Use delta rather than markers in gridpoints_in_polygon

benchmark/autofill.jl

@@ -28,4 +28,37 @@ function coarse_and_fine_autofill()
     return
 end
 
-SUITE["autofill"] = @benchmarkable coarse_and_fine_autofill()
+SUITE["autofill"] = BenchmarkGroup()
+SUITE["autofill"]["coarse_and_fine"] = @benchmarkable coarse_and_fine_autofill()
+
+# gridpoints_in_polygon microbenchmarks at varying grid densities
+using DeviceLayout.Polygons: difference2d
+
+# Polygon with a cutout (exercises both horizontal edge and winding number paths)
+const _gip_r1 = Rectangle(200μm, 200μm)
+const _gip_r2 = Rectangle(100μm, 100μm) + Point(50μm, 50μm)
+const _gip_poly = [difference2d(_gip_r1, _gip_r2)]
+
+const _gip_grids = [
+    (
+        "33x25",
+        collect(range(-10μm, 210μm, length=33)),
+        collect(range(-10μm, 210μm, length=25))
+    ),
+    (
+        "321x241",
+        collect(range(-10μm, 210μm, length=321)),
+        collect(range(-10μm, 210μm, length=241))
+    ),
+    (
+        "3201x2401",
+        collect(range(-10μm, 210μm, length=3201)),
+        collect(range(-10μm, 210μm, length=2401))
+    )
+]
+
+SUITE["autofill"]["gridpoints_in_polygon"] = BenchmarkGroup()
+for (label, gx, gy) in _gip_grids
+    SUITE["autofill"]["gridpoints_in_polygon"][label] =
+        @benchmarkable gridpoints_in_polygon($_gip_poly, $gx, $gy)
+end

benchmark/tune.json

@@ -1 +1 @@
-[{"Julia":"1.12.0","BenchmarkTools":{"major":1,"minor":6,"patch":2,"prerelease":[],"build":[]}},[["BenchmarkGroup",{"data":{"schematic":["BenchmarkGroup",{"data":{"plan":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"curves":["BenchmarkGroup",{"data":{"turns_render":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"bsplines_render":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"offset_bspline_approximation":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"polygons":["BenchmarkGroup",{"data":{"rectangles":["BenchmarkGroup",{"data":{"render_units":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render_units_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render_convertunits":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"circles":["BenchmarkGroup",{"data":{"direct":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"entity_delta":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"direct_units_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"direct_units":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"entity_atol":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"direct_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}]},"tags":[]}],"flatten":["BenchmarkGroup",{"data":{"shallow_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"array_reference":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"nested_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"clipping":["BenchmarkGroup",{"data":{"difference2d_offset":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"difference2d":["BenchmarkGroup",{"data":{"100_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"3600_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"961_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"19881_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"8000_skew":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"10000_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}]},"tags":[]}],"bounds":["BenchmarkGroup",{"data":{"random_rectangles":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"shallow_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"random_circles":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"nested_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"intersection":["BenchmarkGroup",{"data":{"straight_lines":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"autofill":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}]]]
+[{"Julia":"1.12.0","BenchmarkTools":{"major":1,"minor":6,"patch":2,"prerelease":[],"build":[]}},[["BenchmarkGroup",{"data":{"schematic":["BenchmarkGroup",{"data":{"plan":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"curves":["BenchmarkGroup",{"data":{"turns_render":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"bsplines_render":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"offset_bspline_approximation":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"polygons":["BenchmarkGroup",{"data":{"rectangles":["BenchmarkGroup",{"data":{"render_units":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render_units_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render_convertunits":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"render_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"circles":["BenchmarkGroup",{"data":{"direct":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"entity_delta":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"direct_units_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"direct_units":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"entity_atol":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"direct_gds":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}]},"tags":[]}],"flatten":["BenchmarkGroup",{"data":{"shallow_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"array_reference":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"nested_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"clipping":["BenchmarkGroup",{"data":{"difference2d_offset":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"difference2d":["BenchmarkGroup",{"data":{"100_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"3600_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"961_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"19881_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"8000_skew":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"10000_square":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}]},"tags":[]}],"bounds":["BenchmarkGroup",{"data":{"random_rectangles":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"shallow_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"random_circles":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"nested_references":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"intersection":["BenchmarkGroup",{"data":{"straight_lines":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}],"autofill":["BenchmarkGroup",{"data":{"coarse_and_fine":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"gridpoints_in_polygon":["BenchmarkGroup",{"data":{"33x25":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"3201x2401":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}],"321x241":["Parameters",{"gctrial":true,"time_tolerance":0.05,"evals_set":false,"samples":10000,"evals":1,"gcsample":false,"seconds":5.0,"overhead":0.0,"memory_tolerance":0.01}]},"tags":[]}]},"tags":[]}]},"tags":[]}]]]

src/polygons.jl

@@ -2100,6 +2100,10 @@ function gridpoints_in_polygon(
     fT = float(T)
     grid_x = convert.(fT, sort(grid_x))
     grid_y = convert.(fT, sort(grid_y))
+    nx = length(grid_x)
+
+    # Pre-strip units once for horizontal edge lookups (avoids per-edge allocation)
+    grid_x_stripped = ustrip.(grid_x)
 
     # Segment endpoints
     p0s = points.(poly)
@@ -2120,24 +2124,22 @@ function gridpoints_in_polygon(
     )
     edge_tree = IntervalTrees.IntervalMap{fT, Tuple{Point{T}, Point{T}}}(edge_iv)
 
+    # Difference array for winding number accumulation (extra element as sentinel)
+    delta = zeros(Int32, nx + 1)
     # For each grid y value, look at the edges containing that value
-    grid_count = zeros(Int32, length(grid_x))
     for (iy, y) in enumerate(grid_y)
         for i2 in intersect(edge_tree, (y, y))
             edge = i2.value
             # Count horizontal edges only if grid point lies on them
             if last(edge).y == first(edge).y
                 if first(edge).y == y
-                    in_poly[
-                        findall(
-                            in(
-                                ustrip(min(first(edge).x, last(edge).x)) ..
-                                ustrip(max(first(edge).x, last(edge).x))
-                            ),
-                            ustrip.(grid_x)
-                        ),
-                        iy
-                    ] .= true
+                    x_lo = ustrip(min(first(edge).x, last(edge).x))
+                    x_hi = ustrip(max(first(edge).x, last(edge).x))
+                    i_lo = searchsortedfirst(grid_x_stripped, x_lo)
+                    i_hi = searchsortedlast(grid_x_stripped, x_hi)
+                    for i = i_lo:i_hi
+                        in_poly[i, iy] = true
+                    end
                 end
                 continue
             end
@@ -2150,18 +2152,25 @@ function gridpoints_in_polygon(
                 first(edge).x +
                 (last(edge).x - first(edge).x) * (y - first(edge).y) /
                 (last(edge).y - first(edge).y)
-            ix = findfirst(grid_x .>= x) # The first index to the right of x
-            x_right_minidx = isnothing(ix) ? length(grid_x) + 1 : ix
+            x_right_minidx = searchsortedfirst(grid_x, x) # Can be nx + 1 if x > last(grid_x)
             # If the grid point lies on the edge, it's in the polygon
-            x_right_minidx <= length(grid_x) &&
+            x_right_minidx <= nx &&
                 x == grid_x[x_right_minidx] &&
                 (in_poly[x_right_minidx, iy] = true)
-            # Add edge count to all grid points to the left (sign depending on direction)
+            # Record edge crossing in difference array (O(1) instead of O(grid_x))
             s = sign(last(edge).y - first(edge).y)
-            @. grid_count[1:(x_right_minidx - 1)] = grid_count[1:(x_right_minidx - 1)] + s
+            delta[1] += s
+            delta[x_right_minidx] -= s # In bounds thanks to sentinel
+        end
+        # Prefix sum over delta to recover winding counts, then mark in_poly
+        acc = zero(Int32)
+        for ix = 1:nx
+            acc += delta[ix]
+            if acc != 0
+                in_poly[ix, iy] = true
+            end
         end
-        @. in_poly[:, iy] = (grid_count != 0) || in_poly[:, iy]
-        grid_count .= 0
+        delta .= 0
     end
 
     return in_poly