Predicate benchmarks + two rounds of implementation optimization guided by dotnet-trace profiling

[Copilot]

Adds a PredicatesBenchmarks class and applies two rounds of implementation-level optimizations to PredicatesAdaptive and PredicatesExact. No mathematical algorithms were changed.

Benchmarks

PredicatesBenchmarks covers all 8 method variants: Orient2d/InCircle × double/float × general-position (Stage A fast path) × near-degenerate (forces full expansion path). Uses ShortRun, MemoryDiagnoser, and EventPipeProfiler.

Round 1 — static hotspot analysis

Optimization	Target
[SkipLocalsInit] on ExpansionSum, ScaleExpansionSum, PredicatesExact.InCircle, PredicatesExact.Orient2d, PredicatesAdaptive.InCircle	Eliminates zero-init of large stack frames (up to 11 KB per call)
[AggressiveInlining] on TwoTwoDiff	Was the only hot expansion helper missing the hint
Inline Split into MultTail and DekkersPresplit	Removes two ValueTuple<double,double> allocations per TwoTwoDiff call
SIMD (Vector256<double>) in Estimate and NegateInto	Vectorises summation/negation of up to 96-element spans

Round 2 — dotnet-trace profiling

Top hotspots from dotnet-trace collect (76 284 events, inclusive call-time):

Frame	Samples	Change
ExpansionSum	10 718	Tiered stackalloc (16/64/400 tiers); bounds-check-free merge + accumulation loops via Unsafe.Add; Unsafe.CopyBlockUnaligned for tail copies and the elen/flen=0 fast paths; [AggressiveOptimization]
ScaleExpansionSum	6 761	[SkipLocalsInit] retained; AggressiveInlining deliberately withheld — see regression note below
SpanHelpers.Memmove	93	Eliminated: was caused by Span.CopyTo in ExpansionSum fast paths
ScaleExpansion	23	[AggressiveInlining | AggressiveOptimization], [SkipLocalsInit], bounds-check-free loop
TwoTwoDiff	9	[SkipLocalsInit]
PredicatesExact.InCircle stack	—	In-place negate eliminates bdetPos[96] + ddetPos[96] (saves 1 536 B per call); [AggressiveOptimization]
PredicatesExact.Orient2d callsite	—	[AggressiveInlining] allows inlining into AdaptiveOrient2d Stage D

Regression caught during R2: adding [AggressiveInlining] to ScaleExpansionSum caused it to inline 3× into AdaptiveInCircle, each copy cascading into 4 inlined ScaleExpansion calls — 12 total. This bloated the JIT frame, polluted the instruction cache, and regressed Adaptive InCircle double by +52% even for Stage-A calls that never touch expansion code. Removed the hint; full recovery confirmed.

Results (N = 4 000 calls/iteration, .NET 10, AVX2)

Method	Baseline	After R1	After R2	Total Δ
Exact – Orient2d double (general)	986 µs	480 µs	486 µs	−51%
Exact – InCircle double (near-degenerate)	7 087 µs	3 767 µs	3 810 µs	−46%
Exact – InCircle double (general)	12 359 µs	7 961 µs	8 035 µs	−35%
Exact – Orient2d double (near-degenerate)	294 µs	265 µs	266 µs	−10%
Adaptive – Orient2d double (general)	22.1 µs	18.4 µs	18.4 µs	−17%
Adaptive – InCircle double (general/hard)	~44 µs	~42 µs	~42 µs	−5%
Float variants (all)	—	—	—	~0% (float exact uses decimal, not expansion arithmetic)

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[github-actions]

Summary

Summary

Generated on:	2/23/2026 - 12:23:35 AM
Coverage date:	2/23/2026 - 12:23:32 AM
Parser:	Cobertura
Assemblies:	1
Classes:	17
Files:	7
Line coverage:	85% (1325 of 1558)
Covered lines:	1325
Uncovered lines:	233
Coverable lines:	1558
Total lines:	3859
Branch coverage:	79.1% (627 of 792)
Covered branches:	627
Total branches:	792
Method coverage:	Feature is only available for sponsors
Tag:	65_22288154652

Coverage

CDT.Core - 85%

Name	Line	Branch
CDT.Core	85%	79.1%
CDT.Box2d`1	73.6%	100%
CDT.CdtUtils	83.1%	72.8%
CDT.CovariantReadOnlyDictionary`3	18.1%	50%
CDT.DictionaryExtensions	100%	100%
CDT.DuplicatesInfo	100%
CDT.DuplicateVertexException	66.6%
CDT.Edge	75%	66.6%
CDT.IntersectingConstraintsException	0%
CDT.KdTree`1	72.7%	65%
CDT.LayerDepth	0%
CDT.Predicates.PredicatesAdaptive	92.6%	83.3%
CDT.Predicates.PredicatesExact	35.3%	100%
CDT.TopologyVerifier	77.2%	63.8%
CDT.Triangle	66.6%	71.4%
CDT.Triangulation`1	93.8%	84.1%
CDT.TriangulationException	100%
CDT.V2d`1	44.4%	25%

[Copilot]

Pull request overview

This pull request adds comprehensive benchmarks for geometric predicates and applies two rounds of profiler-guided performance optimizations to the predicate implementation. The optimizations target the expensive expansion arithmetic used in exact geometric predicates, achieving significant speedups (10-51%) without changing any mathematical algorithms.

Changes:

• Added PredicatesBenchmarks class covering all 8 predicate method variants (Orient2d/InCircle × double/float × general/near-degenerate cases)
• Applied micro-optimizations including [SkipLocalsInit], aggressive inlining, SIMD vectorization, tiered stackalloc, and bounds-check elimination using unsafe code
• Eliminated unnecessary temporary stack allocations in PredicatesExact.InCircle through in-place negation

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
benchmark/CDT.Benchmarks/PredicatesBenchmarks.cs	Adds comprehensive benchmark suite with 16 benchmark methods covering all predicate variants, plus helper methods for generating test data
src/CDT.Core/Predicates/PredicatesExact.cs	Adds optimization attributes and eliminates 1536 bytes of stack allocation per InCircle call via in-place negation
src/CDT.Core/Predicates/PredicatesAdaptive.cs	Extensive optimization: SIMD vectorization in Estimate/NegateInto, tiered stackalloc in ExpansionSum, bounds-check elimination throughout, and Split function inlining

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.