Merge #28

28: Performance tweak for orient2d r=urschrei a=tinko92 This PR proposes a minor change for the test of the approximate result in orient2d. It removes the if-statements based on detleft and detright and uses abs directly. This reduces the number of branches in the common case (the approximate result is larger in magnitude than error bound) to one well-predictable branch, which improves performance on modern CPUs. This behavior was described in https://doi.org/10.1007/s10543-015-0574-9 by Ozaki et al. for the orient2d-predicate. Tested with `RUSTFLAGS='-C target-cpu=native' cargo bench` on the reverse dependencies spade (2-5% improvements for most benchmarks) and delaunator-rs (~12-13% improvement for most benchmarks). Co-authored-by: Tinko Bartels <[email protected]>
georust · May 25, 2023 · 3f4e3ad · 3f4e3ad
2 parents 6cead7c + e1f3d3c
commit 3f4e3ad
Showing 1 changed file with 8 additions and 17 deletions.
diff --git a/src/lib.rs b/src/lib.rs
@@ -51,7 +51,6 @@ pub struct Coord3D<T: Into<f64>> {
 const SPLITTER: f64 = 134_217_729f64;
 const EPSILON: f64 = 0.000_000_000_000_000_111_022_302_462_515_65;
 const RESULTERRBOUND: f64 = (3.0 + 8.0 * EPSILON) * EPSILON;
-const CCWERRBOUND_A: f64 = (3.0 + 16.0 * EPSILON) * EPSILON;
 const CCWERRBOUND_B: f64 = (2.0 + 12.0 * EPSILON) * EPSILON;
 const CCWERRBOUND_C: f64 = (9.0 + 64.0 * EPSILON) * EPSILON * EPSILON;
 const O3DERRBOUND_A: f64 = (7.0 + 56.0 * EPSILON) * EPSILON;
@@ -86,22 +85,14 @@ pub fn orient2d<T: Into<f64>>(pa: Coord<T>, pb: Coord<T>, pc: Coord<T>) -> f64 {
     let detright = (pa.y - pc.y) * (pb.x - pc.x);
     let det = detleft - detright;
 
-    let detsum = if detleft > 0.0 {
-        if detright <= 0.0 {
-            return det;
-        } else {
-            detleft + detright
-        }
-    } else if detleft < 0.0 {
-        if detright >= 0.0 {
-            return det;
-        } else {
-            -detleft - detright
-        }
-    } else {
-        return det;
-    };
-    let errbound = CCWERRBOUND_A * detsum;
+    // The errbound calculation was changed to require only one branch on the likely execution path.
+    // This improves performance on modern processors as discussed by Ozaki et al in
+    // https://doi.org/10.1007/s10543-015-0574-9
+    // The underflow guard "+ u_N" was omitted because orient2dadapt(...) would not guarantee
+    // correct results in cases of underflow, the derivation of THETA is given in the reference.
+    let detsum = abs(detleft + detright);
+    const THETA: f64 = 3.3306690621773722e-16;
+    let errbound = THETA * detsum;
     if det >= errbound || -det >= errbound {
         det
     } else {