rename h3go.go -> constants.go and add h3.go... push justinhwang github
100.0
27727545624 justinhwang/purego-phase-g feat: add pure-Go regions (polygon/cells, multipolygon) to x/h3go Implements the full H3 regions API in pure Go, verified against the cgo reference: - PolygonToCells (legacy flood-fill polyfill) and GeoPolygon.Cells, with the GeoPolygon/GeoLoo... Pull #130 justinhwang github
100.0
27665848677 justinhwang/purego-phase-f feat: add pure-Go vertexes, directed edges & edge length to x/h3go Add the Vertex and DirectedEdge index types with their full method and free-function surfaces: Cell.Vertex/Vertexes and Vertex.LatLng/IsValid/ String/Resolution/IndexDigit; Cell.D... Pull #129 justinhwang github
100.0
27664523708 justinhwang/purego-phase-e feat: add pure-Go local IJ coordinates & grid metrics to x/h3go Add CellToLocalIJ/LocalIJToCell, GridDistance, GridPath (and Cell method forms) plus the CoordIJ type, implementing the local IJK/IJ coordinate machinery, cube-coordinate line interp... Pull #128 justinhwang github
100.0
See All Builds (332)