GPU usage restored to duplicate CL exceptions.

This reverts commit a84445ef2a.

.vscode/settings.json

@@ -9,5 +9,7 @@
 		"Mersenne",
 		"primality",
 		"suff"
-	]
+	],
+	"geminicodeassist.project": "evenperfectbitscanner",
+	"todo-tree.tree.showCountsInTree": true
 }

EvenPerfectBitScanner.Benchmarks/GpuModularArithmeticBenchmarks.cs

@@ -1,4 +1,3 @@
-using System;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/MersenneDivisorCycleLengthGpuBenchmarks.cs

@@ -1,6 +1,5 @@
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
-using PerfectNumbers.Core;
 using BenchmarkDotNet.Jobs;
 
 namespace EvenPerfectBitScanner.Benchmarks;

EvenPerfectBitScanner.Benchmarks/Mod10Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod10_8_5_3Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using PerfectNumbers.Core;

EvenPerfectBitScanner.Benchmarks/Mod11Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod128Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod3Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod5Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod6Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod7Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/Mod8Benchmarks.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;

EvenPerfectBitScanner.Benchmarks/ModBenchmarkCase.cs

@@ -1,5 +1,3 @@
-using System;
-
 namespace EvenPerfectBitScanner.Benchmarks;
 
 public enum ModNumericType

EvenPerfectBitScanner.Benchmarks/MontgomeryMultiplyBenchmarks.cs

@@ -1,4 +1,3 @@
-using System;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
 

EvenPerfectBitScanner.Benchmarks/MulMod64Benchmarks.Gpu.cs

@@ -1,12 +1,8 @@
-using System;
-using System.Collections.Generic;
-using System.Linq;
 using System.Runtime.CompilerServices;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
 using ILGPU;
 using ILGPU.Runtime;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
 
 namespace EvenPerfectBitScanner.Benchmarks;

EvenPerfectBitScanner.Benchmarks/Pow2ModBenchmarks.cs

@@ -1,4 +1,3 @@
-using System;
 using System.Buffers.Binary;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
@@ -101,7 +100,7 @@ public class Pow2ModBenchmarks
 
         for (int i = 0; i < SampleCount; i++)
         {
-            checksum ^= exponents[i].Pow2MontgomeryModWindowedGpu(divisors[i], keepMontgomery: false);
+            checksum ^= exponents[i].Pow2MontgomeryModWindowedConvertGpu(divisors[i]);
         }
 
         return checksum;

EvenPerfectBitScanner.Benchmarks/Pow2MontgomeryModBenchmarks.cs

@@ -77,7 +77,7 @@ public class Pow2MontgomeryModBenchmarks
 
         for (int i = 0; i < SampleCount; i++)
         {
-            checksum ^= exponents[i].Pow2MontgomeryModWindowedGpu(divisors[i], keepMontgomery: false);
+            checksum ^= exponents[i].Pow2MontgomeryModWindowedConvertGpu(divisors[i]);
         }
 
         return checksum;

EvenPerfectBitScanner.Benchmarks/Pow2MontgomeryModWindowedBenchmarkData.cs

@@ -1,4 +1,3 @@
-using System;
 using PerfectNumbers.Core;
 
 namespace EvenPerfectBitScanner.Benchmarks;

EvenPerfectBitScanner.Benchmarks/Pow2MontgomeryModWindowedKeepMontgomeryGpuBenchmarks.cs

@@ -1,4 +1,3 @@
-using System;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
 using PerfectNumbers.Core;
@@ -34,7 +33,7 @@ public class Pow2MontgomeryModWindowedKeepMontgomeryGpuBenchmarks
         for (int i = 0; i < cases.Length; i++)
         {
             ref readonly Pow2MontgomeryModWindowedBenchmarkCase current = ref cases[i];
-            checksum ^= current.Exponent.Pow2MontgomeryModWindowedGpu(current.Divisor, keepMontgomery: false);
+            checksum ^= current.Exponent.Pow2MontgomeryModWindowedConvertGpu(current.Divisor);
         }
 
         return checksum;
@@ -52,7 +51,7 @@ public class Pow2MontgomeryModWindowedKeepMontgomeryGpuBenchmarks
         for (int i = 0; i < cases.Length; i++)
         {
             ref readonly Pow2MontgomeryModWindowedBenchmarkCase current = ref cases[i];
-            checksum ^= current.Exponent.Pow2MontgomeryModWindowedGpu(current.Divisor, keepMontgomery: true);
+            checksum ^= current.Exponent.Pow2MontgomeryModWindowedKeepGpu(current.Divisor);
         }
 
         return checksum;

EvenPerfectBitScanner.Benchmarks/PrimeTesterBenchmarks.cs

@@ -1,10 +1,8 @@
-using System;
-using System.Collections.Generic;
-using System.Threading;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
 using Open.Numeric.Primes;
 using PerfectNumbers.Core;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace EvenPerfectBitScanner.Benchmarks;
 
@@ -12,7 +10,6 @@ namespace EvenPerfectBitScanner.Benchmarks;
 [MemoryDiagnoser]
 public class PrimeTesterBenchmarks
 {
-    private static readonly HeuristicCombinedPrimeTester _heuristicCombinedTester = new();
     private static readonly HeuristicPrimeTester _heuristicTester = new();
     private ulong[] _candidates = [];
 
@@ -29,12 +26,14 @@ public class PrimeTesterBenchmarks
             return;
         }
 
+		PrimeOrderCalculatorAccelerator.WarmUp();
+		// HeuristicGroupABPrimeTesterAccelerator.WarmUp();
         _ = _heuristicTester.IsPrimeCpu(_candidates[0]);
         _ = _heuristicTester.IsPrimeGpu(_candidates[0]);
         _ = Prime.Numbers.IsPrime(_candidates[0]);
     }
 
-    [Benchmark(Baseline = true)]
+    // [Benchmark]
     public int HeuristicCpu()
     {
         HeuristicPrimeTester tester = _heuristicTester;
@@ -104,18 +103,15 @@ public class PrimeTesterBenchmarks
         return primeCount;
     }
 
-	private static readonly PrimeTester _nonHeuristicTester = new();
-
     [Benchmark]
     public int NonHeuristicGpu()
 	{
-		PrimeTester tester = _nonHeuristicTester;
         ulong[] values = _candidates;
         int primeCount = 0;
 
         for (int i = 0; i < values.Length; i++)
         {
-            if (tester.IsPrimeGpu(values[i]))
+            if (PrimeTester.IsPrimeGpu(values[i]))
             {
                 primeCount++;
             }
@@ -124,7 +120,7 @@ public class PrimeTesterBenchmarks
         return primeCount;
     }
 
-    [Benchmark]
+    [Benchmark(Baseline = true)]
     public int NonHeuristicCpu()
     {
         ulong[] values = _candidates;
@@ -141,7 +137,7 @@ public class PrimeTesterBenchmarks
         return primeCount;
     }
 
-    [Benchmark]
+    // [Benchmark]
     public int OpenNumericPrimes()
     {
         ulong[] values = _candidates;

EvenPerfectBitScanner.Tests/ProgramTests.cs

@@ -3,13 +3,9 @@ using Xunit;
 using System.Reflection;
 using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
-
-using EvenPerfectBitScanner;
 using EvenPerfectBitScanner.Candidates;
 using EvenPerfectBitScanner.Candidates.Transforms;
 using System.Numerics;
-using System.Threading;
-using System.Threading.Tasks;
 
 namespace EvenPerfectBitScanner.Tests;
 
@@ -102,13 +98,11 @@ public class ProgramTests
         var testerField = typeof(Program).GetField("_divisorTester", BindingFlags.NonPublic | BindingFlags.Static)!;
         var candidatesField = typeof(MersenneNumberDivisorGpuTester).GetField("_divisorCandidates", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         testerField.SetValue(null, new MersenneNumberDivisorGpuTester());
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
         candidatesField.SetValue(null, new (ulong, uint)[] { (7UL, 3U), (23UL, 11U) });
-        forceCpuProp!.SetValue(null, true);
 
         try
         {
@@ -122,7 +116,6 @@ public class ProgramTests
             candidatesField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -135,13 +128,11 @@ public class ProgramTests
         var testerField = typeof(Program).GetField("_divisorTester", BindingFlags.NonPublic | BindingFlags.Static)!;
         var candidatesField = typeof(MersenneNumberDivisorGpuTester).GetField("_divisorCandidates", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         testerField.SetValue(null, new MersenneNumberDivisorGpuTester());
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
         candidatesField.SetValue(null, Array.Empty<(ulong, uint)>());
-        forceCpuProp!.SetValue(null, true);
 
         try
         {
@@ -155,7 +146,6 @@ public class ProgramTests
             candidatesField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -168,12 +158,10 @@ public class ProgramTests
 
         var testerField = typeof(Program).GetField("_divisorTester", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         testerField.SetValue(null, new MersenneNumberDivisorGpuTester());
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, true);
 
         try
         {
@@ -186,7 +174,6 @@ public class ProgramTests
             testerField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -199,12 +186,10 @@ public class ProgramTests
 
         var testerField = typeof(Program).GetField("_divisorTester", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         testerField.SetValue(null, new MersenneNumberDivisorGpuTester());
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, false);
 
         try
         {
@@ -221,7 +206,6 @@ public class ProgramTests
             testerField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -233,12 +217,10 @@ public class ProgramTests
 
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(useIncremental: true, useResidue: true, maxK: 1_000UL), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, false);
 
         try
         {
@@ -255,7 +237,6 @@ public class ProgramTests
             mersenneField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -267,7 +248,6 @@ public class ProgramTests
 
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(
             useIncremental: true,
@@ -275,7 +255,6 @@ public class ProgramTests
             maxK: 1_024UL), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, false);
 
         try
         {
@@ -292,7 +271,6 @@ public class ProgramTests
             mersenneField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -305,7 +283,6 @@ public class ProgramTests
 
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(
             useIncremental: true,
@@ -313,7 +290,6 @@ public class ProgramTests
             maxK: 5_000_000UL), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, false);
 
         try
         {
@@ -330,7 +306,6 @@ public class ProgramTests
             mersenneField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -342,7 +317,6 @@ public class ProgramTests
 
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(
             useIncremental: true,
@@ -350,7 +324,6 @@ public class ProgramTests
             maxK: 5_000_000UL), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, false);
 
         try
         {
@@ -367,7 +340,6 @@ public class ProgramTests
             mersenneField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -380,7 +352,6 @@ public class ProgramTests
 
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(
             useIncremental: true,
@@ -390,7 +361,6 @@ public class ProgramTests
             maxK: 1_024UL), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, exponent, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, true);
 
         try
         {
@@ -402,7 +372,6 @@ public class ProgramTests
             mersenneField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }
@@ -417,7 +386,6 @@ public class ProgramTests
 
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(
             useIncremental: false,
@@ -428,7 +396,6 @@ public class ProgramTests
             maxK: configuration.ResidueMaxK), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, exponent, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, false);
 
         try
         {
@@ -440,9 +407,7 @@ public class ProgramTests
             mersenneField.SetValue(null, null);
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
-            GpuContextPool.DisposeAll();
         }
     }
 
@@ -506,12 +471,10 @@ public class ProgramTests
         var mersenneField = typeof(Program).GetField("MersenneTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var primeField = typeof(Program).GetField("PrimeTesters", BindingFlags.NonPublic | BindingFlags.Static)!;
         var compositeField = typeof(Program).GetField("_lastCompositeP", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var forceCpuProp = typeof(GpuContextPool).GetProperty("ForceCpu");
 
         mersenneField.SetValue(null, new ThreadLocal<MersenneNumberTester>(() => new MersenneNumberTester(useIncremental: true, useResidue: false, maxK: 1_000UL), trackAllValues: true));
         primeField.SetValue(null, new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true));
         CandidatesCalculator.OverrideResidueTrackers(new ThreadLocal<ModResidueTracker>(() => new ModResidueTracker(ResidueModel.Identity, 2UL, true), trackAllValues: true));
-        forceCpuProp!.SetValue(null, true);
 
         try
         {
@@ -532,7 +495,6 @@ public class ProgramTests
             primeField.SetValue(null, null);
             CandidatesCalculator.ResetResidueTrackers();
             compositeField.SetValue(null, false);
-            forceCpuProp!.SetValue(null, false);
             ResetCliArguments();
         }
     }

EvenPerfectBitScanner.Tests/ResultsFileNameTests.cs

@@ -1,6 +1,3 @@
-using System.Collections.Generic;
-using System.IO;
-using System.Numerics;
 using FluentAssertions;
 using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;

EvenPerfectBitScanner/CalculationResultHandler.cs

@@ -1,8 +1,6 @@
 using System.Buffers.Text;
-using System.IO;
 using System.Runtime.CompilerServices;
 using System.Text;
-using System.Threading;
 using PerfectNumbers.Core;
 
 namespace EvenPerfectBitScanner;

EvenPerfectBitScanner/Candidates/CalculationTestTime.cs

@@ -1,5 +1,4 @@
 using System.Globalization;
-using System.IO;
 
 namespace EvenPerfectBitScanner.Candidates;
 

EvenPerfectBitScanner/Candidates/CandidatesCalculator.cs

@@ -1,5 +1,4 @@
 using System.Runtime.CompilerServices;
-using System.Threading;
 using Open.Numeric.Primes;
 using PerfectNumbers.Core;
 using EvenPerfectBitScanner.Candidates.Transforms;

EvenPerfectBitScanner/Candidates/Transforms/CandidateAddPrimesTransform.cs

@@ -1,6 +1,4 @@
 using System.Runtime.CompilerServices;
-using System.Threading;
-using EvenPerfectBitScanner.Candidates;
 using Open.Numeric.Primes;
 
 namespace EvenPerfectBitScanner.Candidates.Transforms;

EvenPerfectBitScanner/Candidates/Transforms/CandidateAddTransform.cs

@@ -1,6 +1,4 @@
 using System.Runtime.CompilerServices;
-using System.Threading;
-using EvenPerfectBitScanner.Candidates;
 
 namespace EvenPerfectBitScanner.Candidates.Transforms;
 

EvenPerfectBitScanner/Candidates/Transforms/CandidateBitTransform.cs

@@ -1,5 +1,4 @@
 using System.Runtime.CompilerServices;
-using System.Threading;
 
 namespace EvenPerfectBitScanner.Candidates.Transforms;
 

EvenPerfectBitScanner/CliArguments.cs

@@ -1,6 +1,4 @@
-using System;
 using System.Globalization;
-using System.Numerics;
 using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
 

EvenPerfectBitScanner/IO/CalculationResultsFile.cs

@@ -1,7 +1,6 @@
 using Open.Collections;
 using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
-using EvenPerfectBitScanner;
 
 namespace EvenPerfectBitScanner.IO;
 

EvenPerfectBitScanner/Program.cs

@@ -12,7 +12,7 @@ namespace EvenPerfectBitScanner;
 
 internal static class Program
 {
-	private static ThreadLocal<PrimeTester> PrimeTesters = null!;
+	// private static ThreadLocal<PrimeTester> PrimeTesters = null!;
 	private static ThreadLocal<MersenneNumberTester> MersenneTesters = null!;
 	private const ulong InitialP = PerfectNumberConstants.BiggestKnownEvenPerfectP;
 	private static MersenneNumberDivisorGpuTester? _divisorTester;
@@ -58,12 +58,14 @@ internal static class Program
 			bool startPrimeProvided = _cliArguments.StartPrimeProvided;
 			int threadCount = Math.Max(1, _cliArguments.ThreadCount);
 			int gpuPrimeBatch = Math.Max(1, _cliArguments.GpuPrimeBatch);
+			// int gpuPrimeThreads = SharedGpuContext.Device.MaxNumThreads;
 			int gpuPrimeThreads = Math.Max(1, _cliArguments.GpuPrimeThreads);
 			UnboundedTaskScheduler.ConfigureThreadCount(threadCount);
 			PrimeTester.GpuBatchSize = gpuPrimeBatch;
 			GpuPrimeWorkLimiter.SetLimit(gpuPrimeThreads);
-			GpuContextPool.WarmUpPool(threadCount);
-			PrimeTester.WarmUpGpuKernels(gpuPrimeThreads);
+			Console.WriteLine("Warming up GPU kernels");
+			PrimeTester.WarmUpGpuKernels(1024);			
+			// PrimeTester.WarmUpGpuKernels(gpuPrimeThreads >> 4);
 			Console.WriteLine("Starting up threads...");
 			_ = UnboundedTaskScheduler.Instance;
 			int blockSize = Math.Max(1, _cliArguments.BlockSize);
@@ -176,7 +178,7 @@ internal static class Program
 			blockSize = useByDivisor ? 1 : blockSize;
 
 			CandidatesCalculator.Initialize(currentP);
-			PrimeTesters = new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true);
+			// PrimeTesters = new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true);
 			// Note: --primes-device controls default device for library kernels; p primality remains CPU here.
 			// Initialize per-thread p residue tracker (Identity model) at currentP
 			if (!useDivisor && !_cliArguments.UseByDivisor)
@@ -626,7 +628,7 @@ internal static class Program
 		{
 			if (!(_runPrimesOnCpu
 					? PrimeTester.IsPrime(p)
-					: PrimeTesters.Value!.IsPrimeGpu(p)))
+					: HeuristicCombinedPrimeTester.IsPrimeGpu(p)))
 			{
 				_lastCompositeP = true;
 				return false;

EvenPerfectBitScanner/Utf8CliParser.cs

@@ -1,4 +1,3 @@
-using System;
 using System.Buffers;
 using System.Buffers.Text;
 using System.Text;

PerfectNumbers.Core.Tests/Gpu/GpuContextPoolTests.cs

@@ -1,7 +1,3 @@
-using FluentAssertions;
-using ILGPU;
-using ILGPU.Runtime;
-using PerfectNumbers.Core.Gpu;
 using Xunit;
 
 namespace PerfectNumbers.Core.Tests.Gpu;

PerfectNumbers.Core.Tests/Gpu/GpuKernelCompilationTests.cs

@@ -1,8 +1,8 @@
 using System.Reflection;
-using FluentAssertions;
 using ILGPU;
 using ILGPU.Runtime;
 using PerfectNumbers.Core.Gpu;
+using PerfectNumbers.Core.Gpu.Accelerators;
 using Xunit;
 
 namespace PerfectNumbers.Core.Tests.Gpu;
@@ -114,14 +114,10 @@ public class GpuKernelCompilationTests
         typeof(PrimeOrderGpuHeuristics).GetMethod("GetPow2ModKernel", BindingFlags.NonPublic | BindingFlags.Static)
         ?? throw new InvalidOperationException("PrimeOrderGpuHeuristics.GetPow2ModKernel not found.");
 
-    private static readonly MethodInfo PrimeOrderGetPow2ModWideKernelMethod =
-        typeof(PrimeOrderGpuHeuristics).GetMethod("GetPow2ModWideKernel", BindingFlags.NonPublic | BindingFlags.Static)
-        ?? throw new InvalidOperationException("PrimeOrderGpuHeuristics.GetPow2ModWideKernel not found.");
-
     private static readonly Action<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> MersenneDivisorCyclesKernel =
         DivisorCycleKernels.GpuDivisorCycleKernel;
 
-    private static readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> PrimeTesterSmallPrimeKernel =
+    private static readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> PrimeTesterSmallPrimeKernel =
         PrimeTesterKernels.SmallPrimeSieveKernel;
 
     private static readonly Action<Index1D, ArrayView<ulong>, ArrayView<byte>> PrimeTesterSharesFactorKernel =
@@ -325,56 +321,67 @@ public class GpuKernelCompilationTests
 
     private static void CompilePrimeOrderPow2ModWideKernel(Accelerator accelerator)
     {
-        _ = PrimeOrderGetPow2ModWideKernelMethod.Invoke(null, new object[] { accelerator });
+        var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(PrimeOrderGpuHeuristics.Pow2ModKernelWide);
+        _ = kernel;
     }
 
     private static void CompileMersenneDivisorCyclesKernel(Accelerator accelerator)
     {
-        var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(MersenneDivisorCyclesKernel);
+        var kernel = accelerator.LoadAutoGroupedStreamKernel(MersenneDivisorCyclesKernel);
         _ = kernel;
     }
 
     private static void CompilePrimeTesterSmallPrimeKernel(Accelerator accelerator)
     {
-        var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterSmallPrimeKernel);
+        var kernel = accelerator.LoadAutoGroupedStreamKernel(PrimeTesterSmallPrimeKernel);
         _ = kernel;
     }
 
     private static void CompilePrimeTesterSharesFactorKernel(Accelerator accelerator)
     {
-        var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterSharesFactorKernel);
+        var kernel = accelerator.LoadAutoGroupedStreamKernel(PrimeTesterSharesFactorKernel);
         _ = kernel;
     }
 
-    private static void CompileGpuKernelPoolKernels(Accelerator accelerator)
-    {
-        _ = accelerator;
-        var lease = GpuKernelPool.GetKernel(useGpuOrder: false);
-        try
-        {
-            _ = lease.OrderKernel;
-            _ = lease.IncrementalKernel;
-            _ = lease.Pow2ModKernel;
-            _ = lease.IncrementalOrderKernel;
-            _ = lease.Pow2ModOrderKernel;
-        }
-        finally
-        {
-            lease.Dispose();
-        }
+	private static KernelType _kernelType = KernelType.Pow2ModKernelScan | KernelType.OrderKernelScan | KernelType.IncrementalKernelScan | KernelType.IncrementalOrderKernelScan | KernelType.Pow2ModOrderKernelScan;
+	private static void CompileGpuKernelPoolKernels(Accelerator accelerator)
+	{
 
-        var orderLease = GpuKernelPool.GetKernel(useGpuOrder: true);
-        try
-        {
-            _ = orderLease.OrderKernel;
-            _ = orderLease.IncrementalKernel;
-            _ = orderLease.Pow2ModKernel;
-            _ = orderLease.IncrementalOrderKernel;
-            _ = orderLease.Pow2ModOrderKernel;
-        }
-        finally
-        {
-            orderLease.Dispose();
-        }
-    }
+		GpuPrimeWorkLimiter.Acquire();
+		int acceleratorIndex = AcceleratorPool.Shared.Rent();
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		try
+		{
+			var kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, _kernelType);
+			_ = kernels.Order;
+			_ = kernels.Incremental;
+			_ = kernels.Pow2Mod;
+			_ = kernels.IncrementalOrder;
+			_ = kernels.Pow2ModOrder;
+		}
+		finally
+		{
+			stream.Synchronize();
+			AcceleratorStreamPool.Return(acceleratorIndex, stream);
+			// GpuPrimeWorkLimiter.Release();
+		}
+
+		stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		GpuPrimeWorkLimiter.Acquire();
+		try
+		{
+			var kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, _kernelType);
+			_ = kernels.Order;
+			_ = kernels.Incremental;
+			_ = kernels.Pow2Mod;
+			_ = kernels.IncrementalOrder;
+			_ = kernels.Pow2ModOrder;
+		}
+		finally
+		{
+			stream.Synchronize();
+			AcceleratorStreamPool.Return(acceleratorIndex, stream);
+			// GpuPrimeWorkLimiter.Release();
+		}
+	}
 }

PerfectNumbers.Core.Tests/Gpu/GpuPrimeLimiterTests.cs

@@ -18,19 +18,19 @@ public class GpuPrimeLimiterTests
         var startedSecond = new TaskCompletionSource<bool>();
         var sw = new Stopwatch();
 
-        var first = GpuPrimeWorkLimiter.Acquire();
+        GpuPrimeWorkLimiter.Acquire();
         sw.Start();
 
         var secondTask = Task.Run(() =>
         {
-            var second = GpuPrimeWorkLimiter.Acquire();
+            GpuPrimeWorkLimiter.Acquire();
             try
             {
                 startedSecond.TrySetResult(true);
             }
             finally
             {
-                second.Dispose();
+				GpuPrimeWorkLimiter.Release();
             }
         });
 
@@ -43,7 +43,7 @@ public class GpuPrimeLimiterTests
         finally
         {
             // Release first after the initial wait so the second task can proceed.
-            first.Dispose();
+			GpuPrimeWorkLimiter.Release();
         }
 
         await secondTask.WaitAsync(TimeSpan.FromSeconds(1));

PerfectNumbers.Core.Tests/Gpu/NttGpuMathTests.cs

@@ -13,367 +13,352 @@ namespace PerfectNumbers.Core.Tests.Gpu;
 [Collection("GpuNtt")]
 public class NttGpuMathTests
 {
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void ForwardGpu_reference_matches_cpu_forward()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Reference;
-            Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
-            cpu[0] = new GpuUInt128(0UL, 1UL);
-            cpu[1] = new GpuUInt128(0UL, 2UL);
-            cpu[2] = new GpuUInt128(0UL, 3UL);
-            cpu[3] = new GpuUInt128(0UL, 4UL);
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
-            cpu.CopyTo(gpu);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            NttGpuMath.Forward(cpu, modulus, primitiveRoot);
-            NttGpuMath.ForwardGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < cpu.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void ForwardGpu_reference_matches_cpu_forward()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Reference;
+			Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
+			cpu[0] = new GpuUInt128(0UL, 1UL);
+			cpu[1] = new GpuUInt128(0UL, 2UL);
+			cpu[2] = new GpuUInt128(0UL, 3UL);
+			cpu[3] = new GpuUInt128(0UL, 4UL);
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
+			cpu.CopyTo(gpu);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			NttGpuMath.Forward(cpu, modulus, primitiveRoot);
+			NttGpuMath.ForwardGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < cpu.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void ForwardGpu_staged_matches_cpu_forward()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Staged;
-            Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
-            cpu[0] = new GpuUInt128(0UL, 1UL);
-            cpu[1] = new GpuUInt128(0UL, 2UL);
-            cpu[2] = new GpuUInt128(0UL, 3UL);
-            cpu[3] = new GpuUInt128(0UL, 4UL);
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
-            cpu.CopyTo(gpu);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            NttGpuMath.Forward(cpu, modulus, primitiveRoot);
-            NttGpuMath.ForwardGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < cpu.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void ForwardGpu_staged_matches_cpu_forward()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Staged;
+			Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
+			cpu[0] = new GpuUInt128(0UL, 1UL);
+			cpu[1] = new GpuUInt128(0UL, 2UL);
+			cpu[2] = new GpuUInt128(0UL, 3UL);
+			cpu[3] = new GpuUInt128(0UL, 4UL);
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
+			cpu.CopyTo(gpu);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			NttGpuMath.Forward(cpu, modulus, primitiveRoot);
+			NttGpuMath.ForwardGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < cpu.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void InverseGpu_reference_matches_cpu_inverse()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Reference;
-            Span<GpuUInt128> input = stackalloc GpuUInt128[4];
-            input[0] = new GpuUInt128(0UL, 5UL);
-            input[1] = new GpuUInt128(0UL, 7UL);
-            input[2] = new GpuUInt128(0UL, 9UL);
-            input[3] = new GpuUInt128(0UL, 11UL);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            Span<GpuUInt128> transformed = stackalloc GpuUInt128[4];
-            input.CopyTo(transformed);
-            NttGpuMath.Forward(transformed, modulus, primitiveRoot);
-            Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
-            transformed.CopyTo(cpu);
-            NttGpuMath.Inverse(cpu, modulus, primitiveRoot);
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
-            transformed.CopyTo(gpu);
-            NttGpuMath.InverseGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < input.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void InverseGpu_reference_matches_cpu_inverse()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Reference;
+			Span<GpuUInt128> input = stackalloc GpuUInt128[4];
+			input[0] = new GpuUInt128(0UL, 5UL);
+			input[1] = new GpuUInt128(0UL, 7UL);
+			input[2] = new GpuUInt128(0UL, 9UL);
+			input[3] = new GpuUInt128(0UL, 11UL);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			Span<GpuUInt128> transformed = stackalloc GpuUInt128[4];
+			input.CopyTo(transformed);
+			NttGpuMath.Forward(transformed, modulus, primitiveRoot);
+			Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
+			transformed.CopyTo(cpu);
+			NttGpuMath.Inverse(cpu, modulus, primitiveRoot);
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
+			transformed.CopyTo(gpu);
+			NttGpuMath.InverseGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < input.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void InverseGpu_staged_matches_cpu_inverse()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Staged;
-            Span<GpuUInt128> input = stackalloc GpuUInt128[4];
-            input[0] = new GpuUInt128(0UL, 5UL);
-            input[1] = new GpuUInt128(0UL, 7UL);
-            input[2] = new GpuUInt128(0UL, 9UL);
-            input[3] = new GpuUInt128(0UL, 11UL);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            Span<GpuUInt128> transformed = stackalloc GpuUInt128[4];
-            input.CopyTo(transformed);
-            NttGpuMath.Forward(transformed, modulus, primitiveRoot);
-            Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
-            transformed.CopyTo(cpu);
-            NttGpuMath.Inverse(cpu, modulus, primitiveRoot);
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
-            transformed.CopyTo(gpu);
-            NttGpuMath.InverseGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < input.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void InverseGpu_staged_matches_cpu_inverse()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Staged;
+			Span<GpuUInt128> input = stackalloc GpuUInt128[4];
+			input[0] = new GpuUInt128(0UL, 5UL);
+			input[1] = new GpuUInt128(0UL, 7UL);
+			input[2] = new GpuUInt128(0UL, 9UL);
+			input[3] = new GpuUInt128(0UL, 11UL);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			Span<GpuUInt128> transformed = stackalloc GpuUInt128[4];
+			input.CopyTo(transformed);
+			NttGpuMath.Forward(transformed, modulus, primitiveRoot);
+			Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
+			transformed.CopyTo(cpu);
+			NttGpuMath.Inverse(cpu, modulus, primitiveRoot);
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
+			transformed.CopyTo(gpu);
+			NttGpuMath.InverseGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < input.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void Montgomery_twiddle_inverses_multiply_to_r2()
-    {
-        var modulus = new GpuUInt128(0UL, 17UL);
-        var primitiveRoot = new GpuUInt128(0UL, 3UL);
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void Montgomery_twiddle_inverses_multiply_to_r2()
+	{
+		var modulus = new GpuUInt128(0UL, 17UL);
+		var primitiveRoot = new GpuUInt128(0UL, 3UL);
 
-        GpuContextPool.GpuContextLease gpu = GpuContextPool.Rent();
-        try
-        {
-            var accelerator = gpu.Accelerator;
+		var accelerator = SharedGpuContext.CreateAccelerator();
+		var stream = accelerator.CreateStream();
 
-            var method = typeof(NttGpuMath).GetMethod("GetSquareCache", BindingFlags.NonPublic | BindingFlags.Static);
-            method.Should().NotBeNull();
-            var cache = method!.Invoke(null, new object[] { accelerator, 8, modulus, primitiveRoot })!;
-            var cacheType = cache.GetType();
+		var method = typeof(NttGpuMath).GetMethod("GetSquareCache", BindingFlags.NonPublic | BindingFlags.Static);
+		method.Should().NotBeNull();
+		var cache = method!.Invoke(null, new object[] { accelerator, 8, modulus, primitiveRoot })!;
+		var cacheType = cache.GetType();
 
-            var twiddlesMont = (MemoryBuffer1D<GpuUInt128, Stride1D.Dense>)cacheType.GetProperty("TwiddlesMont")!.GetValue(cache)!;
-            var twiddlesInvMont = (MemoryBuffer1D<GpuUInt128, Stride1D.Dense>)cacheType.GetProperty("TwiddlesInvMont")!.GetValue(cache)!;
-            var r2 = (ulong)cacheType.GetProperty("MontR2Mod64")!.GetValue(cache)!;
-            var mod = (ulong)cacheType.GetProperty("ModulusLow")!.GetValue(cache)!;
+		var twiddlesMont = (MemoryBuffer1D<GpuUInt128, Stride1D.Dense>)cacheType.GetProperty("TwiddlesMont")!.GetValue(cache)!;
+		var twiddlesInvMont = (MemoryBuffer1D<GpuUInt128, Stride1D.Dense>)cacheType.GetProperty("TwiddlesInvMont")!.GetValue(cache)!;
+		var r2 = (ulong)cacheType.GetProperty("MontR2Mod64")!.GetValue(cache)!;
+		var mod = (ulong)cacheType.GetProperty("ModulusLow")!.GetValue(cache)!;
 
-            int count = 7; // length 8 -> 7 twiddles
-            var forward = new GpuUInt128[count];
-            var inverse = new GpuUInt128[count];
-            twiddlesMont.CopyToCPU(forward);
-            twiddlesInvMont.CopyToCPU(inverse);
+		int count = 7; // length 8 -> 7 twiddles
+		var forward = new GpuUInt128[count];
+		var inverse = new GpuUInt128[count];
+		twiddlesMont.CopyToCPU(stream, forward);
+		twiddlesInvMont.CopyToCPU(stream, inverse);
+		stream.Synchronize();
+		stream.Dispose();
+		accelerator.Dispose();
 
-            for (int i = 0; i < count; i++)
-            {
-                ulong prod = (ulong)((((BigInteger)forward[i].Low) * inverse[i].Low) % mod);
-                prod.Should().Be(r2);
-            }
-        }
-        finally
-        {
-            gpu.Dispose();
-        }
-    }
+		for (int i = 0; i < count; i++)
+		{
+			ulong prod = (ulong)((((BigInteger)forward[i].Low) * inverse[i].Low) % mod);
+			prod.Should().Be(r2);
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void ConvolveGpu_reference_matches_cpu_convolve()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Reference;
-            Span<GpuUInt128> leftCpu = stackalloc GpuUInt128[4];
-            Span<GpuUInt128> rightCpu = stackalloc GpuUInt128[4];
-            leftCpu[0] = new GpuUInt128(0UL, 1UL);
-            leftCpu[1] = new GpuUInt128(0UL, 2UL);
-            leftCpu[2] = new GpuUInt128(0UL, 3UL);
-            leftCpu[3] = new GpuUInt128(0UL, 4UL);
-            rightCpu[0] = new GpuUInt128(0UL, 5UL);
-            rightCpu[1] = new GpuUInt128(0UL, 6UL);
-            rightCpu[2] = new GpuUInt128(0UL, 7UL);
-            rightCpu[3] = new GpuUInt128(0UL, 8UL);
-            Span<GpuUInt128> leftGpu = stackalloc GpuUInt128[4];
-            Span<GpuUInt128> rightGpu = stackalloc GpuUInt128[4];
-            leftCpu.CopyTo(leftGpu);
-            rightCpu.CopyTo(rightGpu);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            NttGpuMath.Convolve(leftCpu, rightCpu, modulus, primitiveRoot);
-            NttGpuMath.ConvolveGpu(leftGpu, rightGpu, modulus, primitiveRoot);
-            for (int i = 0; i < leftCpu.Length; i++)
-            {
-                leftGpu[i].Should().Be(leftCpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void ConvolveGpu_reference_matches_cpu_convolve()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Reference;
+			Span<GpuUInt128> leftCpu = stackalloc GpuUInt128[4];
+			Span<GpuUInt128> rightCpu = stackalloc GpuUInt128[4];
+			leftCpu[0] = new GpuUInt128(0UL, 1UL);
+			leftCpu[1] = new GpuUInt128(0UL, 2UL);
+			leftCpu[2] = new GpuUInt128(0UL, 3UL);
+			leftCpu[3] = new GpuUInt128(0UL, 4UL);
+			rightCpu[0] = new GpuUInt128(0UL, 5UL);
+			rightCpu[1] = new GpuUInt128(0UL, 6UL);
+			rightCpu[2] = new GpuUInt128(0UL, 7UL);
+			rightCpu[3] = new GpuUInt128(0UL, 8UL);
+			Span<GpuUInt128> leftGpu = stackalloc GpuUInt128[4];
+			Span<GpuUInt128> rightGpu = stackalloc GpuUInt128[4];
+			leftCpu.CopyTo(leftGpu);
+			rightCpu.CopyTo(rightGpu);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			NttGpuMath.Convolve(leftCpu, rightCpu, modulus, primitiveRoot);
+			NttGpuMath.ConvolveGpu(leftGpu, rightGpu, modulus, primitiveRoot);
+			for (int i = 0; i < leftCpu.Length; i++)
+			{
+				leftGpu[i].Should().Be(leftCpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void ConvolveGpu_staged_matches_cpu_convolve()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Staged;
-            Span<GpuUInt128> leftCpu = stackalloc GpuUInt128[4];
-            Span<GpuUInt128> rightCpu = stackalloc GpuUInt128[4];
-            leftCpu[0] = new GpuUInt128(0UL, 1UL);
-            leftCpu[1] = new GpuUInt128(0UL, 2UL);
-            leftCpu[2] = new GpuUInt128(0UL, 3UL);
-            leftCpu[3] = new GpuUInt128(0UL, 4UL);
-            rightCpu[0] = new GpuUInt128(0UL, 5UL);
-            rightCpu[1] = new GpuUInt128(0UL, 6UL);
-            rightCpu[2] = new GpuUInt128(0UL, 7UL);
-            rightCpu[3] = new GpuUInt128(0UL, 8UL);
-            Span<GpuUInt128> leftGpu = stackalloc GpuUInt128[4];
-            Span<GpuUInt128> rightGpu = stackalloc GpuUInt128[4];
-            leftCpu.CopyTo(leftGpu);
-            rightCpu.CopyTo(rightGpu);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            NttGpuMath.Convolve(leftCpu, rightCpu, modulus, primitiveRoot);
-            NttGpuMath.ConvolveGpu(leftGpu, rightGpu, modulus, primitiveRoot);
-            for (int i = 0; i < leftCpu.Length; i++)
-            {
-                leftGpu[i].Should().Be(leftCpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void ConvolveGpu_staged_matches_cpu_convolve()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Staged;
+			Span<GpuUInt128> leftCpu = stackalloc GpuUInt128[4];
+			Span<GpuUInt128> rightCpu = stackalloc GpuUInt128[4];
+			leftCpu[0] = new GpuUInt128(0UL, 1UL);
+			leftCpu[1] = new GpuUInt128(0UL, 2UL);
+			leftCpu[2] = new GpuUInt128(0UL, 3UL);
+			leftCpu[3] = new GpuUInt128(0UL, 4UL);
+			rightCpu[0] = new GpuUInt128(0UL, 5UL);
+			rightCpu[1] = new GpuUInt128(0UL, 6UL);
+			rightCpu[2] = new GpuUInt128(0UL, 7UL);
+			rightCpu[3] = new GpuUInt128(0UL, 8UL);
+			Span<GpuUInt128> leftGpu = stackalloc GpuUInt128[4];
+			Span<GpuUInt128> rightGpu = stackalloc GpuUInt128[4];
+			leftCpu.CopyTo(leftGpu);
+			rightCpu.CopyTo(rightGpu);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			NttGpuMath.Convolve(leftCpu, rightCpu, modulus, primitiveRoot);
+			NttGpuMath.ConvolveGpu(leftGpu, rightGpu, modulus, primitiveRoot);
+			for (int i = 0; i < leftCpu.Length; i++)
+			{
+				leftGpu[i].Should().Be(leftCpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void SquareGpu_reference_matches_cpu_square()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Reference;
-            Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
-            cpu[0] = new GpuUInt128(0UL, 1UL);
-            cpu[1] = new GpuUInt128(0UL, 2UL);
-            cpu[2] = new GpuUInt128(0UL, 3UL);
-            cpu[3] = new GpuUInt128(0UL, 4UL);
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
-            cpu.CopyTo(gpu);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            NttGpuMath.Square(cpu, modulus, primitiveRoot);
-            NttGpuMath.SquareGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < cpu.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void SquareGpu_reference_matches_cpu_square()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Reference;
+			Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
+			cpu[0] = new GpuUInt128(0UL, 1UL);
+			cpu[1] = new GpuUInt128(0UL, 2UL);
+			cpu[2] = new GpuUInt128(0UL, 3UL);
+			cpu[3] = new GpuUInt128(0UL, 4UL);
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
+			cpu.CopyTo(gpu);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			NttGpuMath.Square(cpu, modulus, primitiveRoot);
+			NttGpuMath.SquareGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < cpu.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void SquareGpu_staged_matches_cpu_square()
-    {
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Staged;
-            Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
-            cpu[0] = new GpuUInt128(0UL, 1UL);
-            cpu[1] = new GpuUInt128(0UL, 2UL);
-            cpu[2] = new GpuUInt128(0UL, 3UL);
-            cpu[3] = new GpuUInt128(0UL, 4UL);
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
-            cpu.CopyTo(gpu);
-            var modulus = new GpuUInt128(0UL, 17UL);
-            var primitiveRoot = new GpuUInt128(0UL, 3UL);
-            NttGpuMath.Square(cpu, modulus, primitiveRoot);
-            NttGpuMath.SquareGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < cpu.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void SquareGpu_staged_matches_cpu_square()
+	{
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Staged;
+			Span<GpuUInt128> cpu = stackalloc GpuUInt128[4];
+			cpu[0] = new GpuUInt128(0UL, 1UL);
+			cpu[1] = new GpuUInt128(0UL, 2UL);
+			cpu[2] = new GpuUInt128(0UL, 3UL);
+			cpu[3] = new GpuUInt128(0UL, 4UL);
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[4];
+			cpu.CopyTo(gpu);
+			var modulus = new GpuUInt128(0UL, 17UL);
+			var primitiveRoot = new GpuUInt128(0UL, 3UL);
+			NttGpuMath.Square(cpu, modulus, primitiveRoot);
+			NttGpuMath.SquareGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < cpu.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void BitReverseGpu_matches_cpu()
-    {
-        GpuUInt128[] cpu = new GpuUInt128[8];
-        for (ulong i = 0; i < (ulong)cpu.Length; i++)
-        {
-            cpu[i] = new GpuUInt128(0UL, i);
-        }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void BitReverseGpu_matches_cpu()
+	{
+		GpuUInt128[] cpu = new GpuUInt128[8];
+		for (ulong i = 0; i < (ulong)cpu.Length; i++)
+		{
+			cpu[i] = new GpuUInt128(0UL, i);
+		}
 
-        GpuUInt128[] gpu = (GpuUInt128[])cpu.Clone();
-        NttGpuMath.BitReverse(cpu);
-        NttGpuMath.BitReverseGpu(gpu);
-        gpu.Should().Equal(cpu);
-    }
+		GpuUInt128[] gpu = (GpuUInt128[])cpu.Clone();
+		NttGpuMath.BitReverse(cpu);
+		NttGpuMath.BitReverseGpu(gpu);
+		gpu.Should().Equal(cpu);
+	}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void DisposeAll_clears_ntt_caches()
-    {
-        // Ensure the cache starts empty.
-        NttGpuMath.ClearCaches();
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void DisposeAll_clears_ntt_caches()
+	{
+		// Ensure the cache starts empty.
+		NttGpuMath.ClearCaches();
 
-        Span<GpuUInt128> values = stackalloc GpuUInt128[2];
-        values[0] = new GpuUInt128(0UL, 1UL);
-        values[1] = new GpuUInt128(0UL, 2UL);
-        var modulus = new GpuUInt128(0UL, 17UL);
-        var primitiveRoot = new GpuUInt128(0UL, 3UL);
-        GpuContextPool.GpuContextLease gpu = GpuContextPool.Rent();
-        try
-        {
-            var buffer = gpu.Accelerator.Allocate1D<GpuUInt128>(values.Length);
-            try
-            {
-                buffer.View.CopyFromCPU(ref values[0], values.Length);
-                NttGpuMath.SquareDevice(gpu.Accelerator, buffer.View, modulus, primitiveRoot);
-            }
-            finally
-            {
-                buffer.Dispose();
-            }
-        }
-        finally
-        {
-            gpu.Dispose();
-        }
+		Span<GpuUInt128> values = stackalloc GpuUInt128[2];
+		values[0] = new GpuUInt128(0UL, 1UL);
+		values[1] = new GpuUInt128(0UL, 2UL);
+		var modulus = new GpuUInt128(0UL, 17UL);
+		var primitiveRoot = new GpuUInt128(0UL, 3UL);
+		Accelerator accelerator = SharedGpuContext.CreateAccelerator();
+		var stream = accelerator.CreateStream();
+		var buffer = accelerator.Allocate1D<GpuUInt128>(values.Length);
+		buffer.View.CopyFromCPU(stream, ref values[0], values.Length);
+		NttGpuMath.SquareDevice(accelerator, stream, buffer.View, modulus, primitiveRoot);
+		stream.Synchronize();
+		stream.Dispose();
+		buffer.Dispose();
+		accelerator.Dispose();
 
-        // Verify that the square cache now holds entries for the accelerator.
-        var field = typeof(NttGpuMath).GetField("SquareCache", BindingFlags.NonPublic | BindingFlags.Static)!;
-        var cacheObj = field.GetValue(null)!;
-        int count = (int)cacheObj.GetType().GetProperty("Count")!.GetValue(cacheObj)!;
-        count.Should().BeGreaterThan(0);
+		// Verify that the square cache now holds entries for the accelerator.
+		var field = typeof(NttGpuMath).GetField("SquareCache", BindingFlags.NonPublic | BindingFlags.Static)!;
+		var cacheObj = field.GetValue(null)!;
+		int count = (int)cacheObj.GetType().GetProperty("Count")!.GetValue(cacheObj)!;
+		count.Should().BeGreaterThan(0);
 
-        GpuContextPool.DisposeAll();
-
-        count = (int)cacheObj.GetType().GetProperty("Count")!.GetValue(cacheObj)!;
-        count.Should().Be(0);
-    }
+		count = (int)cacheObj.GetType().GetProperty("Count")!.GetValue(cacheObj)!;
+		count.Should().Be(0);
+	}
 }
 

PerfectNumbers.Core.Tests/Gpu/NttMontgomeryTests.cs

@@ -9,86 +9,83 @@ namespace PerfectNumbers.Core.Tests.Gpu;
 [Trait("Category", "Fast")]
 public class NttMontgomeryTests
 {
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void SquareDevice_montgomery64_matches_cpu_square_on_small_input()
-    {
-        // Small 64-bit modulus and tiny input to keep runtime trivial.
-        var modulus = new GpuUInt128(0UL, 17UL);
-        var primitiveRoot = new GpuUInt128(0UL, 3UL);
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void SquareDevice_montgomery64_matches_cpu_square_on_small_input()
+	{
+		// Small 64-bit modulus and tiny input to keep runtime trivial.
+		var modulus = new GpuUInt128(0UL, 17UL);
+		var primitiveRoot = new GpuUInt128(0UL, 3UL);
 
-        Span<GpuUInt128> cpu = stackalloc GpuUInt128[8];
-        for (ulong i = 0; i < (ulong)cpu.Length; i++)
-        {
-            cpu[(int)i] = new GpuUInt128(0UL, i + 1UL);
-        }
-        Span<GpuUInt128> gpu = stackalloc GpuUInt128[16];
-        cpu.CopyTo(gpu);
-        for (int i = cpu.Length; i < gpu.Length; i++)
-        {
-            gpu[i] = new GpuUInt128(0UL, 0UL);
-        }
+		Span<GpuUInt128> cpu = stackalloc GpuUInt128[8];
+		for (ulong i = 0; i < (ulong)cpu.Length; i++)
+		{
+			cpu[(int)i] = new GpuUInt128(0UL, i + 1UL);
+		}
+		Span<GpuUInt128> gpu = stackalloc GpuUInt128[16];
+		cpu.CopyTo(gpu);
+		for (int i = cpu.Length; i < gpu.Length; i++)
+		{
+			gpu[i] = new GpuUInt128(0UL, 0UL);
+		}
 
-        // CPU reference
-        NttGpuMath.Square(cpu, modulus, primitiveRoot);
+		// CPU reference
+		NttGpuMath.Square(cpu, modulus, primitiveRoot);
 
-        // Device execution
-        GpuContextPool.GpuContextLease context = GpuContextPool.Rent();
-        try
-        {
-            var accelerator = context.Accelerator;
-            using var buffer = accelerator.Allocate1D<GpuUInt128>(gpu.Length);
-            buffer.View.CopyFromCPU(ref gpu[0], gpu.Length);
-            NttGpuMath.SquareDevice(accelerator, buffer.View, modulus, primitiveRoot);
-            buffer.View.CopyToCPU(ref gpu[0], gpu.Length);
+		// Device execution
+		var accelerator = SharedGpuContext.CreateAccelerator();
+		var stream = accelerator.CreateStream();
+		using var buffer = accelerator.Allocate1D<GpuUInt128>(gpu.Length);
+		buffer.View.CopyFromCPU(stream, ref gpu[0], gpu.Length);
+		NttGpuMath.SquareDevice(accelerator, stream, buffer.View, modulus, primitiveRoot);
+		buffer.View.CopyToCPU(stream, ref gpu[0], gpu.Length);
+		stream.Synchronize();
+		stream.Dispose();
+		accelerator.Dispose();
 
-            for (int i = 0; i < cpu.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            context.Dispose();
-        }
-    }
+		for (int i = 0; i < cpu.Length; i++)
+		{
+			gpu[i].Should().Be(cpu[i]);
+		}
 
-    [Fact]
-    [Trait("Category", "Fast")]
-    public void InverseDevice_montgomery64_matches_cpu_inverse_on_small_input()
-    {
-        var modulus = new GpuUInt128(0UL, 17UL);
-        var primitiveRoot = new GpuUInt128(0UL, 3UL);
+	}
 
-        Span<GpuUInt128> input = stackalloc GpuUInt128[8];
-        for (ulong i = 0; i < (ulong)input.Length; i++)
-        {
-            input[(int)i] = new GpuUInt128(0UL, i + 1UL);
-        }
+	[Fact]
+	[Trait("Category", "Fast")]
+	public void InverseDevice_montgomery64_matches_cpu_inverse_on_small_input()
+	{
+		var modulus = new GpuUInt128(0UL, 17UL);
+		var primitiveRoot = new GpuUInt128(0UL, 3UL);
 
-        Span<GpuUInt128> transformed = stackalloc GpuUInt128[input.Length];
-        input.CopyTo(transformed);
-        NttGpuMath.Forward(transformed, modulus, primitiveRoot);
+		Span<GpuUInt128> input = stackalloc GpuUInt128[8];
+		for (ulong i = 0; i < (ulong)input.Length; i++)
+		{
+			input[(int)i] = new GpuUInt128(0UL, i + 1UL);
+		}
 
-        Span<GpuUInt128> cpu = stackalloc GpuUInt128[input.Length];
-        transformed.CopyTo(cpu);
-        NttGpuMath.Inverse(cpu, modulus, primitiveRoot);
+		Span<GpuUInt128> transformed = stackalloc GpuUInt128[input.Length];
+		input.CopyTo(transformed);
+		NttGpuMath.Forward(transformed, modulus, primitiveRoot);
 
-        var original = NttGpuMath.GpuTransformBackend;
-        try
-        {
-            NttGpuMath.GpuTransformBackend = NttBackend.Staged;
-            Span<GpuUInt128> gpu = stackalloc GpuUInt128[input.Length];
-            transformed.CopyTo(gpu);
-            NttGpuMath.InverseGpu(gpu, modulus, primitiveRoot);
-            for (int i = 0; i < input.Length; i++)
-            {
-                gpu[i].Should().Be(cpu[i]);
-            }
-        }
-        finally
-        {
-            NttGpuMath.GpuTransformBackend = original;
-        }
-    }
+		Span<GpuUInt128> cpu = stackalloc GpuUInt128[input.Length];
+		transformed.CopyTo(cpu);
+		NttGpuMath.Inverse(cpu, modulus, primitiveRoot);
+
+		var original = NttGpuMath.GpuTransformBackend;
+		try
+		{
+			NttGpuMath.GpuTransformBackend = NttBackend.Staged;
+			Span<GpuUInt128> gpu = stackalloc GpuUInt128[input.Length];
+			transformed.CopyTo(gpu);
+			NttGpuMath.InverseGpu(gpu, modulus, primitiveRoot);
+			for (int i = 0; i < input.Length; i++)
+			{
+				gpu[i].Should().Be(cpu[i]);
+			}
+		}
+		finally
+		{
+			NttGpuMath.GpuTransformBackend = original;
+		}
+	}
 }

PerfectNumbers.Core.Tests/Gpu/PrimeTesterGpuTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core.Gpu;
 using Xunit;
 
 namespace PerfectNumbers.Core.Tests.Gpu;
@@ -10,37 +9,23 @@ public class PrimeTesterGpuTests
     [Fact]
     public void IsPrimeGpu_accepts_known_primes()
     {
-        try
-        {
-            var tester = new PrimeTester();
-            ulong[] primes = [31UL, 61UL, 89UL, 107UL, 127UL, 521UL];
-            foreach (ulong prime in primes)
-            {
-                tester.IsPrimeGpu(prime).Should().BeTrue();
-            }
-        }
-        finally
-        {
-            GpuContextPool.DisposeAll();
-        }
+		var tester = new PrimeTester();
+		ulong[] primes = [31UL, 61UL, 89UL, 107UL, 127UL, 521UL];
+		foreach (ulong prime in primes)
+		{
+			PrimeTester.IsPrimeGpu(prime).Should().BeTrue();
+		}
     }
 
     [Fact]
     public void IsPrimeGpu_rejects_composites()
     {
-        try
-        {
-            var tester = new PrimeTester();
-            ulong[] composites = [33UL, 39UL, 51UL, 77UL, 91UL];
-            foreach (ulong composite in composites)
-            {
-                tester.IsPrimeGpu(composite).Should().BeFalse();
-            }
-        }
-        finally
-        {
-            GpuContextPool.DisposeAll();
-        }
+		var tester = new PrimeTester();
+		ulong[] composites = [33UL, 39UL, 51UL, 77UL, 91UL];
+		foreach (ulong composite in composites)
+		{
+			PrimeTester.IsPrimeGpu(composite).Should().BeFalse();
+		}
     }
 
 }

PerfectNumbers.Core.Tests/Gpu/ULongExtensionsGpuTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core;
 using Xunit;
 
 namespace PerfectNumbers.Core.Tests.Gpu;

PerfectNumbers.Core.Tests/HeuristicCombinedPrimeTesterTests.cs

@@ -1,7 +1,3 @@
-using System;
-using System.Numerics;
-using System.Collections.Generic;
-using PerfectNumbers.Core;
 using FluentAssertions;
 using Xunit;
 

PerfectNumbers.Core.Tests/MersenneNumberDivisorByDivisorTesterTests.cs

@@ -83,7 +83,7 @@ public class MersenneNumberDivisorByDivisorTesterTests
                 throw new NotSupportedException();
             }
 
-            public void Dispose()
+            public void Return()
             {
             }
         }

PerfectNumbers.Core.Tests/MersenneNumberDivisorCpuTesterTests.cs

@@ -47,7 +47,7 @@ public class MersenneNumberDivisorCpuTesterTests
                 }
                 finally
                 {
-                        session.Dispose();
+                        session.Return();
                 }
         }
 
@@ -72,7 +72,7 @@ public class MersenneNumberDivisorCpuTesterTests
                 }
                 finally
                 {
-                        session.Dispose();
+                        session.Return();
                 }
         }
 
@@ -97,7 +97,7 @@ public class MersenneNumberDivisorCpuTesterTests
                 }
                 finally
                 {
-                        session.Dispose();
+                        session.Return();
                 }
         }
 }

PerfectNumbers.Core.Tests/MersenneNumberDivisorGpuTesterTests.cs

@@ -161,7 +161,7 @@ public class MersenneNumberDivisorGpuTesterTests
         }
         finally
         {
-            session.Dispose();
+            session.Return();
         }
     }
 
@@ -189,7 +189,7 @@ public class MersenneNumberDivisorGpuTesterTests
         }
         finally
         {
-            session.Dispose();
+            session.Return();
         }
     }
 
@@ -217,7 +217,7 @@ public class MersenneNumberDivisorGpuTesterTests
         }
         finally
         {
-            session.Dispose();
+            session.Return();
         }
     }
 
@@ -246,7 +246,7 @@ public class MersenneNumberDivisorGpuTesterTests
         }
         finally
         {
-            session.Dispose();
+            session.Return();
         }
     }
 
@@ -275,7 +275,7 @@ public class MersenneNumberDivisorGpuTesterTests
         }
         finally
         {
-            session.Dispose();
+            session.Return();
         }
     }
 
@@ -317,7 +317,6 @@ public class MersenneNumberDivisorGpuTesterTests
             tableField.SetValue(cycles, originalTable);
             smallCyclesField.SetValue(cycles, originalSmall);
             DivisorCycleCache.Shared.RefreshSnapshot();
-            GpuContextPool.DisposeAll();
         }
     }
 
@@ -345,7 +344,7 @@ public class MersenneNumberDivisorGpuTesterTests
         }
         finally
         {
-            session.Dispose();
+            session.Return();
         }
     }
 }

PerfectNumbers.Core.Tests/MersenneNumberIncrementalCpuTesterTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Cpu;
 using Xunit;
 using UInt128 = System.UInt128;

PerfectNumbers.Core.Tests/MersenneNumberIncrementalGpuTesterTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
 using Xunit;
 using UInt128 = System.UInt128;

PerfectNumbers.Core.Tests/MersenneNumberOrderCpuTesterTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Cpu;
 using Xunit;
 using UInt128 = System.UInt128;

PerfectNumbers.Core.Tests/MersenneNumberOrderGpuTesterTests.cs

@@ -1,6 +1,5 @@
 using System.Reflection;
 using FluentAssertions;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
 using Xunit;
 using UInt128 = System.UInt128;
@@ -55,7 +54,6 @@ public class MersenneNumberOrderGpuTesterTests
             ulong[]? backupSmall = originalSmall is null ? null : (ulong[])originalSmall.Clone();
 
             cycles.LoadFrom(tempPath);
-            GpuContextPool.DisposeAll();
 
             try
             {
@@ -75,7 +73,6 @@ public class MersenneNumberOrderGpuTesterTests
             {
                 tableField.SetValue(cycles, originalTable);
                 smallCyclesField.SetValue(cycles, backupSmall);
-                GpuContextPool.DisposeAll();
             }
         }
         finally

PerfectNumbers.Core.Tests/MersenneNumberResidueCpuTesterTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Cpu;
 using Xunit;
 using UInt128 = System.UInt128;

PerfectNumbers.Core.Tests/MersenneNumberResidueGpuTesterTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
 using Xunit;
 using UInt128 = System.UInt128;

PerfectNumbers.Core.Tests/PrimeTester/PrimeTesterIsPrimeTests.cs

@@ -39,7 +39,7 @@ public class PrimeTesterIsPrimeTests
     {
         var tester = new PrimeTester();
 
-        tester.IsPrimeGpu(n).Should().Be(expected);
+		PrimeTester.IsPrimeGpu(n).Should().Be(expected);
     }
 
     [Fact]
@@ -64,7 +64,7 @@ public class PrimeTesterIsPrimeTests
         var testers = new ThreadLocal<PrimeTester>(() => new PrimeTester(), trackAllValues: true);
         Parallel.For(0, primes.Length, i =>
         {
-            results[i] = testers.Value!.IsPrimeGpu(primes[i]);
+            results[i] = PrimeTester.IsPrimeGpu(primes[i]);
         });
 
         results.Should().AllBeEquivalentTo(true);
@@ -73,7 +73,7 @@ public class PrimeTesterIsPrimeTests
         {
             foreach (ulong prime in primes)
             {
-                testers.Value!.IsPrimeGpu(prime).Should().BeTrue();
+				PrimeTester.IsPrimeGpu(prime).Should().BeTrue();
             }
         });
     }

PerfectNumbers.Core/Cpu/DeterministicRandomCpu.cs

@@ -1,6 +1,6 @@
-namespace PerfectNumbers.Core;
+namespace PerfectNumbers.Core.Cpu;
 
-internal static class DeterministicRandom
+internal static class DeterministicRandomCpu
 {
     [ThreadStatic]
     private static ulong s_threadState;

PerfectNumbers.Core/Cpu/MersenneCpuDivisorScanSession.cs

@@ -62,7 +62,7 @@ internal sealed class MersenneCpuDivisorScanSession : IMersenneNumberDivisorByDi
         }
     }
 
-    public void Dispose()
+    public void Return()
     {
         ThreadStaticPools.ReturnMersenneCpuDivisorSession(this);
     }

PerfectNumbers.Core/Cpu/MersenneNumberDivisorByDivisorCpuTester.cs

@@ -1,487 +1,495 @@
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core.Cpu;
 
 public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDivisorByDivisorTester
 {
-    private ulong _divisorLimit;
-    private int _batchSize = 1_024;
-
-    [ThreadStatic]
-    private static GpuUInt128WorkSet _divisorScanGpuWorkSet;
-
-    private struct GpuUInt128WorkSet
-    {
-        public GpuUInt128 Step;
-        public GpuUInt128 Divisor;
-        public GpuUInt128 Limit;
-    }
-
-    public int BatchSize
-    {
-        get => _batchSize;
-        set => _batchSize = Math.Max(1, value);
-    }
-
-    public void ConfigureFromMaxPrime(ulong maxPrime)
-    {
-        _divisorLimit = ComputeDivisorLimitFromMaxPrime(maxPrime);
-    }
-
-    public ulong DivisorLimit
-    {
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        get
-        {
-            return _divisorLimit;
-        }
-    }
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public ulong GetAllowedMaxDivisor(ulong prime) => ComputeAllowedMaxDivisor(prime, _divisorLimit);
-
-    public bool IsPrime(ulong prime, out bool divisorsExhausted)
-    {
-        ulong allowedMax = ComputeAllowedMaxDivisor(prime, _divisorLimit);
-
-        // The CPU by-divisor run always hands us primes with enormous divisor limits, so the fallback below never executes.
-        // if (allowedMax < 3UL)
-        // {
-        //     // EvenPerfectBitScanner routes primes below the small-divisor cutoff to the GPU path, so the CPU path still sees
-        //     // trivial candidates during targeted tests. Short-circuit here to keep those runs aligned with the production flow.
-        //     divisorsExhausted = true;
-        //     return true;
-        // }
-
-        bool processedAll;
-
-        bool composite = CheckDivisors(
-            prime,
-            allowedMax,
-            out processedAll);
-
-        if (composite)
-        {
-            divisorsExhausted = true;
-            return false;
-        }
-
-        divisorsExhausted = processedAll || composite;
-        return true;
-    }
-
-    public void PrepareCandidates(in ReadOnlySpan<ulong> primes, Span<ulong> allowedMaxValues)
-    {
-        ulong divisorLimit = _divisorLimit;
-        int length = primes.Length;
-        for (int index = 0; index < length; index++)
-        {
-            allowedMaxValues[index] = ComputeAllowedMaxDivisor(primes[index], divisorLimit);
-        }
-    }
-
-    public IMersenneNumberDivisorByDivisorTester.IDivisorScanSession CreateDivisorSession()
-    {
-        MersenneCpuDivisorScanSession? session = ThreadStaticPools.RentMersenneCpuDivisorSession();
-        if (session is not null)
-        {
-            return session;
-        }
-
-        return new MersenneCpuDivisorScanSession();
-    }
-
-    private bool CheckDivisors(
-        ulong prime,
-        ulong allowedMax,
-        out bool processedAll)
-    {
-        processedAll = false;
-
-        // The EvenPerfectBitScanner feeds primes >= 138,000,000 here, so allowedMax >= 3 in production runs.
-        // Keeping the guard commented out documents the reasoning for benchmarks and tests.
-        // if (allowedMax < 3UL)
-        // {
-        //     return false;
-        // }
-
-        ref GpuUInt128WorkSet workSet = ref _divisorScanGpuWorkSet;
-
-        ref GpuUInt128 step = ref workSet.Step;
-        step.High = 0UL;
-        step.Low = prime;
-        step.ShiftLeft(1);
-
-        ref GpuUInt128 limit = ref workSet.Limit;
-        limit.High = 0UL;
-        limit.Low = allowedMax;
-
-        ref GpuUInt128 divisor = ref workSet.Divisor;
-        divisor.High = step.High;
-        divisor.Low = step.Low;
-        divisor.Add(1UL);
-        if (divisor.CompareTo(limit) > 0)
-        {
-            processedAll = true;
-            return false;
-        }
-
-        // Intentionally recomputes factorizations without a per-thread cache.
-        // The previous factor cache recorded virtually no hits and only slowed down the scan.
-        DivisorCycleCache cycleCache = DivisorCycleCache.Shared;
-
-        ulong stepHigh = step.High;
-        ulong stepLow = step.Low;
-        ulong limitHigh = limit.High;
-        ulong divisorLow = divisor.Low;
-
-        // Keep the divisibility filters aligned with the divisor-cycle generator so the
-        // CPU path never requests cycles that were skipped during cache creation.
-        LastDigit lastDigit = (prime & 3UL) == 3UL ? LastDigit.Seven : LastDigit.One;
-        ushort decimalMask = DivisorGenerator.GetDecimalMask(lastDigit);
-
-        byte step10;
-        byte step8;
-        byte step3;
-        byte step7;
-        byte step11;
-
-        byte remainder10;
-        byte remainder8;
-        byte remainder3;
-        byte remainder7;
-        byte remainder11;
-
-        if (stepHigh == 0UL && limitHigh == 0UL)
-        {
-            step10 = (byte)(stepLow % 10UL);
-            step8 = (byte)(stepLow & 7UL);
-            step3 = (byte)(stepLow % 3UL);
-            step7 = (byte)(stepLow % 7UL);
-            step11 = (byte)(stepLow % 11UL);
-
-            remainder10 = (byte)(divisorLow % 10UL);
-            remainder8 = (byte)(divisorLow & 7UL);
-            remainder3 = (byte)(divisorLow % 3UL);
-            remainder7 = (byte)(divisorLow % 7UL);
-            remainder11 = (byte)(divisorLow % 11UL);
-
-            return CheckDivisors64(
-                prime,
-                stepLow,
-                limit.Low,
-                divisorLow,
-                decimalMask,
-                step10,
-                step8,
-                step3,
-                step7,
-                step11,
-                remainder10,
-                remainder8,
-                remainder3,
-                remainder7,
-                remainder11,
-                out processedAll);
-        }
-
-        ulong divisorHigh = divisor.High;
-        step10 = (byte)((((stepHigh % 10UL) * 6UL) + (stepLow % 10UL)) % 10UL);
-        step8 = (byte)(stepLow % 8UL);
-        step3 = (byte)(((stepHigh % 3UL) + (stepLow % 3UL)) % 3UL);
-        step7 = (byte)((((stepHigh % 7UL) * 2UL) + (stepLow % 7UL)) % 7UL);
-        step11 = (byte)((((stepHigh % 11UL) * 5UL) + (stepLow % 11UL)) % 11UL);
-
-        remainder10 = (byte)((((divisorHigh % 10UL) * 6UL) + (divisorLow % 10UL)) % 10UL);
-        remainder8 = (byte)(divisorLow % 8UL);
-        remainder3 = (byte)(((divisorHigh % 3UL) + (divisorLow % 3UL)) % 3UL);
-        remainder7 = (byte)((((divisorHigh % 7UL) * 2UL) + (divisorLow % 7UL)) % 7UL);
-        remainder11 = (byte)((((divisorHigh % 11UL) * 5UL) + (divisorLow % 11UL)) % 11UL);
-
-        while (divisor.CompareTo(limit) <= 0)
-        {
-            bool passesSmallModuli = remainder3 != 0 && remainder7 != 0 && remainder11 != 0;
-            if (passesSmallModuli && (remainder8 == 1 || remainder8 == 7) && ((decimalMask >> remainder10) & 1) != 0)
-            {
-                ulong candidate = divisor.Low;
-                MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(candidate);
-                ulong divisorCycle;
-                // Divisors generated from 2 * k * p + 1 exceed the small-cycle snapshot when p >= 138,000,000, so the short path below never runs.
-                // if (candidate <= PerfectNumberConstants.MaxQForDivisorCycles)
-                // {
-                //     divisorCycle = cycleCache.GetCycleLength(candidate);
-                // }
-                // else
-                {
-                    if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
-                            candidate,
-                            prime,
-                            divisorData,
-                            out ulong computedCycle,
-                            out bool primeOrderFailed) || computedCycle == 0UL)
-                    {
-                        // Divisors produced by 2 * k * p + 1 always exceed PerfectNumberConstants.MaxQForDivisorCycles
-                        // for the exponents scanned here, so skip the unused cache fallback and compute directly.
-                        divisorCycle = MersenneDivisorCycles.CalculateCycleLength(
-                            candidate,
-                            divisorData,
-                            skipPrimeOrderHeuristic: primeOrderFailed);
-                    }
-                    else
-                    {
-                        divisorCycle = computedCycle;
-                    }
-                }
-
-                if (divisorCycle == prime)
-                {
-                    // A cycle equal to the tested exponent (which is prime in this path) guarantees that the candidate divides
-                    // the corresponding Mersenne number because the order of 2 modulo the divisor is exactly p.
-                    processedAll = true;
-                    return true;
-                }
-
-                if (divisorCycle == 0UL)
-                {
-                    Console.WriteLine($"Divisor cycle was not calculated for {prime}");
-                }
-            }
-
-            divisor.Add(step);
-            remainder10 = AddMod10(remainder10, step10);
-            remainder8 = AddMod8(remainder8, step8);
-            remainder3 = AddMod3(remainder3, step3);
-            remainder7 = AddMod7(remainder7, step7);
-            remainder11 = AddMod11(remainder11, step11);
-        }
-        processedAll = true;
-        return false;
-    }
-
-    private static bool CheckDivisors64(
-        ulong prime,
-        ulong step,
-        ulong limit,
-        ulong divisor,
-        ushort decimalMask,
-        byte step10,
-        byte step8,
-        byte step3,
-        byte step7,
-        byte step11,
-        byte remainder10,
-        byte remainder8,
-        byte remainder3,
-        byte remainder7,
-        byte remainder11,
-        out bool processedAll)
-    {
-        processedAll = false;
-
-        bool canAdvance = step <= limit;
-        if (!canAdvance)
-        {
-            if (divisor <= limit)
-            {
-                if (remainder3 != 0 && remainder7 != 0 && remainder11 != 0 && (remainder8 == 1 || remainder8 == 7) && ((decimalMask >> remainder10) & 1) != 0)
-                {
-                    MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
-                    ulong divisorCycle;
-                    if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
-                            divisor,
-                            prime,
-                            divisorData,
-                            out ulong computedCycle,
-                            out bool primeOrderFailed) || computedCycle == 0UL)
-                    {
-                        divisorCycle = MersenneDivisorCycles.CalculateCycleLength(
-                            divisor,
-                            divisorData,
-                            skipPrimeOrderHeuristic: primeOrderFailed);
-                    }
-                    else
-                    {
-                        divisorCycle = computedCycle;
-                    }
-
-                    if (divisorCycle == prime)
-                    {
-                        processedAll = true;
-                        return true;
-                    }
-
-                    if (divisorCycle == 0UL)
-                    {
-                        Console.WriteLine($"Divisor cycle was not calculated for {prime}");
-                    }
-                }
-            }
-
-            processedAll = true;
-            return false;
-        }
-
-        if (divisor > limit)
-        {
-            processedAll = true;
-            return false;
-        }
-
-        ulong remainingIterations = ((limit - divisor) / step) + 1UL;
-
-        while (true)
-        {
-            bool passesSmallModuli = remainder3 != 0 && remainder7 != 0 && remainder11 != 0;
-            if (passesSmallModuli && (remainder8 == 1 || remainder8 == 7) && ((decimalMask >> remainder10) & 1) != 0)
-            {
-                MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
-                ulong divisorCycle;
-                if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
-                        divisor,
-                        prime,
-                        divisorData,
-                        out ulong computedCycle,
-                        out bool primeOrderFailed) || computedCycle == 0UL)
-                {
-                    divisorCycle = MersenneDivisorCycles.CalculateCycleLength(
-                        divisor,
-                        divisorData,
-                        skipPrimeOrderHeuristic: primeOrderFailed);
-                }
-                else
-                {
-                    divisorCycle = computedCycle;
-                }
-
-                if (divisorCycle == prime)
-                {
-                    processedAll = true;
-                    return true;
-                }
-
-                if (divisorCycle == 0UL)
-                {
-                    Console.WriteLine($"Divisor cycle was not calculated for {prime}");
-                }
-            }
-
-            remainingIterations--;
-            if (remainingIterations == 0UL)
-            {
-                processedAll = true;
-                return false;
-            }
-
-            divisor += step;
-            remainder10 = AddMod10(remainder10, step10);
-            remainder8 = AddMod8(remainder8, step8);
-            remainder3 = AddMod3(remainder3, step3);
-            remainder7 = AddMod7(remainder7, step7);
-            remainder11 = AddMod11(remainder11, step11);
-        }
-    }
-
-    private static byte CheckDivisor(ulong prime, ulong divisorCycle, in MontgomeryDivisorData divisorData)
-    {
-        ulong residue = prime.Pow2MontgomeryModWithCycleCpu(divisorCycle, divisorData);
-        return residue == 1UL ? (byte)1 : (byte)0;
-    }
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static byte AddMod3(byte value, byte delta)
-    {
-        const int Modulus = 3;
-        int sum = value + delta;
-
-        if (sum >= Modulus)
-        {
-            sum -= Modulus;
-        }
-
-        return (byte)sum;
-    }
-
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static byte AddMod7(byte value, byte delta)
-    {
-        const int Modulus = 7;
-        int sum = value + delta;
-
-        if (sum >= Modulus)
-        {
-            sum -= Modulus;
-        }
-
-        return (byte)sum;
-    }
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static byte AddMod10(byte value, byte delta)
-    {
-        const int Modulus = 10;
-        int sum = value + delta;
-
-        if (sum >= Modulus)
-        {
-            sum -= Modulus;
-        }
-
-        return (byte)sum;
-    }
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static byte AddMod11(byte value, byte delta)
-    {
-        const int Modulus = 11;
-        int sum = value + delta;
-
-        if (sum >= Modulus)
-        {
-            sum -= Modulus;
-        }
-
-        return (byte)sum;
-    }
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static byte AddMod8(byte value, byte delta)
-    {
-        return (byte)((value + delta) & 7);
-    }
-
-    private static ulong ComputeDivisorLimitFromMaxPrime(ulong maxPrime)
-    {
-        // The by-divisor CPU configuration only supplies primes greater than 1, so the guard below never trips.
-        // if (maxPrime <= 1UL)
-        // {
-        //     return 0UL;
-        // }
-        if (maxPrime - 1UL >= 64UL)
-        {
-            return ulong.MaxValue;
-        }
-
-        return (1UL << (int)(maxPrime - 1UL)) - 1UL;
-    }
-
-    private static ulong ComputeAllowedMaxDivisor(ulong prime, ulong divisorLimit)
-    {
-        // Production by-divisor scans only handle primes, so inputs never fall below 2.
-        // if (prime <= 1UL)
-        // {
-        //     return 0UL;
-        // }
-        if (prime - 1UL >= 64UL)
-        {
-            return divisorLimit;
-        }
-
-        return Math.Min((1UL << (int)(prime - 1UL)) - 1UL, divisorLimit);
-    }
+	private ulong _divisorLimit;
+	private int _batchSize = 1_024;
+
+	[ThreadStatic]
+	private static GpuUInt128WorkSet _divisorScanGpuWorkSet;
+
+	private struct GpuUInt128WorkSet
+	{
+		public GpuUInt128 Step;
+		public GpuUInt128 Divisor;
+		public GpuUInt128 Limit;
+	}
+
+	public int BatchSize
+	{
+		get => _batchSize;
+		set => _batchSize = Math.Max(1, value);
+	}
+
+	public void ConfigureFromMaxPrime(ulong maxPrime)
+	{
+		_divisorLimit = ComputeDivisorLimitFromMaxPrime(maxPrime);
+	}
+
+	public ulong DivisorLimit
+	{
+		[MethodImpl(MethodImplOptions.AggressiveInlining)]
+		get
+		{
+			return _divisorLimit;
+		}
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ulong GetAllowedMaxDivisor(ulong prime) => ComputeAllowedMaxDivisor(prime, _divisorLimit);
+
+	public bool IsPrime(ulong prime, out bool divisorsExhausted)
+	{
+		ulong allowedMax = ComputeAllowedMaxDivisor(prime, _divisorLimit);
+
+		// The CPU by-divisor run always hands us primes with enormous divisor limits, so the fallback below never executes.
+		// if (allowedMax < 3UL)
+		// {
+		//     // EvenPerfectBitScanner routes primes below the small-divisor cutoff to the GPU path, so the CPU path still sees
+		//     // trivial candidates during targeted tests. Short-circuit here to keep those runs aligned with the production flow.
+		//     divisorsExhausted = true;
+		//     return true;
+		// }
+
+		bool processedAll;
+
+		bool composite = CheckDivisors(
+			prime,
+			allowedMax,
+			out processedAll);
+
+		if (composite)
+		{
+			divisorsExhausted = true;
+			return false;
+		}
+
+		divisorsExhausted = processedAll || composite;
+		return true;
+	}
+
+	public void PrepareCandidates(in ReadOnlySpan<ulong> primes, Span<ulong> allowedMaxValues)
+	{
+		ulong divisorLimit = _divisorLimit;
+		int length = primes.Length;
+		for (int index = 0; index < length; index++)
+		{
+			allowedMaxValues[index] = ComputeAllowedMaxDivisor(primes[index], divisorLimit);
+		}
+	}
+
+	public IMersenneNumberDivisorByDivisorTester.IDivisorScanSession CreateDivisorSession()
+	{
+		MersenneCpuDivisorScanSession? session = ThreadStaticPools.RentMersenneCpuDivisorSession();
+		if (session is not null)
+		{
+			return session;
+		}
+
+		return new MersenneCpuDivisorScanSession();
+	}
+
+	private bool CheckDivisors(
+		ulong prime,
+		ulong allowedMax,
+		out bool processedAll)
+	{
+		// The EvenPerfectBitScanner feeds primes >= 138,000,000 here, so allowedMax >= 3 in production runs.
+		// Keeping the guard commented out documents the reasoning for benchmarks and tests.
+		// if (allowedMax < 3UL)
+		// {
+		//     return false;
+		// }
+
+		ref GpuUInt128WorkSet workSet = ref _divisorScanGpuWorkSet;
+
+		ref GpuUInt128 step = ref workSet.Step;
+		step.High = 0UL;
+		step.Low = prime;
+		step.ShiftLeft(1);
+
+		ref GpuUInt128 limit = ref workSet.Limit;
+		limit.High = 0UL;
+		limit.Low = allowedMax;
+
+		ref GpuUInt128 divisor = ref workSet.Divisor;
+		divisor.High = step.High;
+		divisor.Low = step.Low;
+		divisor.Add(1UL);
+		if (divisor.CompareTo(limit) > 0)
+		{
+			processedAll = true;
+			return false;
+		}
+
+		// Intentionally recomputes factorizations without a per-thread cache.
+		// The previous factor cache recorded virtually no hits and only slowed down the scan.
+		_ = DivisorCycleCache.Shared;
+
+		ulong stepHigh = step.High;
+		ulong stepLow = step.Low;
+		ulong limitHigh = limit.High;
+		ulong divisorLow = divisor.Low;
+
+		// Keep the divisibility filters aligned with the divisor-cycle generator so the
+		// CPU path never requests cycles that were skipped during cache creation.
+		LastDigit lastDigit = (prime & 3UL) == 3UL ? LastDigit.Seven : LastDigit.One;
+		ushort decimalMask = DivisorGenerator.GetDecimalMask(lastDigit);
+
+		byte step10;
+		byte step8;
+		byte step3;
+		byte step7;
+		byte step11;
+
+		byte remainder10;
+		byte remainder8;
+		byte remainder3;
+		byte remainder7;
+		byte remainder11;
+
+		if (stepHigh == 0UL && limitHigh == 0UL)
+		{
+			step10 = (byte)(stepLow % 10UL);
+			step8 = (byte)(stepLow & 7UL);
+			step3 = (byte)(stepLow % 3UL);
+			step7 = (byte)(stepLow % 7UL);
+			step11 = (byte)(stepLow % 11UL);
+
+			remainder10 = (byte)(divisorLow % 10UL);
+			remainder8 = (byte)(divisorLow & 7UL);
+			remainder3 = (byte)(divisorLow % 3UL);
+			remainder7 = (byte)(divisorLow % 7UL);
+			remainder11 = (byte)(divisorLow % 11UL);
+
+			return CheckDivisors64(
+				prime,
+				stepLow,
+				limit.Low,
+				divisorLow,
+				decimalMask,
+				step10,
+				step8,
+				step3,
+				step7,
+				step11,
+				remainder10,
+				remainder8,
+				remainder3,
+				remainder7,
+				remainder11,
+				out processedAll);
+		}
+
+		ulong divisorHigh = divisor.High;
+		step10 = (byte)((((stepHigh % 10UL) * 6UL) + (stepLow % 10UL)) % 10UL);
+		step8 = (byte)(stepLow % 8UL);
+		step3 = (byte)(((stepHigh % 3UL) + (stepLow % 3UL)) % 3UL);
+		step7 = (byte)((((stepHigh % 7UL) * 2UL) + (stepLow % 7UL)) % 7UL);
+		step11 = (byte)((((stepHigh % 11UL) * 5UL) + (stepLow % 11UL)) % 11UL);
+
+		remainder10 = (byte)((((divisorHigh % 10UL) * 6UL) + (divisorLow % 10UL)) % 10UL);
+		remainder8 = (byte)(divisorLow % 8UL);
+		remainder3 = (byte)(((divisorHigh % 3UL) + (divisorLow % 3UL)) % 3UL);
+		remainder7 = (byte)((((divisorHigh % 7UL) * 2UL) + (divisorLow % 7UL)) % 7UL);
+		remainder11 = (byte)((((divisorHigh % 11UL) * 5UL) + (divisorLow % 11UL)) % 11UL);
+
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+
+		while (divisor.CompareTo(limit) <= 0)
+		{
+			bool passesSmallModuli = remainder3 != 0 && remainder7 != 0 && remainder11 != 0;
+			if (passesSmallModuli && (remainder8 == 1 || remainder8 == 7) && ((decimalMask >> remainder10) & 1) != 0)
+			{
+				ulong candidate = divisor.Low;
+				MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(candidate);
+				ulong divisorCycle;
+				// Divisors generated from 2 * k * p + 1 exceed the small-cycle snapshot when p >= 138,000,000, so the short path below never runs.
+				// if (candidate <= PerfectNumberConstants.MaxQForDivisorCycles)
+				// {
+				//     divisorCycle = cycleCache.GetCycleLength(candidate);
+				// }
+				// else
+				{
+					if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
+							gpu,
+							candidate,
+							prime,
+							divisorData,
+							out ulong computedCycle,
+							out bool primeOrderFailed) || computedCycle == 0UL)
+					{
+						// Divisors produced by 2 * k * p + 1 always exceed PerfectNumberConstants.MaxQForDivisorCycles
+						// for the exponents scanned here, so skip the unused cache fallback and compute directly.
+						divisorCycle = MersenneDivisorCycles.CalculateCycleLength(
+							candidate,
+							divisorData,
+							skipPrimeOrderHeuristic: primeOrderFailed);
+					}
+					else
+					{
+						divisorCycle = computedCycle;
+					}
+				}
+
+				if (divisorCycle == prime)
+				{
+					// A cycle equal to the tested exponent (which is prime in this path) guarantees that the candidate divides
+					// the corresponding Mersenne number because the order of 2 modulo the divisor is exactly p.
+					processedAll = true;
+					PrimeOrderCalculatorAccelerator.Return(gpu);
+					return true;
+				}
+
+				if (divisorCycle == 0UL)
+				{
+					Console.WriteLine($"Divisor cycle was not calculated for {prime}");
+				}
+			}
+
+			divisor.Add(step);
+			remainder10 = AddMod10(remainder10, step10);
+			remainder8 = AddMod8(remainder8, step8);
+			remainder3 = AddMod3(remainder3, step3);
+			remainder7 = AddMod7(remainder7, step7);
+			remainder11 = AddMod11(remainder11, step11);
+		}
+
+		PrimeOrderCalculatorAccelerator.Return(gpu);
+		processedAll = true;
+		return false;
+	}
+
+	private static bool CheckDivisors64(
+		ulong prime,
+		ulong step,
+		ulong limit,
+		ulong divisor,
+		ushort decimalMask,
+		byte step10,
+		byte step8,
+		byte step3,
+		byte step7,
+		byte step11,
+		byte remainder10,
+		byte remainder8,
+		byte remainder3,
+		byte remainder7,
+		byte remainder11,
+		out bool processedAll)
+	{
+		processedAll = false;
+
+		bool canAdvance = step <= limit;
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+		if (!canAdvance)
+		{
+			if (divisor <= limit)
+			{
+				if (remainder3 != 0 && remainder7 != 0 && remainder11 != 0 && (remainder8 == 1 || remainder8 == 7) && ((decimalMask >> remainder10) & 1) != 0)
+				{
+					MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
+					ulong divisorCycle;
+					if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
+							gpu,
+							divisor,
+							prime,
+							divisorData,
+							out ulong computedCycle,
+							out bool primeOrderFailed) || computedCycle == 0UL)
+					{
+						divisorCycle = MersenneDivisorCycles.CalculateCycleLength(
+							divisor,
+							divisorData,
+							skipPrimeOrderHeuristic: primeOrderFailed);
+					}
+					else
+					{
+						divisorCycle = computedCycle;
+					}
+
+					if (divisorCycle == prime)
+					{
+						processedAll = true;
+						PrimeOrderCalculatorAccelerator.Return(gpu);
+						return true;
+					}
+
+					if (divisorCycle == 0UL)
+					{
+						Console.WriteLine($"Divisor cycle was not calculated for {prime}");
+					}
+				}
+			}
+
+			processedAll = true;
+			return false;
+		}
+
+		if (divisor > limit)
+		{
+			processedAll = true;
+			return false;
+		}
+
+		ulong remainingIterations = ((limit - divisor) / step) + 1UL;
+
+		while (true)
+		{
+			bool passesSmallModuli = remainder3 != 0 && remainder7 != 0 && remainder11 != 0;
+			if (passesSmallModuli && (remainder8 == 1 || remainder8 == 7) && ((decimalMask >> remainder10) & 1) != 0)
+			{
+				MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
+				ulong divisorCycle;
+				if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
+						gpu,
+						divisor,
+						prime,
+						divisorData,
+						out ulong computedCycle,
+						out bool primeOrderFailed) || computedCycle == 0UL)
+				{
+					divisorCycle = MersenneDivisorCycles.CalculateCycleLength(
+						divisor,
+						divisorData,
+						skipPrimeOrderHeuristic: primeOrderFailed);
+				}
+				else
+				{
+					divisorCycle = computedCycle;
+				}
+
+				if (divisorCycle == prime)
+				{
+					processedAll = true;
+					PrimeOrderCalculatorAccelerator.Return(gpu);
+					return true;
+				}
+
+				if (divisorCycle == 0UL)
+				{
+					Console.WriteLine($"Divisor cycle was not calculated for {prime}");
+				}
+			}
+
+			remainingIterations--;
+			if (remainingIterations == 0UL)
+			{
+				processedAll = true;
+				return false;
+			}
+
+			divisor += step;
+			remainder10 = AddMod10(remainder10, step10);
+			remainder8 = AddMod8(remainder8, step8);
+			remainder3 = AddMod3(remainder3, step3);
+			remainder7 = AddMod7(remainder7, step7);
+			remainder11 = AddMod11(remainder11, step11);
+		}
+	}
+
+	private static byte CheckDivisor(ulong prime, ulong divisorCycle, in MontgomeryDivisorData divisorData)
+	{
+		ulong residue = prime.Pow2MontgomeryModWithCycleCpu(divisorCycle, divisorData);
+		return residue == 1UL ? (byte)1 : (byte)0;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static byte AddMod3(byte value, byte delta)
+	{
+		const int Modulus = 3;
+		int sum = value + delta;
+
+		if (sum >= Modulus)
+		{
+			sum -= Modulus;
+		}
+
+		return (byte)sum;
+	}
+
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static byte AddMod7(byte value, byte delta)
+	{
+		const int Modulus = 7;
+		int sum = value + delta;
+
+		if (sum >= Modulus)
+		{
+			sum -= Modulus;
+		}
+
+		return (byte)sum;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static byte AddMod10(byte value, byte delta)
+	{
+		const int Modulus = 10;
+		int sum = value + delta;
+
+		if (sum >= Modulus)
+		{
+			sum -= Modulus;
+		}
+
+		return (byte)sum;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static byte AddMod11(byte value, byte delta)
+	{
+		const int Modulus = 11;
+		int sum = value + delta;
+
+		if (sum >= Modulus)
+		{
+			sum -= Modulus;
+		}
+
+		return (byte)sum;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static byte AddMod8(byte value, byte delta)
+	{
+		return (byte)((value + delta) & 7);
+	}
+
+	private static ulong ComputeDivisorLimitFromMaxPrime(ulong maxPrime)
+	{
+		// The by-divisor CPU configuration only supplies primes greater than 1, so the guard below never trips.
+		// if (maxPrime <= 1UL)
+		// {
+		//     return 0UL;
+		// }
+		if (maxPrime - 1UL >= 64UL)
+		{
+			return ulong.MaxValue;
+		}
+
+		return (1UL << (int)(maxPrime - 1UL)) - 1UL;
+	}
+
+	private static ulong ComputeAllowedMaxDivisor(ulong prime, ulong divisorLimit)
+	{
+		// Production by-divisor scans only handle primes, so inputs never fall below 2.
+		// if (prime <= 1UL)
+		// {
+		//     return 0UL;
+		// }
+		if (prime - 1UL >= 64UL)
+		{
+			return divisorLimit;
+		}
+
+		return Math.Min((1UL << (int)(prime - 1UL)) - 1UL, divisorLimit);
+	}
 
 
 }

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -1,329 +1,377 @@
-using System.Buffers;
-using System.Collections.Concurrent;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
 using ILGPU;
 using ILGPU.Runtime;
 using PerfectNumbers.Core.Gpu;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core;
 
 public sealed class DivisorCycleCache
 {
-    private const byte ByteZero = 0;
-    private const byte ByteOne = 1;
-    private const int StackBufferThreshold = 128;
+	private const byte ByteZero = 0;
+	private const byte ByteOne = 1;
+	private const int StackBufferThreshold = 128;
+	private readonly Accelerator _accelerator;
+	private readonly int _acceleratorIndex;
+	private readonly Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> _gpuKernel;
+	private volatile ulong[] _snapshot;
+	private volatile bool _useGpuGeneration = true;
+	private readonly int _divisorCyclesBatchSize;
+	private static int _sharedDivisorCyclesBatchSize = GpuConstants.GpuCycleStepsPerInvocation;
 
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>> _gpuKernelCache = new(AcceleratorReferenceComparer.Instance);
-    private volatile ulong[] _snapshot;
-    private volatile bool _useGpuGeneration = true;
-    private readonly int _divisorCyclesBatchSize;
-    private static int _sharedDivisorCyclesBatchSize = GpuConstants.GpuCycleStepsPerInvocation;
+	private const int CycleCacheTrackingLimit = 500_000;
+	private static readonly HashSet<ulong> CycleCacheTrackedDivisors = new();
+	private static readonly object CycleCacheTrackingLock = new();
+	private static ulong CycleCacheHitCount;
+	private static bool CycleCacheTrackingDisabled;
 
-    private const int CycleCacheTrackingLimit = 500_000;
-    private static readonly HashSet<ulong> CycleCacheTrackedDivisors = new();
-    private static readonly object CycleCacheTrackingLock = new();
-    private static ulong CycleCacheHitCount;
-    private static bool CycleCacheTrackingDisabled;
+	public static void SetDivisorCyclesBatchSize(int divisorCyclesBatchSize)
+	{
+		_sharedDivisorCyclesBatchSize = Math.Max(1, divisorCyclesBatchSize);
+	}
 
-    public static void SetDivisorCyclesBatchSize(int divisorCyclesBatchSize)
-    {
-        _sharedDivisorCyclesBatchSize = Math.Max(1, divisorCyclesBatchSize);
-    }
+	public static DivisorCycleCache Shared { get; } = new DivisorCycleCache(_sharedDivisorCyclesBatchSize);
 
-    public static DivisorCycleCache Shared { get; } = new DivisorCycleCache(_sharedDivisorCyclesBatchSize);
+	public int PreferredBatchSize => _divisorCyclesBatchSize;
 
-    public int PreferredBatchSize => _divisorCyclesBatchSize;
+	// private static readonly Accelerator[] _accelerators;
 
-    private DivisorCycleCache(int divisorCyclesBatchSize)
-    {
-        _divisorCyclesBatchSize = Math.Max(1, divisorCyclesBatchSize);
-        _snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
-    }
+	private DivisorCycleCache(int divisorCyclesBatchSize)
+	{
+		_divisorCyclesBatchSize = Math.Max(1, divisorCyclesBatchSize);
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		_acceleratorIndex = acceleratorIndex;
+		Accelerator accelerator = AcceleratorPool.Shared.Accelerators[acceleratorIndex];
+		_accelerator = accelerator;
+		_gpuKernel = LoadKernel(accelerator);
+		_snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
+	}
 
-    public void ConfigureGeneratorDevice(bool useGpu)
-    {
-        _useGpuGeneration = useGpu;
-    }
+	public void ConfigureGeneratorDevice(bool useGpu)
+	{
+		_useGpuGeneration = useGpu;
+	}
 
-    public void RefreshSnapshot()
-    {
-        _snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
-    }
+	public void RefreshSnapshot()
+	{
+		_snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
+	}
 
-    public ulong GetCycleLength(ulong divisor, bool skipPrimeOrderHeuristic = false)
-    {
-        if (!skipPrimeOrderHeuristic)
-        {
-            Span<ulong> result = stackalloc ulong[1];
-            ReadOnlySpan<ulong> singleDivisor = stackalloc ulong[1] { divisor };
-            GetCycleLengths(singleDivisor, result);
-            return result[0];
-        }
+	public ulong GetCycleLength(ulong divisor, bool skipPrimeOrderHeuristic = false)
+	{
+		if (!skipPrimeOrderHeuristic)
+		{
+			Span<ulong> result = stackalloc ulong[1];
+			ReadOnlySpan<ulong> singleDivisor = stackalloc ulong[1] { divisor };
+			GetCycleLengths(singleDivisor, result);
+			return result[0];
+		}
 
-        // The skipPrimeOrderHeuristic path is only used for divisors produced by the CPU by-divisor scanner,
-        // which always generates values greater than one. Keep this guard disabled here to avoid repeating the
-        // validation already performed by the general GetCycleLengths path.
-        // if (divisor <= 1UL)
-        // {
-        //     throw new InvalidDataException("Divisor must be > 1");
-        // }
+		// The skipPrimeOrderHeuristic path is only used for divisors produced by the CPU by-divisor scanner,
+		// which always generates values greater than one. Keep this guard disabled here to avoid repeating the
+		// validation already performed by the general GetCycleLengths path.
+		// if (divisor <= 1UL)
+		// {
+		//     throw new InvalidDataException("Divisor must be > 1");
+		// }
 
-        ulong[] snapshot = _snapshot;
-        if (divisor < (ulong)snapshot.Length)
-        {
-            ulong cached = snapshot[divisor];
-            if (cached == 0UL)
-            {
-                throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
-            }
+		ulong[] snapshot = _snapshot;
+		if (divisor < (ulong)snapshot.Length)
+		{
+			ulong cached = snapshot[divisor];
+			if (cached == 0UL)
+			{
+				throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
+			}
 
-            // ObserveCycleCacheDivisor(divisor);
-            return cached;
-        }
+			return cached;
+		}
 
-        MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
-        return MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: true);
-    }
+		MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
+		return MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: true);
+	}
 
-    public void GetCycleLengths(ReadOnlySpan<ulong> divisors, Span<ulong> cycles)
-    {
-        ulong[] snapshot = _snapshot;
-        int length = divisors.Length;
-        int[]? rentedMissing = null;
-        Span<int> missingBuffer = length <= StackBufferThreshold
-            ? stackalloc int[length]
-            : new Span<int>(rentedMissing = ArrayPool<int>.Shared.Rent(length), 0, length);
+	public void GetCycleLengths(ReadOnlySpan<ulong> divisors, Span<ulong> cycles)
+	{
+		ulong[] snapshot = _snapshot;
+		int length = divisors.Length;
+		int[]? rentedMissing = null;
+		Span<int> missingBuffer = length <= StackBufferThreshold
+			? stackalloc int[length]
+			: new Span<int>(rentedMissing = ThreadStaticPools.IntPool.Rent(length), 0, length);
 
-        int missingCount = 0;
+		int missingCount = 0;
 
-        for (int i = 0; i < length; i++)
-        {
-            ulong divisor = divisors[i];
-            if (divisor <= 1UL)
-            {
-                throw new InvalidDataException("Divisor must be > 1");
-            }
+		for (int i = 0; i < length; i++)
+		{
+			ulong divisor = divisors[i];
+			if (divisor <= 1UL)
+			{
+				throw new InvalidDataException("Divisor must be > 1");
+			}
 
-            if (divisor < (ulong)snapshot.Length)
-            {
-                ulong cached = snapshot[divisor];
-                if (cached == 0UL)
-                {
-                    throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
-                }
+			if (divisor < (ulong)snapshot.Length)
+			{
+				ulong cached = snapshot[divisor];
+				if (cached == 0UL)
+				{
+					throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
+				}
 
-                // ObserveCycleCacheDivisor(divisor);
-                cycles[i] = cached;
-            }
-            else
-            {
-                missingBuffer[missingCount++] = i;
-            }
-        }
+				cycles[i] = cached;
+			}
+			else
+			{
+				missingBuffer[missingCount++] = i;
+			}
+		}
 
-        if (missingCount != 0)
-        {
-            ReadOnlySpan<int> missingIndices = missingBuffer[..missingCount];
-            if (_useGpuGeneration)
-            {
-                ComputeCyclesGpu(divisors, cycles, missingIndices);
-            }
-            else
-            {
-                ComputeCyclesCpu(divisors, cycles, missingIndices);
-            }
-        }
+		if (missingCount != 0)
+		{
+			ReadOnlySpan<int> missingIndices = missingBuffer[..missingCount];
+			if (_useGpuGeneration)
+			{
+				ComputeCyclesGpu(divisors, cycles, missingIndices);
+			}
+			else
+			{
+				ComputeCyclesCpu(divisors, cycles, missingIndices);
+			}
+		}
 
-        if (rentedMissing is not null)
-        {
-            ArrayPool<int>.Shared.Return(rentedMissing, clearArray: false);
-        }
-    }
+		if (rentedMissing is not null)
+		{
+			ThreadStaticPools.IntPool.Return(rentedMissing, clearArray: false);
+		}
+	}
 
-    private static void ObserveCycleCacheDivisor(ulong divisor)
-    {
-        bool logHit = false;
-        ulong hitNumber = 0UL;
-        bool trackingDisabledNow = false;
+	private static void ObserveCycleCacheDivisor(ulong divisor)
+	{
+		bool logHit = false;
+		ulong hitNumber = 0UL;
+		bool trackingDisabledNow = false;
 
-        lock (CycleCacheTrackingLock)
-        {
-            if (CycleCacheTrackingDisabled)
-            {
-                return;
-            }
+		lock (CycleCacheTrackingLock)
+		{
+			if (CycleCacheTrackingDisabled)
+			{
+				return;
+			}
 
-            if (!CycleCacheTrackedDivisors.Add(divisor))
-            {
-                CycleCacheHitCount++;
-                hitNumber = CycleCacheHitCount;
-                logHit = true;
-            }
-            else if (CycleCacheTrackedDivisors.Count >= CycleCacheTrackingLimit)
-            {
-                CycleCacheTrackingDisabled = true;
-                CycleCacheTrackedDivisors.Clear();
-                trackingDisabledNow = true;
-            }
-        }
+			if (!CycleCacheTrackedDivisors.Add(divisor))
+			{
+				CycleCacheHitCount++;
+				hitNumber = CycleCacheHitCount;
+				logHit = true;
+			}
+			else if (CycleCacheTrackedDivisors.Count >= CycleCacheTrackingLimit)
+			{
+				CycleCacheTrackingDisabled = true;
+				CycleCacheTrackedDivisors.Clear();
+				trackingDisabledNow = true;
+			}
+		}
 
-        if (logHit)
-        {
-            Console.WriteLine($"Cycle cache hit for divisor {divisor} ({hitNumber})");
-        }
-        else if (trackingDisabledNow)
-        {
-            Console.WriteLine("Cycle cache hit tracking disabled after exceeding the tracking limit.");
-        }
-    }
+		if (logHit)
+		{
+			Console.WriteLine($"Cycle cache hit for divisor {divisor} ({hitNumber})");
+		}
+		else if (trackingDisabledNow)
+		{
+			Console.WriteLine("Cycle cache hit tracking disabled after exceeding the tracking limit.");
+		}
+	}
 
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static void ComputeCyclesCpu(ReadOnlySpan<ulong> divisors, Span<ulong> cycles, ReadOnlySpan<int> indices)
-    {
-        for (int i = 0; i < indices.Length; i++)
-        {
-            int targetIndex = indices[i];
-            ulong divisor = divisors[targetIndex];
-            MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
-            cycles[targetIndex] = MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData);
-        }
-    }
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static void ComputeCyclesCpu(ReadOnlySpan<ulong> divisors, Span<ulong> cycles, ReadOnlySpan<int> indices)
+	{
+		for (int i = 0; i < indices.Length; i++)
+		{
+			int targetIndex = indices[i];
+			ulong divisor = divisors[targetIndex];
+			MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
+			cycles[targetIndex] = MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData);
+		}
+	}
 
-    private void ComputeCyclesGpu(ReadOnlySpan<ulong> divisors, Span<ulong> cycles, ReadOnlySpan<int> indices)
-    {
-        int count = indices.Length;
-        ulong[]? rentedDivisors = null;
-        Span<ulong> workDivisors = count <= StackBufferThreshold
-            ? stackalloc ulong[count]
-            : new Span<ulong>(rentedDivisors = ArrayPool<ulong>.Shared.Rent(count), 0, count);
+	private void ComputeCyclesGpu(ReadOnlySpan<ulong> divisors, Span<ulong> cycles, ReadOnlySpan<int> indices)
+	{
+		int count = indices.Length;
+		ulong[]? rentedDivisors = null;
+		Span<ulong> workDivisors = count <= StackBufferThreshold
+			? stackalloc ulong[count]
+			: new Span<ulong>(rentedDivisors = ThreadStaticPools.UlongPool.Rent(count), 0, count);
 
-        ulong[]? rentedResults = null;
-        Span<ulong> workResults = count <= StackBufferThreshold
-            ? stackalloc ulong[count]
-            : new Span<ulong>(rentedResults = ArrayPool<ulong>.Shared.Rent(count), 0, count);
+		ulong[]? rentedResults = null;
+		Span<ulong> workResults = count <= StackBufferThreshold
+			? stackalloc ulong[count]
+			: new Span<ulong>(rentedResults = ThreadStaticPools.UlongPool.Rent(count), 0, count);
 
-        for (int i = 0; i < count; i++)
-        {
-            workDivisors[i] = divisors[indices[i]];
-        }
+		for (int i = 0; i < count; i++)
+		{
+			workDivisors[i] = divisors[indices[i]];
+		}
 
-        ComputeCyclesGpuCore(workDivisors, workResults);
+		ComputeCyclesGpuCore(workDivisors, workResults);
 
-        for (int i = 0; i < count; i++)
-        {
-            cycles[indices[i]] = workResults[i];
-        }
+		for (int i = 0; i < count; i++)
+		{
+			cycles[indices[i]] = workResults[i];
+		}
 
-        if (rentedDivisors is not null)
-        {
-            ArrayPool<ulong>.Shared.Return(rentedDivisors, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedResults!, clearArray: false);
-        }
-    }
+		if (rentedDivisors is not null)
+		{
+			ThreadStaticPools.UlongPool.Return(rentedDivisors, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedResults!, clearArray: false);
+		}
+	}
 
-    private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
-    {
-        int length = divisors.Length;
-        var gpuLease = GpuKernelPool.GetKernel(useGpuOrder: true);
+	private const int DefaultCapacity = 128;
 
-        Accelerator accelerator = gpuLease.Accelerator;
-        var kernel = _gpuKernelCache.GetOrAdd(accelerator, LoadKernel);
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? divisorBuffer;
 
-        using MemoryBuffer1D<ulong, Stride1D.Dense> divisorBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<ulong, Stride1D.Dense> powBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<ulong, Stride1D.Dense> orderBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<byte, Stride1D.Dense> statusBuffer = accelerator.Allocate1D<byte>(length);
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? powBuffer;
 
-        ulong[]? rentedPow = null;
-        Span<ulong> powSpan = length <= StackBufferThreshold
-            ? stackalloc ulong[length]
-            : new Span<ulong>(rentedPow = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? orderBuffer;
 
-        ulong[]? rentedOrder = null;
-        Span<ulong> orderSpan = length <= StackBufferThreshold
-            ? stackalloc ulong[length]
-            : new Span<ulong>(rentedOrder = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? resultBuffer;
 
-        ulong[]? rentedResult = null;
-        Span<ulong> resultSpan = length <= StackBufferThreshold
-            ? stackalloc ulong[length]
-            : new Span<ulong>(rentedResult = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+	[ThreadStatic]
+	private static MemoryBuffer1D<byte, Stride1D.Dense>? statusBuffer;
 
-        byte[]? rentedStatus = null;
-        Span<byte> statusSpan = length <= StackBufferThreshold
-            ? stackalloc byte[length]
-            : new Span<byte>(rentedStatus = ArrayPool<byte>.Shared.Rent(length), 0, length);
+	private void EnsureCapacity(int requiredCapacity)
+	{
+		if (requiredCapacity <= (divisorBuffer?.Length ?? 0))
+		{
+			return;
+		}
 
-        ref ulong divisorRef = ref MemoryMarshal.GetReference(divisors);
-        divisorBuffer.View.CopyFromCPU(ref divisorRef, length);
+		if (divisorBuffer != null)
+		{
+			divisorBuffer.Dispose();
+			powBuffer!.Dispose();
+			orderBuffer!.Dispose();
+			resultBuffer!.Dispose();
+			statusBuffer!.Dispose();
+		}
 
-        for (int i = 0; i < length; i++)
-        {
-            ulong divisor = divisors[i];
-            ulong initialPow = 2UL;
-            if (initialPow >= divisor)
-            {
-                initialPow -= divisor;
-            }
+		Accelerator accelerator = _accelerator;
+		lock(accelerator)
+		{
+			divisorBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+			powBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+			orderBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+			resultBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+			statusBuffer = accelerator.Allocate1D<byte>(requiredCapacity);
+		}
+	}
 
-            powSpan[i] = initialPow;
-            orderSpan[i] = 1UL;
-            resultSpan[i] = 0UL;
-            statusSpan[i] = ByteZero;
-        }
+	private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
+	{
+		int length = divisors.Length;
 
-        powBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(powSpan), length);
-        orderBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(orderSpan), length);
-        resultBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(resultSpan), length);
-        statusBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(statusSpan), length);
+		EnsureCapacity(length);
 
-        int pending;
-        do
-        {
-            kernel(length, _divisorCyclesBatchSize, divisorBuffer.View, powBuffer.View, orderBuffer.View, resultBuffer.View, statusBuffer.View);
-            accelerator.Synchronize();
+		ulong[]? rentedPow = null;
+		Span<ulong> powSpan = length <= StackBufferThreshold
+			? stackalloc ulong[length]
+			: new Span<ulong>(rentedPow = ThreadStaticPools.UlongPool.Rent(length), 0, length);
 
-            statusBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(statusSpan), length);
-            pending = 0;
-            for (int i = 0; i < length; i++)
-            {
-                if (statusSpan[i] == ByteZero)
-                {
-                    pending++;
-                }
-            }
-        }
-        while (pending > 0);
+		ulong[]? rentedOrder = null;
+		Span<ulong> orderSpan = length <= StackBufferThreshold
+			? stackalloc ulong[length]
+			: new Span<ulong>(rentedOrder = ThreadStaticPools.UlongPool.Rent(length), 0, length);
 
-        resultBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(resultSpan), length);
-        resultSpan.CopyTo(destination);
-        if (rentedPow is not null)
-        {
-            ArrayPool<ulong>.Shared.Return(rentedPow, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedOrder!, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedResult!, clearArray: false);
-            ArrayPool<byte>.Shared.Return(rentedStatus!, clearArray: false);
-        }
+		ulong[]? rentedResult = null;
+		Span<ulong> resultSpan = length <= StackBufferThreshold
+			? stackalloc ulong[length]
+			: new Span<ulong>(rentedResult = ThreadStaticPools.UlongPool.Rent(length), 0, length);
 
-        gpuLease.Dispose();
-    }
+		byte[]? rentedStatus = null;
+		Span<byte> statusSpan = length <= StackBufferThreshold
+			? stackalloc byte[length]
+			: new Span<byte>(rentedStatus = ThreadStaticPools.BytePool.Rent(length), 0, length);
 
-    private static Action<Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> LoadKernel(Accelerator accelerator)
-    {
-        return accelerator.LoadAutoGroupedStreamKernel<Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>(DivisorCycleKernels.GpuAdvanceDivisorCyclesKernel);
-    }
+		for (int i = 0; i < length; i++)
+		{
+			ulong divisor = divisors[i];
+			ulong initialPow = 2UL;
+			if (initialPow >= divisor)
+			{
+				initialPow -= divisor;
+			}
 
-    private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
-    {
-        internal static AcceleratorReferenceComparer Instance { get; } = new();
+			powSpan[i] = initialPow;
+			orderSpan[i] = 1UL;
+			resultSpan[i] = 0UL;
+			statusSpan[i] = ByteZero;
+		}
 
-        public bool Equals(Accelerator? x, Accelerator? y) => ReferenceEquals(x, y);
+		// GpuPrimeWorkLimiter.Acquire();
+		int pending;
+		var kernel = _gpuKernel;
+		var acceleratorIndex = _acceleratorIndex;
+		var stream = AcceleratorStreamPool.Rent(acceleratorIndex);
 
-        public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
-    }
+		divisorBuffer!.View.CopyFromCPU(stream, divisors);
+		powBuffer!.View.CopyFromCPU(stream, powSpan);
+		orderBuffer!.View.CopyFromCPU(stream, orderSpan);
+		resultBuffer!.View.CopyFromCPU(stream, resultSpan);
+		statusBuffer!.View.CopyFromCPU(stream, statusSpan);
+
+		do
+		{
+			kernel(stream, length, _divisorCyclesBatchSize, divisorBuffer.View, powBuffer.View, orderBuffer.View, resultBuffer.View, statusBuffer.View);
+
+			statusBuffer.View.CopyToCPU(stream, statusSpan);
+			stream.Synchronize();
+
+			pending = 0;
+			for (int i = 0; i < length; i++)
+			{
+				if (statusSpan[i] == ByteZero)
+				{
+					pending++;
+				}
+			}
+		}
+		while (pending > 0);
+
+		resultBuffer.View.CopyToCPU(stream, resultSpan);
+		stream.Synchronize();
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
+
+		resultSpan.CopyTo(destination);
+		if (rentedPow is not null)
+		{
+			ThreadStaticPools.UlongPool.Return(rentedPow, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedOrder!, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedResult!, clearArray: false);
+			ThreadStaticPools.BytePool.Return(rentedStatus!, clearArray: false);
+		}
+
+		// GpuPrimeWorkLimiter.Release();
+	}
+
+	private static Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> LoadKernel(Accelerator accelerator)
+	{
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>(DivisorCycleKernels.GpuAdvanceDivisorCyclesKernel);
+		var kernel = KernelUtil.GetKernel(loaded);
+		return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>>();
+	}
+
+	private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
+	{
+		internal static AcceleratorReferenceComparer Instance { get; } = new();
+
+		public bool Equals(Accelerator? x, Accelerator? y) => ReferenceEquals(x, y);
+
+		public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
+	}
 }
 
 

PerfectNumbers.Core/DivisorGenerator.cs

@@ -1,5 +1,3 @@
-using System;
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
 
 namespace PerfectNumbers.Core;

PerfectNumbers.Core/Gpu/AcceleratorPool.cs

@@ -0,0 +1,61 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu
+{
+	public struct AcceleratorPool
+	{
+		internal readonly Context Context;
+		internal readonly Device Device;
+		internal readonly Accelerator[] Accelerators;
+
+		private int _index;
+
+		public int Rent()
+		{
+			Atomic.CompareExchange(ref _index, Capacity, 0);
+			var index = Atomic.Add(ref _index, 1);
+			if (index >= Capacity)
+			{
+				index -= Capacity;
+			}
+
+			return index;
+		}
+
+		public void Return(int acceleratorIndex)
+		{
+			// Intentionally left empty - there's nothing to do here. We're not really renting anything.
+		}
+
+		public static readonly AcceleratorPool Shared = new(PerfectNumberConstants.RollingAccelerators);
+
+		private readonly int Capacity;
+
+		public AcceleratorPool(int capacity)
+		{
+			Console.WriteLine("Creating accelerators");
+			Capacity = capacity;
+			var context = Context.CreateDefault();
+			var device = context.GetPreferredDevice(false);
+			Context = context;
+			Device = device;
+			int acceleratorsInContext = 0;
+			int acceleratorIndex = 0;
+			Accelerators = [.. Enumerable.Range(0, capacity).Select(_ =>
+			{
+				Console.WriteLine($"Preparing accelerator {acceleratorIndex++}");
+				Accelerator accelerator = device.CreateAccelerator(context);
+				acceleratorsInContext++;
+				if (acceleratorsInContext >= 128)
+				{
+					context = Context.CreateDefault();
+					device = context.GetPreferredDevice(false);
+					acceleratorsInContext = 0;
+				}
+
+				return accelerator;
+			})];
+		}
+	}
+}

PerfectNumbers.Core/Gpu/AcceleratorReferenceComparer.cs

@@ -0,0 +1,13 @@
+using System.Runtime.CompilerServices;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu;
+
+internal sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
+{
+	internal static AcceleratorReferenceComparer Instance { get; } = new();
+
+	public bool Equals(Accelerator? x, Accelerator? y) => ReferenceEquals(x, y);
+
+	public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
+}

PerfectNumbers.Core/Gpu/Accelerators/AcceleratorExtensions.cs

@@ -0,0 +1,34 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+public static class AcceleratorExtensions
+{
+	public static MemoryBuffer1D<ulong, Stride1D.Dense> CopySpanToDevice(this Accelerator accelerator, AcceleratorStream stream, in ulong[] span)
+	{
+		MemoryBuffer1D<ulong, Stride1D.Dense>? buffer;
+		buffer = accelerator.Allocate1D(stream, span);
+
+		return buffer;
+	}
+
+	public static MemoryBuffer1D<ulong, Stride1D.Dense> CopyUintSpanToDevice(this Accelerator accelerator, AcceleratorStream stream, ReadOnlySpan<uint> span)
+	{
+		MemoryBuffer1D<ulong, Stride1D.Dense>? buffer;
+		buffer = accelerator.Allocate1D<ulong>(span.Length);
+
+		if (!span.IsEmpty)
+		{
+			var converted = new ulong[span.Length];
+			for (int i = 0; i < span.Length; i++)
+			{
+				converted[i] = span[i];
+			}
+
+			buffer.View.CopyFromCPU(stream, converted);
+		}
+
+		return buffer;
+	}
+}

PerfectNumbers.Core/Gpu/Accelerators/AcceleratorStreamPool.cs

@@ -0,0 +1,42 @@
+using System.Collections.Concurrent;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+public static class AcceleratorStreamPool
+{
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+	private static readonly SemaphoreSlim[] _locks = new SemaphoreSlim[PerfectNumberConstants.RollingAccelerators];
+	private static readonly ConcurrentQueue<AcceleratorStream>[] _streams = new ConcurrentQueue<AcceleratorStream> [PerfectNumberConstants.RollingAccelerators];
+
+	private static SemaphoreSlim CreateLock() => new(PerfectNumberConstants.ThreadsByAccelerator);
+
+
+	// private static int _rented;
+
+	public static AcceleratorStream Rent(int acceleratorIndex)
+	{
+		var streamLock = _locks[acceleratorIndex];
+		var queue = _streams[acceleratorIndex];
+
+		streamLock.Wait();
+
+		return queue.TryDequeue(out var stream)
+			? stream
+			: _accelerators[acceleratorIndex].CreateStream();
+	}
+
+	public static void Return(int acceleratorIndex, AcceleratorStream stream)
+	{
+		var streamLock = _locks[acceleratorIndex];
+		var queue = _streams[acceleratorIndex];
+		queue.Enqueue(stream);
+		streamLock.Release();
+	}
+
+	public static void WarmUp(int acceleratorIndex)
+	{
+		_streams[acceleratorIndex] = new();
+		_locks[acceleratorIndex] = CreateLock();
+	}
+}

PerfectNumbers.Core/Gpu/Accelerators/HeuristicCombinedGpuTables.cs

@@ -0,0 +1,69 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+internal readonly struct HeuristicCombinedGpuTables
+{
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
+	private static readonly HeuristicCombinedGpuTables[] _sharedTables = new HeuristicCombinedGpuTables[PerfectNumberConstants.RollingAccelerators];
+
+	internal static void WarmUp(int acceleratorIndex, AcceleratorStream stream) => _sharedTables[acceleratorIndex] = new(_accelerators[acceleratorIndex], stream);
+
+	internal static HeuristicCombinedGpuTables GetStaticTables(int acceleratorIndex) => _sharedTables[acceleratorIndex];
+
+	internal HeuristicCombinedGpuTables(Accelerator accelerator, AcceleratorStream stream)
+	{
+		var combinedEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1;
+		HeuristicCombinedDivisorsEnding1 = accelerator.CopySpanToDevice(stream, combinedEnding1);
+		var combinedSquaresEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1Squares;
+		HeuristicCombinedDivisorSquaresEnding1 = accelerator.CopySpanToDevice(stream, combinedSquaresEnding1);
+
+		var combinedEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3;
+		HeuristicCombinedDivisorsEnding3 = accelerator.CopySpanToDevice(stream, combinedEnding3);
+		var combinedSquaresEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3Squares;
+		HeuristicCombinedDivisorSquaresEnding3 = accelerator.CopySpanToDevice(stream, combinedSquaresEnding3);
+
+		var combinedEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7;
+		HeuristicCombinedDivisorsEnding7 = accelerator.CopySpanToDevice(stream, combinedEnding7);
+		var combinedSquaresEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7Squares;
+		HeuristicCombinedDivisorSquaresEnding7 = accelerator.CopySpanToDevice(stream, combinedSquaresEnding7);
+
+		var combinedEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9;
+		HeuristicCombinedDivisorsEnding9 = accelerator.CopySpanToDevice(stream, combinedEnding9);
+		var combinedSquaresEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9Squares;
+		HeuristicCombinedDivisorSquaresEnding9 = accelerator.CopySpanToDevice(stream, combinedSquaresEnding9);
+	}
+
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding1;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding1;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding3;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding3;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding7;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding7;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding9;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding9;
+
+	internal HeuristicCombinedGpuViews CreateViews() => new(
+			HeuristicCombinedDivisorsEnding1.View,
+			HeuristicCombinedDivisorsEnding3.View,
+			HeuristicCombinedDivisorsEnding7.View,
+			HeuristicCombinedDivisorsEnding9.View,
+			HeuristicCombinedDivisorSquaresEnding1.View,
+			HeuristicCombinedDivisorSquaresEnding3.View,
+			HeuristicCombinedDivisorSquaresEnding7.View,
+			HeuristicCombinedDivisorSquaresEnding9.View);
+
+	internal void Dispose()
+	{
+		HeuristicCombinedDivisorsEnding1.Dispose();
+		HeuristicCombinedDivisorSquaresEnding1.Dispose();
+		HeuristicCombinedDivisorsEnding3.Dispose();
+		HeuristicCombinedDivisorSquaresEnding3.Dispose();
+		HeuristicCombinedDivisorsEnding7.Dispose();
+		HeuristicCombinedDivisorSquaresEnding7.Dispose();
+		HeuristicCombinedDivisorsEnding9.Dispose();
+		HeuristicCombinedDivisorSquaresEnding9.Dispose();
+	}
+}

PerfectNumbers.Core/Gpu/Accelerators/HeuristicCombinedGpuViews.cs

@@ -0,0 +1,45 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+public readonly struct HeuristicCombinedGpuViews(
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding1,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding3,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding7,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding9,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding1,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding3,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding7,
+		ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding9)
+{
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding1 = combinedDivisorsEnding1;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding3 = combinedDivisorsEnding3;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding7 = combinedDivisorsEnding7;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding9 = combinedDivisorsEnding9;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding1 = combinedDivisorSquaresEnding1;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding3 = combinedDivisorSquaresEnding3;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding7 = combinedDivisorSquaresEnding7;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding9 = combinedDivisorSquaresEnding9;
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ArrayView1D<ulong, Stride1D.Dense> SelectDivisors(byte ending) => ending switch
+	{
+		1 => CombinedDivisorsEnding1,
+		3 => CombinedDivisorsEnding3,
+		7 => CombinedDivisorsEnding7,
+		9 => CombinedDivisorsEnding9,
+		_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
+	};
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ArrayView1D<ulong, Stride1D.Dense> SelectDivisorSquares(byte ending) => ending switch
+	{
+		1 => CombinedDivisorSquaresEnding1,
+		3 => CombinedDivisorSquaresEnding3,
+		7 => CombinedDivisorSquaresEnding7,
+		9 => CombinedDivisorSquaresEnding9,
+		_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
+	};
+}

PerfectNumbers.Core/Gpu/Accelerators/HeuristicGroupABGpuTables.cs

@@ -0,0 +1,79 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+internal sealed class HeuristicGroupABGpuTables
+{
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
+	private static readonly HeuristicGroupABGpuTables[] _sharedTables = new HeuristicGroupABGpuTables[PerfectNumberConstants.RollingAccelerators];
+
+	private static readonly SemaphoreSlim[] _locks = [..Enumerable.Range(1, PerfectNumberConstants.RollingAccelerators).Select(_ => new SemaphoreSlim(1))];
+
+	internal static HeuristicGroupABGpuTables EnsureStaticTables(int acceleratorIndex, AcceleratorStream stream)
+	{
+		var @lock = _locks[acceleratorIndex];
+		@lock.Wait();
+		if (_sharedTables[acceleratorIndex] is {} existing)
+		{
+			@lock.Release();
+			return existing;
+		}
+
+		var accelerator = _accelerators[acceleratorIndex];
+		existing = new(accelerator, stream);
+		_sharedTables[acceleratorIndex] = existing;
+		@lock.Release();
+		return existing;
+	}
+
+	internal HeuristicGroupABGpuTables(Accelerator accelerator, AcceleratorStream stream)
+	{
+		var heuristicGroupA = HeuristicPrimeSieves.GroupADivisorsStorage;
+		HeuristicGroupADivisors = accelerator.CopySpanToDevice(stream, heuristicGroupA);
+
+		var heuristicGroupASquares = HeuristicPrimeSieves.GroupADivisorSquaresStorage;
+		HeuristicGroupADivisorSquares = accelerator.CopySpanToDevice(stream, heuristicGroupASquares);
+
+		HeuristicGroupBDivisorsEnding1 = accelerator.CopyUintSpanToDevice(stream, DivisorGenerator.SmallPrimesLastOneWithoutLastThree);
+		HeuristicGroupBDivisorSquaresEnding1 = accelerator.CopySpanToDevice(stream, DivisorGenerator.SmallPrimesPow2LastOneWithoutLastThree);
+
+		HeuristicGroupBDivisorsEnding7 = accelerator.CopyUintSpanToDevice(stream, DivisorGenerator.SmallPrimesLastSevenWithoutLastThree);
+		HeuristicGroupBDivisorSquaresEnding7 = accelerator.CopySpanToDevice(stream, DivisorGenerator.SmallPrimesPow2LastSevenWithoutLastThree);
+
+		HeuristicGroupBDivisorsEnding9 = accelerator.CopyUintSpanToDevice(stream, DivisorGenerator.SmallPrimesLastNineWithoutLastThree);
+		HeuristicGroupBDivisorSquaresEnding9 = accelerator.CopySpanToDevice(stream, DivisorGenerator.SmallPrimesPow2LastNineWithoutLastThree);
+	}
+
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisors;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisorSquares;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding1;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding1;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding7;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding7;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding9;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding9;
+
+	internal HeuristicGroupABGpuViews CreateViews() => new(
+			HeuristicGroupADivisors.View,
+			HeuristicGroupADivisorSquares.View,
+			HeuristicGroupBDivisorsEnding1.View,
+			HeuristicGroupBDivisorSquaresEnding1.View,
+			HeuristicGroupBDivisorsEnding7.View,
+			HeuristicGroupBDivisorSquaresEnding7.View,
+			HeuristicGroupBDivisorsEnding9.View,
+			HeuristicGroupBDivisorSquaresEnding9.View);
+
+	internal void Dispose()
+	{
+		HeuristicGroupADivisors.Dispose();
+		HeuristicGroupADivisorSquares.Dispose();
+		HeuristicGroupBDivisorsEnding1.Dispose();
+		HeuristicGroupBDivisorSquaresEnding1.Dispose();
+		HeuristicGroupBDivisorsEnding7.Dispose();
+		HeuristicGroupBDivisorSquaresEnding7.Dispose();
+		HeuristicGroupBDivisorsEnding9.Dispose();
+		HeuristicGroupBDivisorSquaresEnding9.Dispose();
+	}
+}

PerfectNumbers.Core/Gpu/Accelerators/HeuristicGroupABGpuViews.cs

@@ -0,0 +1,63 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+public readonly struct HeuristicGroupABGpuViews(
+		ArrayView1D<ulong, Stride1D.Dense> groupADivisors,
+		ArrayView1D<ulong, Stride1D.Dense> groupADivisorSquares,
+		ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding1,
+		ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding1,
+		ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding7,
+		ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding7,
+		ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding9,
+		ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding9)
+{
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisors = groupADivisors;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisorSquares = groupADivisorSquares;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding1 = groupBDivisorsEnding1;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding1 = groupBDivisorSquaresEnding1;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding7 = groupBDivisorsEnding7;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding7 = groupBDivisorSquaresEnding7;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding9 = groupBDivisorsEnding9;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding9 = groupBDivisorSquaresEnding9;
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ArrayView1D<ulong, Stride1D.Dense> SelectGroupB(byte ending) => ending switch
+	{
+		1 => GroupBDivisorsEnding1,
+		7 => GroupBDivisorsEnding7,
+		9 => GroupBDivisorsEnding9,
+		_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
+	};
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ArrayView1D<ulong, Stride1D.Dense> SelectGroupBSquares(byte ending) => ending switch
+	{
+		1 => GroupBDivisorSquaresEnding1,
+		7 => GroupBDivisorSquaresEnding7,
+		9 => GroupBDivisorSquaresEnding9,
+		_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
+	};
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ArrayView1D<ulong, Stride1D.Dense> SelectDivisors(HeuristicGpuDivisorTableKind kind) => kind switch
+	{
+		HeuristicGpuDivisorTableKind.GroupA => GroupADivisors,
+		HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorsEnding1,
+		HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorsEnding7,
+		HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorsEnding9,
+		_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
+	};
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public ArrayView1D<ulong, Stride1D.Dense> SelectDivisorSquares(HeuristicGpuDivisorTableKind kind) => kind switch
+	{
+		HeuristicGpuDivisorTableKind.GroupA => GroupADivisorSquares,
+		HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorSquaresEnding1,
+		HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorSquaresEnding7,
+		HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorSquaresEnding9,
+		_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
+	};
+}

PerfectNumbers.Core/Gpu/Accelerators/HeuristicGroupABPrimeTesterAccelerator.cs

@@ -0,0 +1,174 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+internal sealed class HeuristicGroupABPrimeTesterAccelerator
+{
+	#region Pool
+
+	[ThreadStatic]
+	private static Queue<HeuristicGroupABPrimeTesterAccelerator>? Pool;
+
+	internal static void WarmUp()
+	{
+		Accelerator[]? accelerators = AcceleratorPool.Shared.Accelerators;
+		int acceleratorCount = accelerators.Length;
+		for (var i = 0; i < acceleratorCount; i++)
+		{
+			Console.WriteLine($"Preparing heuristic accelerator {i}...");
+			var accelerator = accelerators[i];
+			AcceleratorStream stream = accelerator.CreateStream();
+			_ = HeuristicGroupABGpuTables.EnsureStaticTables(i, stream);
+			// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
+			// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
+			// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
+			stream.Synchronize();
+			stream.Dispose();
+			Console.WriteLine($"Heuristic accelerator {i} is ready");
+		}
+	}
+
+	internal static HeuristicGroupABPrimeTesterAccelerator Rent(int minBufferCapacity = 1)
+	{
+		var pool = Pool ??= new();
+		if (!pool.TryDequeue(out var lease))
+		{
+			lease = new HeuristicGroupABPrimeTesterAccelerator(minBufferCapacity);
+		}
+		else
+		{
+			lease.EnsureCapacity(minBufferCapacity);
+		}
+
+		return lease;
+	}
+
+	internal static void Return(HeuristicGroupABPrimeTesterAccelerator lease)
+	{
+		var pool = Pool ??= new();
+		pool.Enqueue(lease);
+	}
+
+	internal static void DisposeAll()
+	{
+		if (Pool != null)
+		{
+			while (Pool.TryDequeue(out var lease))
+			{
+				lease.DisposeResources();
+			}
+		}
+	}
+
+	internal static void Clear()
+	{
+		if (Pool != null)
+		{
+			var retained = new List<HeuristicGroupABPrimeTesterAccelerator>();
+			while (Pool.TryDequeue(out var lease))
+			{
+				lease.DisposeResources();
+			}
+		}
+	}
+
+	#endregion
+
+	public readonly HeuristicGroupABGpuViews DivisorTables;
+	internal readonly Context Context;
+	public readonly Accelerator Accelerator;
+
+	[ThreadStatic]
+	public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>? SmallPrimeSieveKernel;
+
+	internal static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetSmallPrimeSieveKernel(Accelerator accelerator)
+	{
+		var pool = SmallPrimeSieveKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	public readonly Kernel HeuristicTrialDivisionKernel;
+
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? Input = null;
+	public MemoryBuffer1D<byte, Stride1D.Dense>? Output = null;
+	public MemoryBuffer1D<int, Stride1D.Dense>? HeuristicFlag = null!;
+	public int BufferCapacity;
+
+	private HeuristicGroupABGpuTables _sharedTables;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisors => _sharedTables.HeuristicGroupADivisors;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisorSquares => _sharedTables.HeuristicGroupADivisorSquares;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding1 => _sharedTables.HeuristicGroupBDivisorsEnding1;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding1 => _sharedTables.HeuristicGroupBDivisorSquaresEnding1;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding7 => _sharedTables.HeuristicGroupBDivisorsEnding7;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding7 => _sharedTables.HeuristicGroupBDivisorSquaresEnding7;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding9 => _sharedTables.HeuristicGroupBDivisorsEnding9;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding9 => _sharedTables.HeuristicGroupBDivisorSquaresEnding9;
+
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+	internal readonly int AcceleratorIndex;
+
+	internal HeuristicGroupABPrimeTesterAccelerator(int minBufferCapacity)
+	{
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		AcceleratorIndex = acceleratorIndex;
+		Accelerator accelerator = _accelerators[acceleratorIndex];
+		Context = accelerator.Context;
+
+		Accelerator = accelerator;
+		EnsureCapacity(minBufferCapacity);
+
+		HeuristicTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGroupABGpuViews>(PrimeTesterKernels.HeuristicTrialGroupABDivisionKernel));
+
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		
+		// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
+
+		_sharedTables = HeuristicGroupABGpuTables.EnsureStaticTables(acceleratorIndex, stream);
+		DivisorTables = _sharedTables.CreateViews();
+
+		stream.Synchronize();
+		AcceleratorStreamPool.Return(acceleratorIndex, stream); 
+	}
+
+	public void EnsureCapacity(int minCapacity)
+	{
+		if (minCapacity <= BufferCapacity)
+		{
+			return;
+		}
+
+		Input?.Dispose();
+		Output?.Dispose();
+		HeuristicFlag?.Dispose();
+
+		Input = Accelerator.Allocate1D<ulong>(minCapacity);
+		Output = Accelerator.Allocate1D<byte>(minCapacity);
+		HeuristicFlag = Accelerator.Allocate1D<int>(minCapacity);
+
+		BufferCapacity = minCapacity;
+	}
+
+	public void Dispose()
+	{
+		Return(this);
+	}
+
+	internal void DisposeResources()
+	{
+		Input?.Dispose();
+		Output?.Dispose();
+		HeuristicFlag?.Dispose();
+		// ReleaseSharedTables(_sharedTables);
+		// These resources are shared between GPU leases
+		// Accelerator.Dispose();
+		// AcceleratorContext.Dispose();
+	}
+}

PerfectNumbers.Core/Gpu/Accelerators/LastDigitGpuTables.cs

@@ -0,0 +1,58 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+internal readonly struct LastDigitGpuTables
+{
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+	private static readonly LastDigitGpuTables[] _sharedTables = new LastDigitGpuTables[PerfectNumberConstants.RollingAccelerators];
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	internal static LastDigitGpuTables GetStaticTables(int acceleratorIndex) => _sharedTables[acceleratorIndex];
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	internal static void WarmUp(int acceleratorIndex, AcceleratorStream stream) => _sharedTables[acceleratorIndex] = new LastDigitGpuTables(_accelerators[acceleratorIndex], stream);
+
+	internal LastDigitGpuTables(Accelerator accelerator, AcceleratorStream stream)
+	{
+		var primesLastOne = DivisorGenerator.SmallPrimesLastOne;
+		var primesLastSeven = DivisorGenerator.SmallPrimesLastSeven;
+		var primesLastThree = DivisorGenerator.SmallPrimesLastThree;
+		var primesLastNine = DivisorGenerator.SmallPrimesLastNine;
+		var primesPow2LastOne = DivisorGenerator.SmallPrimesPow2LastOne;
+		var primesPow2LastSeven = DivisorGenerator.SmallPrimesPow2LastSeven;
+		var primesPow2LastThree = DivisorGenerator.SmallPrimesPow2LastThree;
+		var primesPow2LastNine = DivisorGenerator.SmallPrimesPow2LastNine;
+		DevicePrimesLastOne = accelerator.Allocate1D(stream, primesLastOne);
+		DevicePrimesLastSeven = accelerator.Allocate1D(stream, primesLastSeven);
+		DevicePrimesLastThree = accelerator.Allocate1D(stream, primesLastThree);
+		DevicePrimesLastNine = accelerator.Allocate1D(stream, primesLastNine);
+		DevicePrimesPow2LastOne = accelerator.Allocate1D(stream, primesPow2LastOne);
+		DevicePrimesPow2LastSeven = accelerator.Allocate1D(stream, primesPow2LastSeven);
+		DevicePrimesPow2LastThree = accelerator.Allocate1D(stream, primesPow2LastThree);
+		DevicePrimesPow2LastNine = accelerator.Allocate1D(stream, primesPow2LastNine);
+	}
+
+	internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
+	internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
+	internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree;
+	internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
+	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
+
+	internal void Dispose()
+	{
+		DevicePrimesLastOne.Dispose();
+		DevicePrimesLastSeven.Dispose();
+		DevicePrimesLastThree.Dispose();
+		DevicePrimesLastNine.Dispose();
+		DevicePrimesPow2LastOne.Dispose();
+		DevicePrimesPow2LastSeven.Dispose();
+		DevicePrimesPow2LastThree.Dispose();
+		DevicePrimesPow2LastNine.Dispose();
+	}
+}

PerfectNumbers.Core/Gpu/Accelerators/PrimeOrderCalculatorAccelerator.cs

@@ -0,0 +1,308 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+using PerfectNumbers.Core.Gpu.Kernels;
+
+namespace PerfectNumbers.Core.Gpu.Accelerators;
+
+public sealed class PrimeOrderCalculatorAccelerator
+{
+	#region Pool
+
+	[ThreadStatic]
+	private static Queue<PrimeOrderCalculatorAccelerator>? _pool;
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static PrimeOrderCalculatorAccelerator Rent(int primeTesterCapacity)
+	{
+		var pool = _pool ??= [];
+		if (!pool.TryDequeue(out var gpu))
+		{
+			int acceleratorIndex = AcceleratorPool.Shared.Rent();
+			gpu = new(acceleratorIndex, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity, PerfectNumberConstants.DefaultSmallPrimeFactorSlotCount, PerfectNumberConstants.DefaultSpecialMaxFactorCapacity);
+		}
+		else
+		{
+			gpu.EnsureCapacity(PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity);
+		}
+
+		return gpu;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static void Return(PrimeOrderCalculatorAccelerator gpu) => _pool!.Enqueue(gpu);
+
+	internal static void Clear()
+	{
+		Queue<PrimeOrderCalculatorAccelerator>? pool = _pool;
+		if (pool != null)
+		{
+			while (pool.TryDequeue(out var gpu))
+			{
+				gpu.Dispose();
+			}
+		}
+	}
+
+	internal static void DisposeAll()
+	{
+		Queue<PrimeOrderCalculatorAccelerator>? pool = _pool;
+		if (pool != null)
+		{
+			while (pool.TryDequeue(out var gpu))
+			{
+				gpu.Dispose();
+			}
+		}
+	}
+
+	internal static void WarmUp()
+	{
+		Accelerator[] accelerators = AcceleratorPool.Shared.Accelerators;
+		int acceleratorCount = accelerators.Length;
+		for (var i = 0; i < acceleratorCount; i++)
+		{
+			Console.WriteLine($"Preparing accelerator {i}...");
+			var accelerator = accelerators[i];
+			AcceleratorStreamPool.WarmUp(i);
+			// Don't take this from the pool as quick uploads of data to the accelerator consumes much of GPU's memory and throws.
+			AcceleratorStream stream = accelerator.CreateStream();
+			// TODO: Review which tables are needed when final execution path is defined.
+			LastDigitGpuTables.WarmUp(i, stream);
+			HeuristicCombinedGpuTables.WarmUp(i, stream);
+			// SmallPrimeFactorGpuTables.WarmUp(i, stream);
+			// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
+			// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
+			// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
+			// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
+			stream.Synchronize();
+			stream.Dispose();
+			accelerator.Synchronize();
+		}
+
+	}
+
+	#endregion
+	#region Static Tables
+
+	public readonly ArrayView1D<uint, Stride1D.Dense> DevicePrimesLastOne;
+	public readonly ArrayView1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
+	public readonly ArrayView1D<uint, Stride1D.Dense> DevicePrimesLastThree;
+	public readonly ArrayView1D<uint, Stride1D.Dense> DevicePrimesLastNine;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
+
+	public readonly HeuristicCombinedGpuViews DivisorTables;
+
+	public readonly ArrayView1D<uint, Stride1D.Dense> SmallPrimeFactorPrimes;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> SmallPrimeFactorSquares;
+
+	#endregion
+
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
+	public readonly Accelerator Accelerator;
+	public readonly int AcceleratorIndex;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> Input;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> OutputUlong;
+	public MemoryBuffer1D<byte, Stride1D.Dense> OutputByte;
+	public MemoryBuffer1D<int, Stride1D.Dense> OutputInt;
+	public MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> Pow2ModEntriesToTestOnDevice;
+	public readonly MemoryBuffer1D<int, Stride1D.Dense> SmallPrimeFactorCountSlot;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxResult;
+
+	public Dictionary<ulong, int> Pow2ModEntriesToTestOnHost = [];
+
+	public readonly Kernel CheckFactorsKernel;
+	public readonly Kernel ConvertToStandardKernel;
+
+	public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernelLauncher;
+
+	public readonly Kernel HeuristicCombinedTrialDivisionKernel;
+	public readonly Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews> HeuristicCombinedTrialDivisionKernelLauncher;
+	public readonly Kernel KeepMontgomeryKernel;
+	public readonly Kernel PollardRhoKernel;
+	public readonly Kernel Pow2ModWideKernel;
+	public readonly Kernel SharesFactorKernel;
+	public readonly Kernel SmallPrimeFactorKernel;
+	public readonly Kernel SmallPrimeSieveKernel;
+	public readonly Kernel SpecialMaxKernel;
+
+	public void EnsureCapacity(int factorsCount, int primeTesterCapacity)
+	{
+		Accelerator accelerator = Accelerator;
+
+		MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> pow2ModEntriesToTestOnDevice = Pow2ModEntriesToTestOnDevice;
+		if (pow2ModEntriesToTestOnDevice.Length < factorsCount)
+		{
+			lock (accelerator)
+			{
+				pow2ModEntriesToTestOnDevice.Dispose();
+				Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
+			}
+		}
+
+		MemoryBuffer1D<ulong, Stride1D.Dense> input = Input;
+		if (input.Length < primeTesterCapacity)
+		{
+			lock (accelerator)
+			{
+				input.Dispose();
+				Input = accelerator.Allocate1D<ulong>(primeTesterCapacity);
+			}
+		}
+
+		MemoryBuffer1D<byte, Stride1D.Dense> outputByte = OutputByte;
+		if (outputByte.Length < primeTesterCapacity)
+		{
+			lock (accelerator)
+			{
+				outputByte.Dispose();
+				OutputByte = accelerator.Allocate1D<byte>(primeTesterCapacity);
+			}
+		}
+
+		MemoryBuffer1D<int, Stride1D.Dense> outputInt = OutputInt;
+		if (outputInt.Length < primeTesterCapacity)
+		{
+			lock (accelerator)
+			{
+				outputInt.Dispose();
+				OutputInt = accelerator.Allocate1D<int>(primeTesterCapacity);
+			}
+		}
+	}
+
+	public void EnsureSmallPrimeFactorSlotsCapacity(int newSize)
+	{
+		// Console.WriteLine($"Resizing GPU scratch buffer from pool ({buffer.SmallPrimeFactorPrimeSlots.Length} / {smallPrimeFactorSlotCount}), ({buffer.SpecialMaxFactors.Length}/{specialMaxFactorCapacity})");
+		var accelerator = Accelerator;
+		MemoryBuffer1D<ulong, Stride1D.Dense> outputUlong = OutputUlong;
+		if (outputUlong.Length < newSize)
+		{
+			lock (accelerator)
+			{
+				outputUlong.Dispose();
+				OutputUlong = accelerator.Allocate1D<ulong>(newSize);
+			}
+		}
+
+		MemoryBuffer1D<int, Stride1D.Dense> outputInt = OutputInt;
+		if (outputInt.Length < newSize)
+		{
+			lock (accelerator)
+			{
+				outputInt.Dispose();
+				OutputInt = accelerator.Allocate1D<int>(newSize);
+			}
+		}
+	}
+
+	public void EnsureSpecialMaxFactorsCapacity(int newSize)
+	{
+		var accelerator = Accelerator;
+		MemoryBuffer1D<ulong, Stride1D.Dense> input = Input;
+		if (input.Length < newSize)
+		{
+			lock (accelerator)
+			{
+				input.Dispose();
+				Input = accelerator.Allocate1D<ulong>(newSize);
+			}
+		}
+
+		MemoryBuffer1D<ulong, Stride1D.Dense> outputUlong = OutputUlong;
+		if (outputUlong.Length < newSize)
+		{
+			lock (accelerator)
+			{
+				outputUlong.Dispose();
+				OutputUlong = accelerator.Allocate1D<ulong>(newSize);
+			}
+		}
+	}
+
+	public PrimeOrderCalculatorAccelerator(int acceleratorIndex, int factorsCount, int primeTesterCapacity, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
+	{
+		var accelerator = _accelerators[acceleratorIndex];
+		Accelerator = accelerator;
+		AcceleratorIndex = acceleratorIndex;
+		lock (accelerator)
+		{
+			Input = accelerator.Allocate1D<ulong>(Math.Max(specialMaxFactorCapacity, primeTesterCapacity));
+			OutputByte = Accelerator.Allocate1D<byte>(primeTesterCapacity);
+			OutputInt = Accelerator.Allocate1D<int>(Math.Max(primeTesterCapacity, smallPrimeFactorSlotCount));
+			OutputUlong = accelerator.Allocate1D<ulong>(Math.Max(specialMaxFactorCapacity, smallPrimeFactorSlotCount));
+			Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
+			SmallPrimeFactorCountSlot = accelerator.Allocate1D<int>(1);
+			SpecialMaxResult = accelerator.Allocate1D<ulong>(1);
+
+			LastDigitGpuTables lastDigitSharedTables = LastDigitGpuTables.GetStaticTables(acceleratorIndex);
+			DevicePrimesLastOne = lastDigitSharedTables.DevicePrimesLastOne.View;
+			DevicePrimesLastSeven = lastDigitSharedTables.DevicePrimesLastSeven.View;
+			DevicePrimesLastThree = lastDigitSharedTables.DevicePrimesLastThree.View;
+			DevicePrimesLastNine = lastDigitSharedTables.DevicePrimesLastNine.View;
+			DevicePrimesPow2LastOne = lastDigitSharedTables.DevicePrimesPow2LastOne.View;
+			DevicePrimesPow2LastSeven = lastDigitSharedTables.DevicePrimesPow2LastSeven.View;
+			DevicePrimesPow2LastThree = lastDigitSharedTables.DevicePrimesPow2LastThree.View;
+			DevicePrimesPow2LastNine = lastDigitSharedTables.DevicePrimesPow2LastNine.View;
+
+			DivisorTables = HeuristicCombinedGpuTables
+				.GetStaticTables(acceleratorIndex)
+				.CreateViews();
+
+			var smallPrimeFactorTables = SmallPrimeFactorGpuTables.GetStaticTables(acceleratorIndex);
+
+			if (smallPrimeFactorTables.HasValue)
+			{
+				SmallPrimeFactorGpuTables tables = smallPrimeFactorTables.Value;
+				SmallPrimeFactorPrimes = tables.Primes.View;
+				SmallPrimeFactorSquares = tables.Squares.View;
+			}
+
+			CheckFactorsKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel));
+
+			Kernel convertToStandardKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelConvertToStandard));
+
+			ConvertToStandardKernel = convertToStandardKernel;
+			ConvertToStandardKernelLauncher = convertToStandardKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
+
+			var heuristicCombinedTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>(PrimeTesterKernels.HeuristicTrialCombinedDivisionKernel));
+
+			HeuristicCombinedTrialDivisionKernel = heuristicCombinedTrialDivisionKernel;
+			HeuristicCombinedTrialDivisionKernelLauncher = heuristicCombinedTrialDivisionKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>>();
+
+			KeepMontgomeryKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery));
+
+			PollardRhoKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.TryPollardRhoKernel));
+
+			Pow2ModWideKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(PrimeOrderGpuHeuristics.Pow2ModKernelWide));
+
+			SharesFactorKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel));
+
+			SmallPrimeFactorKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan));
+
+			SmallPrimeSieveKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel));
+
+			SpecialMaxKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel));
+		}
+	}
+
+	public void Dispose()
+	{
+		Input.Dispose();
+		OutputByte.Dispose();
+		OutputInt.Dispose();
+		OutputUlong.Dispose();
+		Pow2ModEntriesToTestOnDevice.Dispose();
+		SmallPrimeFactorCountSlot.Dispose();
+		SpecialMaxResult.Dispose();
+
+		// These resources are shared between GPU leases
+		// Stream.Dispose();
+		// Accelerator.Dispose();
+		// Context.Dispose();
+	}
+}

PerfectNumbers.Core/Gpu/DeterministicRandomGpu.cs

@@ -0,0 +1,75 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+
+namespace PerfectNumbers.Core.Gpu;
+
+public static class ThreadStaticDeterministicRandomGpu
+{
+	[ThreadStatic]
+	private static DeterministicRandomGpu? _exclusive;
+	internal static DeterministicRandomGpu Exclusive => _exclusive ?? new DeterministicRandomGpu();
+}
+
+internal struct DeterministicRandomGpu
+{
+	private static long s_seedCounter;
+
+	private ulong _state;
+	public readonly ulong State
+	{
+		[MethodImpl(MethodImplOptions.AggressiveInlining)]
+		get
+		{
+			return _state;
+		}
+	}
+
+	public DeterministicRandomGpu()
+	{
+		_state = CreateThreadState();
+	}
+
+	public DeterministicRandomGpu(ulong initialState)
+	{
+		_state = initialState;
+	}
+
+	public ulong NextUInt64()
+	{
+		ulong state = _state;
+		state ^= state >> 12;
+		state ^= state << 25;
+		state ^= state >> 27;
+		_state = state;
+		return state * 2685821657736338717UL;
+	}
+
+	public GpuUInt128 NextUInt128()
+	{
+		GpuUInt128 high = (GpuUInt128)NextUInt64();
+		ReadOnlyGpuUInt128 low = (ReadOnlyGpuUInt128)NextUInt64();
+		high.ShiftLeft(64);
+		high.Add(low);
+		return high;
+	}
+
+	public void SetState(ulong state)
+	{
+		_state = state;
+	}
+
+	private static ulong CreateThreadState()
+	{
+		ulong seed = (ulong)Atomic.Add(ref s_seedCounter, 1);
+		seed += 0x9E3779B97F4A7C15UL;
+		seed = (seed ^ (seed >> 30)) * 0xBF58476D1CE4E5B9UL;
+		seed = (seed ^ (seed >> 27)) * 0x94D049BB133111EBUL;
+		seed ^= seed >> 31;
+		if (seed == 0UL)
+		{
+			seed = 0x9E3779B97F4A7C15UL;
+		}
+
+		return seed;
+	}
+}

PerfectNumbers.Core/Gpu/ExponentRemainderStepperGpu.cs

@@ -1,5 +1,4 @@
 using System.Runtime.CompilerServices;
-using PerfectNumbers.Core;
 
 namespace PerfectNumbers.Core.Gpu;
 
@@ -24,17 +23,14 @@ internal struct ExponentRemainderStepperGpu
         _currentResidue = GpuUInt128.One;
     }
 
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public bool MatchesDivisor(in GpuDivisorPartialData divisor)
-    {
-        return _modulus == divisor.Modulus;
-    }
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public readonly bool MatchesDivisor(in GpuDivisorPartialData divisor) => _modulus == divisor.Modulus;
 
-    /// <summary>
-    /// Clears the cached residue so the stepper can be reused for another divisor.
-    /// Callers must reinitialize the stepper before consuming any residues after a reset.
-    /// </summary>
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+	/// <summary>
+	/// Clears the cached residue so the stepper can be reused for another divisor.
+	/// Callers must reinitialize the stepper before consuming any residues after a reset.
+	/// </summary>
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
     public void Reset()
     {
         PreviousExponent = 0UL;

PerfectNumbers.Core/Gpu/GpuContextPool.cs

@@ -1,129 +0,0 @@
-using System;
-using System.Collections.Concurrent;
-using ILGPU;
-using ILGPU.Runtime;
-
-namespace PerfectNumbers.Core.Gpu;
-
-public static class GpuContextPool
-{
-	// The pool intentionally skips pre-loading ProcessEightBitWindows kernels so accelerator initialization
-	// stays stable; GpuKernelPool JIT-compiles them on first use alongside the small-cycle uploads.
-	// Default device preference for generic GPU kernels (prime scans, NTT, etc.)
-	internal sealed class PooledContext
-	{
-		public readonly Context Context;
-		public readonly Accelerator Accelerator;
-
-		private static readonly Context _sharedContext = Context.CreateDefault();
-		private static readonly Device _sharedDevice = _sharedContext.GetPreferredDevice(false);
-		private static readonly Accelerator _sharedAccelerator = _sharedDevice.CreateAccelerator(_sharedContext);
-
-		public PooledContext()
-		{
-			Context = _sharedContext;
-			Accelerator = _sharedAccelerator;
-			// NOTE: Avoid loading/compiling any kernel here to prevent implicit
-			// CL stream/queue creation during accelerator construction.
-			// Some OpenCL drivers are fragile when a queue is created immediately
-			// after device init. Real kernels will be JIT-loaded at first use.
-		}
-
-		public void Dispose()
-		{
-			// Ensure all cached GPU buffers for this accelerator are released.
-			NttGpuMath.ClearCaches(Accelerator);
-			// Release PrimeTester GPU state for this accelerator (kernel/device primes)
-			// Clear any per-accelerator cached resources to avoid releasing
-			// after the accelerator is destroyed.
-			PrimeTester.ClearGpuCaches(Accelerator);
-
-			// These resources are shared between GPU leases
-			//  Accelerator.Dispose();
-			// 	Context.Dispose();
-		}
-	}
-
-	private static readonly ConcurrentQueue<PooledContext> GpuPool = new();
-	private static int WarmedGpuContextCount;
-
-	// Allows callers to choose CPU/GPU per use-case, decoupled from ForceCpu.
-        	public static GpuContextLease Rent()
-	{
-		if (GpuPool.TryDequeue(out var gpu))
-		{
-			return new GpuContextLease(gpu);
-		}
-
-		// Create a new accelerator when the pool does not have one available.
-		return new GpuContextLease(new PooledContext());
-	}
-
-
-	public static void WarmUpPool(int threadCount)
-	{
-		if (threadCount <= WarmedGpuContextCount)
-		{
-			return;
-		}
-
-		int toCreate = threadCount - WarmedGpuContextCount;
-		Console.WriteLine($"Warming up GPU context pool to {toCreate} contexts...");
-		var contexts = new PooledContext[toCreate];
-		var pool = GpuPool;
-		PooledContext context;
-		for (int i = 0; i < toCreate; i++)
-		{
-			context = new PooledContext();
-			contexts[i] = context;
-			pool.Enqueue(context);
-		}
-
-		WarmedGpuContextCount = threadCount;
-	}
-
-	public static void DisposeAll()
-	{
-		while (GpuPool.TryDequeue(out var gpu))
-		{
-			gpu.Dispose();
-		}
-
-		PrimeTester.DisposeGpuContexts();
-		WarmedGpuContextCount = 0;
-	}
-
-	private static void Return(PooledContext ctx)
-	{
-		ctx.Accelerator.Synchronize();
-		GpuPool.Enqueue(ctx);
-	}
-
-	public struct GpuContextLease
-	{
-		private PooledContext? _ctx;
-		private bool _disposed;
-
-		internal GpuContextLease(PooledContext ctx)
-		{
-			_ctx = ctx;
-			_disposed = false;
-		}
-
-		public Context Context => _ctx!.Context;
-		public Accelerator Accelerator => _ctx!.Accelerator;
-
-		public void Dispose()
-		{
-			if (_disposed)
-			{
-				return;
-			}
-
-			_disposed = true;
-			Return(_ctx!);
-			_ctx = null;
-		}
-	}
-}
-

PerfectNumbers.Core/Gpu/GpuKernelLease.cs

@@ -1,153 +0,0 @@
-using System;
-using System.Collections.Concurrent;
-using ILGPU;
-using ILGPU.Runtime;
-using static PerfectNumbers.Core.Gpu.GpuContextPool;
-
-namespace PerfectNumbers.Core.Gpu;
-
-public sealed class GpuKernelLease
-{
-    private static readonly ConcurrentQueue<GpuKernelLease> Pool = new();
-
-    private Action? _releaseLimiter;
-    private GpuContextLease? _gpu;
-    private KernelContainer? _kernels;
-    private AcceleratorStream? _stream;
-
-    private GpuKernelLease()
-    {
-    }
-
-    internal static GpuKernelLease Rent(GpuPrimeWorkLimiter.Lease limiter, GpuContextLease gpu, KernelContainer kernels)
-    {
-        if (!Pool.TryDequeue(out var lease))
-        {
-            lease = new GpuKernelLease();
-        }
-
-        lease._releaseLimiter = limiter.Dispose;
-        lease._gpu = gpu;
-        lease._kernels = kernels;
-        lease._stream = null;
-        return lease;
-    }
-
-    public Accelerator Accelerator => _gpu?.Accelerator ?? throw new NullReferenceException("GPU context is already released");
-
-    public AcceleratorStream Stream
-    {
-        get
-        {
-            if (_stream is { } stream)
-            {
-                return stream;
-            }
-
-            // It feels that we can combine these into 1 return statement, but that increases the max stack frame size to 3 from 2
-            stream = Accelerator.CreateStream();
-            _stream = stream;
-            return stream;
-        }
-    }
-
-    public Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>> OrderKernel
-    {
-        get
-        {
-            var accel = Accelerator; // avoid capturing 'this' in lambda
-            return KernelContainer.InitOnce(ref _kernels!.Order, () =>
-            {
-                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
-                var kernel = KernelUtil.GetKernel(loaded);
-                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>>();
-            });
-        }
-    }
-
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> Pow2ModKernel
-    {
-        get
-        {
-            var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.Pow2Mod, () =>
-            {
-                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModKernelScan);
-                var kernel = KernelUtil.GetKernel(loaded);
-                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
-            });
-        }
-    }
-
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>> IncrementalKernel
-    {
-        get
-        {
-            var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.Incremental, () =>
-            {
-                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
-                var kernel = KernelUtil.GetKernel(loaded);
-                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>>();
-            });
-        }
-    }
-
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>> IncrementalOrderKernel
-    {
-        get
-        {
-            var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.IncrementalOrder, () =>
-            {
-                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
-                var kernel = KernelUtil.GetKernel(loaded);
-                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
-            });
-        }
-    }
-
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>> Pow2ModOrderKernel
-    {
-        get
-        {
-            var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.Pow2ModOrder, () =>
-            {
-                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
-                var kernel = KernelUtil.GetKernel(loaded);
-                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
-            });
-        }
-    }
-
-    public Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SmallPrimeFactorKernel
-    {
-        get
-        {
-            var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.SmallPrimeFactor, () =>
-            {
-                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan);
-                var kernel = KernelUtil.GetKernel(loaded);
-                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
-            });
-        }
-    }
-
-    public void Dispose()
-    {
-        _stream?.Dispose();
-        _stream = null;
-
-        _gpu?.Dispose();
-        _gpu = null;
-
-        _releaseLimiter?.Invoke();
-        _releaseLimiter = null;
-
-        _kernels = null;
-        Pool.Enqueue(this);
-    }
-
-}

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -1,303 +1,232 @@
-using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
-using static PerfectNumbers.Core.Gpu.GpuContextPool;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core.Gpu;
 
-public readonly struct ResidueAutomatonArgs
-{
-    public readonly ulong Q0M10;
-    public readonly ulong Step10;
-    public readonly ulong Q0M8;
-    public readonly ulong Step8;
-    public readonly ulong Q0M3;
-    public readonly ulong Step3;
-    public readonly ulong Q0M5;
-    public readonly ulong Step5;
-
-    public ResidueAutomatonArgs(ulong q0m10, ulong step10, ulong q0m8, ulong step8, ulong q0m3, ulong step3, ulong q0m5, ulong step5)
-    {
-        Q0M10 = q0m10; Step10 = step10; Q0M8 = q0m8; Step8 = step8; Q0M3 = q0m3; Step3 = step3; Q0M5 = q0m5; Step5 = step5;
-    }
-}
-
-public sealed class KernelContainer
-{
-    // Serializes first-time initialization of kernels/buffers per accelerator.
-    public Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>? Order;
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
-    ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>? Incremental;
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
-        ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>,
-        ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>,
-        ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>? Pow2Mod;
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
-            ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>? IncrementalOrder;
-    public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
-        ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>? Pow2ModOrder;
-    public Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>? SmallPrimeFactor;
-
-    // Optional device buffer with small divisor cycles (<= 4M). Index = divisor, value = cycle length.
-    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallCycles;
-    public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimeFactorsPrimes;
-    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorsSquares;
-    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorPrimeSlots;
-    public MemoryBuffer1D<int, Stride1D.Dense>? SmallPrimeFactorExponentSlots;
-    public MemoryBuffer1D<int, Stride1D.Dense>? SmallPrimeFactorCountSlot;
-    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorRemainingSlot;
-    public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastOne;
-    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimesPow2LastOne;
-    public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastSeven;
-    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimesPow2LastSeven;
-
-    public static T InitOnce<T>(ref T? slot, Func<T> factory) where T : class
-    {
-        var current = Volatile.Read(ref slot);
-        if (current is not null)
-        {
-            return current;
-        }
-
-        current = Volatile.Read(ref slot);
-        if (current is null)
-        {
-            current = factory();
-            Volatile.Write(ref slot, current);
-        }
-        return current;
-    }
-}
-
-public readonly struct ResiduePrimeViews(
-    ArrayView1D<uint, Stride1D.Dense> lastOne,
-    ArrayView1D<uint, Stride1D.Dense> lastSeven,
-    ArrayView1D<ulong, Stride1D.Dense> lastOnePow2,
-    ArrayView1D<ulong, Stride1D.Dense> lastSevenPow2)
-{
-    public readonly ArrayView1D<uint, Stride1D.Dense> LastOne = lastOne;
-    public readonly ArrayView1D<uint, Stride1D.Dense> LastSeven = lastSeven;
-    public readonly ArrayView1D<ulong, Stride1D.Dense> LastOnePow2 = lastOnePow2;
-    public readonly ArrayView1D<ulong, Stride1D.Dense> LastSevenPow2 = lastSevenPow2;
-}
-
-public readonly struct SmallPrimeFactorTables
-{
-    public readonly MemoryBuffer1D<uint, Stride1D.Dense> Primes;
-    public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Squares;
-    public readonly int Count;
-
-    public SmallPrimeFactorTables(
-        MemoryBuffer1D<uint, Stride1D.Dense> primes,
-        MemoryBuffer1D<ulong, Stride1D.Dense> squares,
-        int count)
-    {
-        Primes = primes;
-        Squares = squares;
-        Count = count;
-    }
-
-    public ArrayView1D<uint, Stride1D.Dense> PrimesView => Primes.View;
-    public ArrayView1D<ulong, Stride1D.Dense> SquaresView => Squares.View;
-}
-
-public readonly struct SmallPrimeFactorScratch
-{
-    public readonly MemoryBuffer1D<ulong, Stride1D.Dense> PrimeSlots;
-    public readonly MemoryBuffer1D<int, Stride1D.Dense> ExponentSlots;
-    public readonly MemoryBuffer1D<int, Stride1D.Dense> CountSlot;
-    public readonly MemoryBuffer1D<ulong, Stride1D.Dense> RemainingSlot;
-
-    public SmallPrimeFactorScratch(
-        MemoryBuffer1D<ulong, Stride1D.Dense> primeSlots,
-        MemoryBuffer1D<int, Stride1D.Dense> exponentSlots,
-        MemoryBuffer1D<int, Stride1D.Dense> countSlot,
-        MemoryBuffer1D<ulong, Stride1D.Dense> remainingSlot)
-    {
-        PrimeSlots = primeSlots;
-        ExponentSlots = exponentSlots;
-        CountSlot = countSlot;
-        RemainingSlot = remainingSlot;
-    }
-
-    public void Clear()
-    {
-        PrimeSlots.MemSetToZero();
-        ExponentSlots.MemSetToZero();
-        CountSlot.MemSetToZero();
-        RemainingSlot.MemSetToZero();
-    }
-}
-
 public class GpuKernelPool
 {
-    private static readonly ConcurrentDictionary<Accelerator, KernelContainer> KernelCache = new(); // TODO: Replace this concurrent map with a simple accelerator-indexed lookup once kernel launchers are prewarmed during startup so we can drop the thread-safe wrapper entirely.
+	[ThreadStatic]
+	private static KernelContainer[]? _pool;
 
-    private static KernelContainer GetKernels(Accelerator accelerator)
-    {
-        return KernelCache.GetOrAdd(accelerator, _ => new KernelContainer());
-    }
+	public static T InitOnce<T>(ref T? target, Func<T> valueFactory) where T : class
+	{
+		if (target is { } value)
+		{
+			return value;
+		}
 
-    // Ensures the small cycles table is uploaded to the device for the given accelerator.
-    // Returns the ArrayView to pass into kernels (when kernels are extended to accept it).
-    public static ArrayView1D<ulong, Stride1D.Dense> EnsureSmallCyclesOnDevice(Accelerator accelerator)
-    {
-        var kernels = GetKernels(accelerator);
-        if (kernels.SmallCycles is { } buffer)
-        {
-            return buffer.View;
-        }
+		var newValue = valueFactory();
+		return Interlocked.CompareExchange(ref target, newValue, null) ?? newValue;
+	}
 
-        // Ensure single upload per accelerator even if multiple threads race here.
-        lock (kernels) // TODO: Remove this lock by pre-uploading the immutable small-cycle snapshot during initialization; once no mutation happens at runtime, the pool must expose a simple reference without synchronization.
-        {
-            if (kernels.SmallCycles is { } existing)
-            {
-                return existing.View;
-            }
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
 
-            var host = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot(); // TODO: Preload this device buffer during startup and keep it immutable so we can delete the lock above in favor of the preloaded snapshot.
-            var device = accelerator.Allocate1D<ulong>(host.Length);
-            device.View.CopyFromCPU(host);
-            kernels.SmallCycles = device;
-            return device.View;
-        }
-    }
 
-    public static ResiduePrimeViews EnsureSmallPrimesOnDevice(Accelerator accelerator)
-    {
-        var kernels = GetKernels(accelerator);
-        if (kernels.SmallPrimesLastOne is { } lastOne &&
-            kernels.SmallPrimesLastSeven is { } lastSeven &&
-            kernels.SmallPrimesPow2LastOne is { } lastOnePow2 &&
-            kernels.SmallPrimesPow2LastSeven is { } lastSevenPow2)
-        {
-            return new ResiduePrimeViews(lastOne.View, lastSeven.View, lastOnePow2.View, lastSevenPow2.View);
-        }
+	internal static KernelContainer GetOrAddKernels(int acceleratorIndex, AcceleratorStream stream, KernelType kernelType)
+	{
+		var pool = _pool ??= [];
+		if (pool[acceleratorIndex] is not { } kernels)
+		{
+			kernels = new();
+		}
 
-        lock (kernels) // TODO: Inline these small-prime uploads into startup initialization alongside the small-cycle snapshot so we can drop runtime locking and keep the GPU pool free of synchronization.
-        {
-            if (kernels.SmallPrimesLastOne is { } existingLastOne &&
-                kernels.SmallPrimesLastSeven is { } existingLastSeven &&
-                kernels.SmallPrimesPow2LastOne is { } existingLastOnePow2 &&
-                kernels.SmallPrimesPow2LastSeven is { } existingLastSevenPow2)
-            {
-                return new ResiduePrimeViews(existingLastOne.View, existingLastSeven.View, existingLastOnePow2.View, existingLastSevenPow2.View);
-            }
+		Accelerator accelerator = _accelerators[acceleratorIndex];
+		if (kernelType.HasFlag(KernelType.OrderKernelScan))
+		{
+			InitOrderKernelScan(accelerator, kernels);
+		}
 
-            var hostLastOne = PrimesGenerator.SmallPrimesLastOne;
-            var hostLastSeven = PrimesGenerator.SmallPrimesLastSeven;
-            var hostLastOnePow2 = PrimesGenerator.SmallPrimesPow2LastOne;
-            var hostLastSevenPow2 = PrimesGenerator.SmallPrimesPow2LastSeven;
+		if (kernelType.HasFlag(KernelType.Pow2ModKernelScan))
+		{
+			InitPow2ModKernelScan(accelerator, kernels);
+		}
 
-            var deviceLastOne = accelerator.Allocate1D<uint>(hostLastOne.Length);
-            deviceLastOne.View.CopyFromCPU(hostLastOne);
-            var deviceLastSeven = accelerator.Allocate1D<uint>(hostLastSeven.Length);
-            deviceLastSeven.View.CopyFromCPU(hostLastSeven);
-            var deviceLastOnePow2 = accelerator.Allocate1D<ulong>(hostLastOnePow2.Length);
-            deviceLastOnePow2.View.CopyFromCPU(hostLastOnePow2);
-            var deviceLastSevenPow2 = accelerator.Allocate1D<ulong>(hostLastSevenPow2.Length);
-            deviceLastSevenPow2.View.CopyFromCPU(hostLastSevenPow2);
+		if (kernelType.HasFlag(KernelType.IncrementalKernelScan))
+		{
+			InitIncrementalKernelScan(accelerator, kernels);
+		}
 
-            kernels.SmallPrimesLastOne = deviceLastOne;
-            kernels.SmallPrimesLastSeven = deviceLastSeven;
-            kernels.SmallPrimesPow2LastOne = deviceLastOnePow2;
-            kernels.SmallPrimesPow2LastSeven = deviceLastSevenPow2;
+		if (kernelType.HasFlag(KernelType.IncrementalOrderKernelScan))
+		{
+			IncrementalOrderKernelScan(accelerator, kernels);
+		}
 
-            return new ResiduePrimeViews(deviceLastOne.View, deviceLastSeven.View, deviceLastOnePow2.View, deviceLastSevenPow2.View);
-        }
-    }
+		if (kernelType.HasFlag(KernelType.Pow2ModOrderKernelScan))
+		{
+			InitPow2ModOrderKernelScan(accelerator, kernels);
+		}
 
-    public static SmallPrimeFactorTables EnsureSmallPrimeFactorTables(Accelerator accelerator)
-    {
-        var kernels = GetKernels(accelerator);
-        if (kernels.SmallPrimeFactorsPrimes is { } primeBuffer && kernels.SmallPrimeFactorsSquares is { } squareBuffer)
-        {
-            return new SmallPrimeFactorTables(primeBuffer, squareBuffer, (int)primeBuffer.Length);
-        }
+		PreloadStaticTables(accelerator, kernels, stream, kernelType);
 
-        lock (kernels)
-        {
-            if (kernels.SmallPrimeFactorsPrimes is { } existingPrimes && kernels.SmallPrimeFactorsSquares is { } existingSquares)
-            {
-                return new SmallPrimeFactorTables(existingPrimes, existingSquares, (int)existingPrimes.Length);
-            }
+		pool[acceleratorIndex] = kernels;
+		return kernels;
 
-            var hostPrimes = PrimesGenerator.SmallPrimes;
-            var hostSquares = PrimesGenerator.SmallPrimesPow2;
+		static void InitOrderKernelScan(Accelerator accelerator, KernelContainer kernels)
+		{
+			InitOnce(ref kernels.Order, () =>
+			{
+				var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
 
-            var devicePrimes = accelerator.Allocate1D<uint>(hostPrimes.Length);
-            devicePrimes.View.CopyFromCPU(hostPrimes);
-            var deviceSquares = accelerator.Allocate1D<ulong>(hostSquares.Length);
-            deviceSquares.View.CopyFromCPU(hostSquares);
+				var kernel = KernelUtil.GetKernel(loaded);
 
-            kernels.SmallPrimeFactorsPrimes = devicePrimes;
-            kernels.SmallPrimeFactorsSquares = deviceSquares;
+				return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>>();
+			});
+		}
 
-            return new SmallPrimeFactorTables(devicePrimes, deviceSquares, hostPrimes.Length);
-        }
-    }
+		static void InitPow2ModKernelScan(Accelerator accelerator, KernelContainer kernels)
+		{
+			InitOnce(ref kernels.Pow2Mod, () =>
+			{
+				var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModKernelScan);
 
-    public static SmallPrimeFactorScratch EnsureSmallPrimeFactorScratch(Accelerator accelerator, int slotCount)
-    {
-        var kernels = GetKernels(accelerator);
-        lock (kernels)
-        {
-            MemoryBuffer1D<ulong, Stride1D.Dense>? primeSlots = kernels.SmallPrimeFactorPrimeSlots;
-            if (primeSlots is null || primeSlots.Length < slotCount)
-            {
-                primeSlots?.Dispose();
-                primeSlots = accelerator.Allocate1D<ulong>(slotCount);
-                kernels.SmallPrimeFactorPrimeSlots = primeSlots;
-            }
+				var kernel = KernelUtil.GetKernel(loaded);
 
-            MemoryBuffer1D<int, Stride1D.Dense>? exponentSlots = kernels.SmallPrimeFactorExponentSlots;
-            if (exponentSlots is null || exponentSlots.Length < slotCount)
-            {
-                exponentSlots?.Dispose();
-                exponentSlots = accelerator.Allocate1D<int>(slotCount);
-                kernels.SmallPrimeFactorExponentSlots = exponentSlots;
-            }
+				return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
+			});
+		}
 
-            MemoryBuffer1D<int, Stride1D.Dense>? countSlot = kernels.SmallPrimeFactorCountSlot;
-            if (countSlot is null)
-            {
-                countSlot = accelerator.Allocate1D<int>(1);
-                kernels.SmallPrimeFactorCountSlot = countSlot;
-            }
+		static void InitIncrementalKernelScan(Accelerator accelerator, KernelContainer kernels)
+		{
+			InitOnce(ref kernels.Incremental, () =>
+			{
+				var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
 
-            MemoryBuffer1D<ulong, Stride1D.Dense>? remainingSlot = kernels.SmallPrimeFactorRemainingSlot;
-            if (remainingSlot is null)
-            {
-                remainingSlot = accelerator.Allocate1D<ulong>(1);
-                kernels.SmallPrimeFactorRemainingSlot = remainingSlot;
-            }
+				var kernel = KernelUtil.GetKernel(loaded);
 
-            return new SmallPrimeFactorScratch(primeSlots!, exponentSlots!, countSlot!, remainingSlot!);
-        }
-    }
+				return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>>();
+			});
+		}
 
-    public static GpuKernelLease GetKernel(bool useGpuOrder)
-    {
-        var limiter = GpuPrimeWorkLimiter.Acquire();
-        var gpu = Rent();
-        var accelerator = gpu.Accelerator;
-        var kernels = GetKernels(accelerator);
-        return GpuKernelLease.Rent(limiter, gpu, kernels);
-    }
+		static void IncrementalOrderKernelScan(Accelerator accelerator, KernelContainer kernels)
+		{
+			InitOnce(ref kernels.IncrementalOrder, () =>
+			{
+				var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
 
-    /// <summary>
-    /// Runs a GPU action with an acquired accelerator and stream.
-    /// </summary>
-    /// <param name="action">Action to run with (Accelerator, Stream).</param>
-    public static void Run(Action<Accelerator, AcceleratorStream> action)
-    {
-        var lease = GetKernel(useGpuOrder: true);
-        var accelerator = lease.Accelerator;
-        var stream = accelerator.CreateStream();
-        action(accelerator, stream);
-        stream.Dispose();
-        lease.Dispose();
+				var kernel = KernelUtil.GetKernel(loaded);
 
-    }
+				return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte,
+				ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
+			});
+		}
+
+		static void InitPow2ModOrderKernelScan(Accelerator accelerator, KernelContainer kernels)
+		{
+			InitOnce(ref kernels.Pow2ModOrder, () =>
+			{
+				var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
+
+				var kernel = KernelUtil.GetKernel(loaded);
+
+				return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
+			});
+		}
+	}
+
+	internal static void PreloadStaticTables(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream, KernelType kernelType)
+	{
+		if (kernelType.HasFlag(KernelType.SmallCycles))
+		{
+			EnsureSmallCyclesOnDevice(accelerator, kernels, stream);
+		}
+
+		if (kernelType.HasFlag(KernelType.SmallPrimes))
+		{
+			EnsureSmallPrimesOnDevice(accelerator, kernels, stream);
+		}
+	}
+
+	public static ArrayView1D<ulong, Stride1D.Dense> GetSmallCyclesOnDevice(KernelContainer kernels) => kernels.SmallCycles!;
+
+	// Ensures the small cycles table is uploaded to the device for the given accelerator.
+	// Returns the ArrayView to pass into kernels (when kernels are extended to accept it).
+	private static void EnsureSmallCyclesOnDevice(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream)
+	{
+		if (kernels.SmallCycles is { })
+		{
+			return;
+		}
+
+		var host = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
+		MemoryBuffer1D<ulong, Stride1D.Dense>? device;
+
+		lock (accelerator)
+		{
+			device = accelerator.Allocate1D<ulong>(host.Length);
+		}
+
+		device.View.CopyFromCPU(stream, host);
+		kernels.SmallCycles = device;
+	}
+
+	public static ResiduePrimeViews GetSmallPrimesOnDevice(KernelContainer kernels) => new(
+		kernels.SmallPrimesLastOne!,
+		kernels.SmallPrimesLastSeven!,
+		kernels.SmallPrimesPow2LastOne!,
+		kernels.SmallPrimesPow2LastSeven!
+	);
+
+	private static void EnsureSmallPrimesOnDevice(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream)
+	{
+		if (kernels.SmallPrimesLastOne is { } &&
+			kernels.SmallPrimesLastSeven is { } &&
+			kernels.SmallPrimesPow2LastOne is { } &&
+			kernels.SmallPrimesPow2LastSeven is { })
+		{
+			return;
+		}
+
+		if (kernels.SmallPrimesLastOne is { } &&
+			kernels.SmallPrimesLastSeven is { } &&
+			kernels.SmallPrimesPow2LastOne is { } &&
+			kernels.SmallPrimesPow2LastSeven is { })
+		{
+			return;
+		}
+
+		var hostLastOne = PrimesGenerator.SmallPrimesLastOne;
+		var hostLastSeven = PrimesGenerator.SmallPrimesLastSeven;
+		var hostLastOnePow2 = PrimesGenerator.SmallPrimesPow2LastOne;
+		var hostLastSevenPow2 = PrimesGenerator.SmallPrimesPow2LastSeven;
+
+		MemoryBuffer1D<uint, Stride1D.Dense>? deviceLastOne;
+		MemoryBuffer1D<uint, Stride1D.Dense>? deviceLastSeven;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? deviceLastOnePow2;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? deviceLastSevenPow2;
+
+		lock (accelerator)
+		{
+			deviceLastOne = accelerator.Allocate1D<uint>(hostLastOne.Length);
+			deviceLastSeven = accelerator.Allocate1D<uint>(hostLastSeven.Length);
+			deviceLastOnePow2 = accelerator.Allocate1D<ulong>(hostLastOnePow2.Length);
+			deviceLastSevenPow2 = accelerator.Allocate1D<ulong>(hostLastSevenPow2.Length);
+		}
+
+		deviceLastOne.View.CopyFromCPU(stream, hostLastOne);
+		deviceLastSeven.View.CopyFromCPU(stream, hostLastSeven);
+		deviceLastOnePow2.View.CopyFromCPU(stream, hostLastOnePow2);
+		deviceLastSevenPow2.View.CopyFromCPU(stream, hostLastSevenPow2);
+		kernels.SmallPrimesLastOne = deviceLastOne;
+		kernels.SmallPrimesLastSeven = deviceLastSeven;
+		kernels.SmallPrimesPow2LastOne = deviceLastOnePow2;
+		kernels.SmallPrimesPow2LastSeven = deviceLastSevenPow2;
+	}
+
+	/// <summary>
+	/// Runs a GPU action with an acquired accelerator and stream.
+	/// </summary>
+	/// <param name="action">Action to run with (Accelerator, Stream).</param>
+	public static void Run(Action<Accelerator, AcceleratorStream> action)
+	{
+		// GpuPrimeWorkLimiter.Acquire();
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+		var stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		action(accelerator, stream);
+		stream.Synchronize();
+		stream.Dispose();
+		accelerator.Dispose();
+		// GpuPrimeWorkLimiter.Release();
+	}
 }

PerfectNumbers.Core/Gpu/GpuPrimeWorkLimiter.cs

@@ -1,23 +1,14 @@
-using System;
-using System.Threading;
-
 namespace PerfectNumbers.Core.Gpu;
 
 public static class GpuPrimeWorkLimiter
 {
-    private static SemaphoreSlim _semaphore = new(1, int.MaxValue);
-    private static int _currentLimit = 1;
+	private static SemaphoreSlim _semaphore = new(1, int.MaxValue);
+	private static int _currentLimit = 1;
 
-    public static Lease Acquire()
-    {
-        // TODO: Inline the pooled limiter guard from the GpuLimiterThroughputBenchmarks so prime-sieve
-        // GPU jobs stop allocating Lease instances on every acquisition.
-        var sem = _semaphore;
-        sem.Wait();
-        return new Lease(sem);
-    }
+	public static void Acquire() { } // => _semaphore.Wait();
+	public static void Release() { }// => _semaphore.Release();
 
-    public static void SetLimit(int value)
+	public static void SetLimit(int value)
     {
         if (value <= 0)
         {
@@ -36,19 +27,4 @@ public static class GpuPrimeWorkLimiter
         _currentLimit = value;
         previous.Dispose();
     }
-
-    public sealed class Lease : IDisposable
-    {
-        private readonly SemaphoreSlim _sem;
-
-        internal Lease(SemaphoreSlim sem)
-        {
-            _sem = sem;
-        }
-
-        public void Dispose()
-        {
-            _sem.Release();
-        }
-    }
 }

PerfectNumbers.Core/Gpu/GpuUInt128.cs

@@ -261,13 +261,13 @@ public struct GpuUInt128 : IComparable<GpuUInt128>, IEquatable<GpuUInt128>
     public override readonly int GetHashCode() => HashCode.Combine(High, Low);
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public void Add(in ReadOnlyGpuUInt128 other) => AddInternal(other.High, other.Low);
+    public void Add(in ReadOnlyGpuUInt128 other) => Add(other.High, other.Low);
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public void Add(GpuUInt128 other) => AddInternal(other.High, other.Low);
+    public void Add(GpuUInt128 other) => Add(other.High, other.Low);
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private void AddInternal(ulong otherHigh, ulong otherLow)
+    public void Add(ulong otherHigh, ulong otherLow)
     {
         ulong originalLow = Low;
         // We're reusing originalLow for carry to avoid an extra local.
@@ -316,7 +316,7 @@ public struct GpuUInt128 : IComparable<GpuUInt128>, IEquatable<GpuUInt128>
             return;
         }
 
-        AddInternal(valueHigh, valueLow);
+        Add(valueHigh, valueLow);
         if (CompareTo(in modulus) >= 0)
         {
             Sub(in modulus);
@@ -589,9 +589,6 @@ public struct GpuUInt128 : IComparable<GpuUInt128>, IEquatable<GpuUInt128>
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static GpuUInt128 BinaryGcd(GpuUInt128 u, GpuUInt128 v)
     {
-        // TODO: Replace this scalar binary GCD with the branchless reduction from
-        // GpuUInt128BinaryGcdBenchmarks so CPU fallbacks stay aligned with the GPU kernel
-        // performance when resolving large divisor residues.
         if (u.IsZero)
         {
             return v;
@@ -602,26 +599,50 @@ public struct GpuUInt128 : IComparable<GpuUInt128>, IEquatable<GpuUInt128>
             return u;
         }
 
-        int shift = TrailingZeroCount(new GpuUInt128(u.High | v.High, u.Low | v.Low));
+        ulong combinedLow = u.Low | v.Low;
+        int shift;
+        if (combinedLow != 0UL)
+        {
+            shift = BitOperations.TrailingZeroCount(combinedLow);
+        }
+        else
+        {
+            ulong combinedHigh = u.High | v.High;
+            shift = 64 + BitOperations.TrailingZeroCount(combinedHigh);
+        }
+
         int zu = TrailingZeroCount(u);
         u >>= zu;
 
-        do
+        GpuUInt128 currentV = v;
+        while (true)
         {
-            int zv = TrailingZeroCount(v);
-            v >>= zv;
-            if (u > v)
+            int zv = TrailingZeroCount(currentV);
+            currentV >>= zv;
+
+            bool swap = u > currentV;
+            GpuUInt128 min = ConditionalSelect(u, currentV, swap);
+            GpuUInt128 max = ConditionalSelect(currentV, u, swap);
+            max.Sub(min);
+
+            if (max.IsZero)
             {
-                (u, v) = (v, u);
+                return min << shift;
             }
 
-            v -= u;
+            u = min;
+            currentV = max;
         }
-        while (!v.IsZero);
-
-        return u << shift;
     }
 
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static GpuUInt128 ConditionalSelect(in GpuUInt128 left, in GpuUInt128 right, bool useRight)
+    {
+        ulong mask = useRight ? ulong.MaxValue : 0UL;
+        return new GpuUInt128(
+            left.High ^ ((left.High ^ right.High) & mask),
+            left.Low ^ ((left.Low ^ right.Low) & mask));
+    }
 
     public void Sub(ulong value)
     {
@@ -1529,6 +1550,15 @@ public struct GpuUInt128 : IComparable<GpuUInt128>, IEquatable<GpuUInt128>
         high = hi;
     }
 
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public void Pow2()
+    {
+        ulong operand = Low;
+        Low = operand * operand;
+        ulong highProduct = operand * High;
+        High = highProduct + MulHigh(operand, operand);
+    }
+
 }
 
 // TODO: Check if the TODO below is still relevant.

PerfectNumbers.Core/Gpu/HeuristicGpuDivisorTables.cs

@@ -1,146 +0,0 @@
-using System.Runtime.CompilerServices;
-using ILGPU;
-using ILGPU.Runtime;
-
-namespace PerfectNumbers.Core.Gpu;
-
-internal readonly struct HeuristicGpuDivisorTables
-{
-        public static readonly HeuristicGpuDivisorTables Shared = default;
-
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding1;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding3;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding7;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding9;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding1;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding3;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding7;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding9;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisors;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisorSquares;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding1;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding1;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding7;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding7;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding9;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding9;
-
-        public HeuristicGpuDivisorTables(
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding1,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding3,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding7,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding9,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding1,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding3,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding7,
-                ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding9,
-                ArrayView1D<ulong, Stride1D.Dense> groupADivisors,
-                ArrayView1D<ulong, Stride1D.Dense> groupADivisorSquares,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding1,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding1,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding7,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding7,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding9,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding9)
-        {
-                CombinedDivisorsEnding1 = combinedDivisorsEnding1;
-                CombinedDivisorsEnding3 = combinedDivisorsEnding3;
-                CombinedDivisorsEnding7 = combinedDivisorsEnding7;
-                CombinedDivisorsEnding9 = combinedDivisorsEnding9;
-                CombinedDivisorSquaresEnding1 = combinedDivisorSquaresEnding1;
-                CombinedDivisorSquaresEnding3 = combinedDivisorSquaresEnding3;
-                CombinedDivisorSquaresEnding7 = combinedDivisorSquaresEnding7;
-                CombinedDivisorSquaresEnding9 = combinedDivisorSquaresEnding9;
-                GroupADivisors = groupADivisors;
-                GroupADivisorSquares = groupADivisorSquares;
-                GroupBDivisorsEnding1 = groupBDivisorsEnding1;
-                GroupBDivisorSquaresEnding1 = groupBDivisorSquaresEnding1;
-                GroupBDivisorsEnding7 = groupBDivisorsEnding7;
-                GroupBDivisorSquaresEnding7 = groupBDivisorSquaresEnding7;
-                GroupBDivisorsEnding9 = groupBDivisorsEnding9;
-                GroupBDivisorSquaresEnding9 = groupBDivisorSquaresEnding9;
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ArrayView1D<ulong, Stride1D.Dense> SelectCombined(byte ending)
-        {
-                return ending switch
-                {
-                        1 => CombinedDivisorsEnding1,
-                        3 => CombinedDivisorsEnding3,
-                        7 => CombinedDivisorsEnding7,
-                        9 => CombinedDivisorsEnding9,
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ArrayView1D<ulong, Stride1D.Dense> SelectCombinedSquares(byte ending)
-        {
-                return ending switch
-                {
-                        1 => CombinedDivisorSquaresEnding1,
-                        3 => CombinedDivisorSquaresEnding3,
-                        7 => CombinedDivisorSquaresEnding7,
-                        9 => CombinedDivisorSquaresEnding9,
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ArrayView1D<ulong, Stride1D.Dense> SelectGroupB(byte ending)
-        {
-                return ending switch
-                {
-                        1 => GroupBDivisorsEnding1,
-                        7 => GroupBDivisorsEnding7,
-                        9 => GroupBDivisorsEnding9,
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ArrayView1D<ulong, Stride1D.Dense> SelectGroupBSquares(byte ending)
-        {
-                return ending switch
-                {
-                        1 => GroupBDivisorSquaresEnding1,
-                        7 => GroupBDivisorSquaresEnding7,
-                        9 => GroupBDivisorSquaresEnding9,
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ArrayView1D<ulong, Stride1D.Dense> SelectDivisors(HeuristicGpuDivisorTableKind kind, byte ending)
-        {
-                return kind switch
-                {
-                        HeuristicGpuDivisorTableKind.GroupA => GroupADivisors,
-                        HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorsEnding1,
-                        HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorsEnding7,
-                        HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorsEnding9,
-                        HeuristicGpuDivisorTableKind.Combined => SelectCombined(ending),
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        [MethodImpl(MethodImplOptions.AggressiveInlining)]
-        public ArrayView1D<ulong, Stride1D.Dense> SelectDivisorSquares(HeuristicGpuDivisorTableKind kind, byte ending)
-        {
-                return kind switch
-                {
-                        HeuristicGpuDivisorTableKind.GroupA => GroupADivisorSquares,
-                        HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorSquaresEnding1,
-                        HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorSquaresEnding7,
-                        HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorSquaresEnding9,
-                        HeuristicGpuDivisorTableKind.Combined => SelectCombinedSquares(ending),
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        internal static void InitializeShared(in HeuristicGpuDivisorTables tables)
-        {
-                Unsafe.AsRef(in Shared) = tables;
-        }
-}

PerfectNumbers.Core/Gpu/KernelContainer.cs

@@ -0,0 +1,58 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu;
+
+public sealed class KernelContainer
+{
+	// Serializes first-time initialization of kernels/buffers per accelerator.
+	public Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>? Order;
+
+	public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
+	ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>? Incremental;
+
+	public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
+		ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>,
+		ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>,
+		ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>? Pow2Mod;
+
+	public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
+			ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>? IncrementalOrder;
+
+	public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
+		ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>? Pow2ModOrder;
+
+	public Kernel? SmallPrimeFactor;
+	
+	public Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>? SpecialMax;
+
+	// Optional device buffer with small divisor cycles (<= 4M). Index = divisor, value = cycle length.
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallCycles;
+
+	public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastOne;
+	public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastSeven;
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimesPow2LastOne;
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimesPow2LastSeven;
+
+	public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimeFactorsPrimes;
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorsSquares;
+
+	public void Dispose()
+	{
+		// Order = null;
+		// Incremental = null;
+		// Pow2Mod = null;
+		// IncrementalOrder = null;
+		// Pow2ModOrder = null;
+		// SmallPrimeFactor = null;
+		// SpecialMax = null;
+
+		// SmallCycles?.Dispose();
+		// SmallPrimesLastOne?.Dispose();
+		// SmallPrimesLastSeven?.Dispose();
+		// SmallPrimesPow2LastOne?.Dispose();
+		// SmallPrimesPow2LastSeven?.Dispose();
+		// SmallPrimeFactorsPrimes?.Dispose();
+		// SmallPrimeFactorsSquares?.Dispose();
+	}
+}

PerfectNumbers.Core/Gpu/KernelType.cs

@@ -0,0 +1,14 @@
+namespace PerfectNumbers.Core.Gpu;
+
+[Flags]
+public enum KernelType
+{
+	None,
+	SmallCycles,
+	SmallPrimes,
+	Pow2ModKernelScan,
+	OrderKernelScan,
+	IncrementalKernelScan,
+	IncrementalOrderKernelScan,
+	Pow2ModOrderKernelScan
+}

PerfectNumbers.Core/Gpu/Kernels/DivisorByDivisorKernels.cs

@@ -1,7 +1,4 @@
 using ILGPU;
-using ILGPU.Algorithms;
-using ILGPU.Runtime;
-using PerfectNumbers.Core;
 
 namespace PerfectNumbers.Core.Gpu;
 

PerfectNumbers.Core/Gpu/Kernels/DivisorKernels.cs

@@ -1,6 +1,5 @@
 using ILGPU;
 using ILGPU.Algorithms;
-using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 

PerfectNumbers.Core/Gpu/Kernels/IncrementalKernels.cs

@@ -1,5 +1,4 @@
 using ILGPU;
-using ILGPU.Algorithms;
 using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;

PerfectNumbers.Core/Gpu/Kernels/LucasLehmerKernels.cs

@@ -1,7 +1,5 @@
-using System;
 using System.Runtime.CompilerServices;
 using ILGPU;
-using ILGPU.Runtime;
 using UInt128 = System.UInt128;
 
 namespace PerfectNumbers.Core.Gpu;
@@ -305,7 +303,7 @@ internal static class LucasLehmerKernels
         states[idx] = s;
     }
 
-    public static void Kernel(Index1D index, ulong exponent, GpuUInt128 modulus, ArrayView<GpuUInt128> state)
+    public static void Kernel(Index1D _, ulong exponent, GpuUInt128 modulus, ArrayView<GpuUInt128> state)
     {
         // Lucas–Lehmer iteration in the field GF(2^p-1).
         // For p < 128, use a Mersenne-specific squaring + reduction to avoid

PerfectNumbers.Core/Gpu/Kernels/NttBarrettKernels.cs

@@ -1,5 +1,6 @@
 using System.Runtime.CompilerServices;
 using ILGPU;
+using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 
@@ -32,7 +33,7 @@ internal static class NttBarrettKernels
     }
 
     // Barrett-based squaring: v = v^2 mod n for 128-bit n
-    public static void SquareBarrett128Kernel(Index1D index, ArrayView<GpuUInt128> data, ulong modHigh, ulong modLow, ulong muHigh, ulong muLow)
+    public static void SquareBarrett128Kernel(Index1D index, AcceleratorStream stream, ArrayView<GpuUInt128> data, ulong modHigh, ulong modLow, ulong muHigh, ulong muLow)
     {
         var v = data[index];
         Mul128Full(v.High, v.Low, v.High, v.Low, out var z3, out var z2, out var z1, out var z0);

PerfectNumbers.Core/Gpu/Kernels/NttMontgomeryKernels.cs

@@ -1,5 +1,6 @@
 using System.Runtime.CompilerServices;
 using ILGPU;
+using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 
@@ -43,7 +44,7 @@ internal static class NttMontgomeryKernels
         data[index] = new GpuUInt128(0UL, v);
     }
 
-    public static void SquareMont64Kernel(Index1D index, ArrayView<GpuUInt128> data, ulong modulus, ulong nPrime)
+    public static void SquareMont64Kernel(Index1D index, AcceleratorStream stream, ArrayView<GpuUInt128> data, ulong modulus, ulong nPrime)
     {
         ulong aR = data[index].Low;
         ulong v = MontMul64(aR, aR, modulus, nPrime);

PerfectNumbers.Core/Gpu/Kernels/OrderKernels.cs

@@ -1,5 +1,4 @@
 using ILGPU;
-using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 

PerfectNumbers.Core/Gpu/Kernels/Pow2ModKernels.cs

@@ -1,5 +1,4 @@
 using ILGPU;
-using ILGPU.Algorithms;
 using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;

PerfectNumbers.Core/Gpu/Kernels/Pow2MontgomeryKernels.cs

@@ -1,26 +1,80 @@
+using System.Runtime.CompilerServices;
 using ILGPU;
 using ILGPU.Runtime;
-using PerfectNumbers.Core;
 
-namespace PerfectNumbers.Core.Gpu;
+namespace PerfectNumbers.Core.Gpu.Kernels;
 
 internal static class Pow2MontgomeryKernels
 {
-    public static void Pow2MontgomeryKernelKeepMontgomery(
-        Index1D index,
-        ArrayView1D<ulong, Stride1D.Dense> exponents,
-        MontgomeryDivisorData divisor,
-        ArrayView1D<ulong, Stride1D.Dense> results)
-    {
-        results[index] = ULongExtensions.Pow2MontgomeryModWindowedGpuKeepMontgomery(divisor, exponents[index]);
-    }
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static GpuUInt128 AdvancePolynomialGpu(GpuUInt128 x, ulong c, ulong modulus)
+	{
+		x.Pow2();
+		x.AddMod(c, modulus);
+		return x;
+	}
 
-    public static void Pow2MontgomeryKernelConvertToStandard(
-        Index1D index,
-        ArrayView1D<ulong, Stride1D.Dense> exponents,
-        MontgomeryDivisorData divisor,
-        ArrayView1D<ulong, Stride1D.Dense> results)
-    {
-        results[index] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponents[index]);
-    }
+	public static void Pow2MontgomeryKernelKeepMontgomery(
+		Index1D index,
+		ArrayView1D<ulong, Stride1D.Dense> exponents,
+		MontgomeryDivisorData divisor,
+		ArrayView1D<ulong, Stride1D.Dense> results)
+	{
+		results[0] = ULongExtensions.Pow2MontgomeryModWindowedGpuKeepMontgomery(divisor, exponents[0]);
+	}
+
+	public static void Pow2MontgomeryKernelConvertToStandard(
+		Index1D index,
+		ArrayView1D<ulong, Stride1D.Dense> exponents,
+		MontgomeryDivisorData divisor,
+		ArrayView1D<ulong, Stride1D.Dense> results)
+	{
+		results[0] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponents[0]);
+	}
+
+
+	public static void TryPollardRhoKernel(ulong n, int limit, ArrayView1D<ulong, Stride1D.Dense> randomState, ArrayView1D<byte, Stride1D.Dense> factored, ArrayView1D<ulong, Stride1D.Dense> factor)
+	{
+		factored[0] = 0;
+		if ((n & 1UL) == 0UL)
+		{
+			factor[0] = 2UL;
+			factored[0] = 1;
+			return;
+		}
+
+		int attempt = 0;
+		GpuUInt128 nGpuUInt128 = (GpuUInt128)n;
+		DeterministicRandomGpu random = new(randomState[0]);
+		while (attempt < limit)
+		{
+			ulong c = (random.NextUInt64() % (n - 1UL)) + 1UL;
+			GpuUInt128 x = (random.NextUInt64() % (n - 2UL)) + 2UL;
+			GpuUInt128 y = x;
+			GpuUInt128 d = GpuUInt128.One;
+
+			while (d.CompareTo(GpuUInt128.One) == 0)
+			{
+				x = AdvancePolynomialGpu(x, c, modulus: n);
+				y = AdvancePolynomialGpu(y, c, n);
+				y = AdvancePolynomialGpu(y, c, n);
+				GpuUInt128 diff = x.CompareTo(y) > 0 ? x - y : y - x;
+				d = GpuUInt128.BinaryGcd(diff, nGpuUInt128);
+			}
+
+			if (d != n)
+			{
+				if (d.High == 0)
+				{
+					factor[0] = d.Low;
+					factored[0] = 1;
+					return;
+				}
+			}
+
+			attempt++;
+		}
+
+		randomState[0] = random.State;
+	}
 }

PerfectNumbers.Core/Gpu/Kernels/PrimeOrderSpecialMaxKernels.cs

@@ -0,0 +1,79 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu;
+
+internal static partial class PrimeOrderGpuHeuristics
+{
+    internal static void EvaluateSpecialMaxCandidatesKernel(
+        Index1D index,
+        ulong phi,
+        ArrayView1D<ulong, Stride1D.Dense> factors,
+        int factorCount,
+        MontgomeryDivisorData divisor,
+        ArrayView1D<ulong, Stride1D.Dense> candidates,
+        ArrayView1D<ulong, Stride1D.Dense> resultOut)
+    {
+        if (index != 0)
+        {
+            return;
+        }
+
+        int actual = 0;
+        for (int i = 0; i < factorCount; i++)
+        {
+            ulong factor = factors[i];
+            if (factor <= 1UL)
+            {
+                continue;
+            }
+
+            ulong reduced = phi / factor;
+            if (reduced == 0UL)
+            {
+                continue;
+            }
+
+            candidates[actual] = reduced;
+            actual++;
+        }
+
+        if (actual == 0)
+        {
+            resultOut[0] = 1;
+            return;
+        }
+
+        for (int i = 1; i < actual; i++)
+        {
+            ulong current = candidates[i];
+            int insert = i - 1;
+            while (insert >= 0 && candidates[insert] > current)
+            {
+                candidates[insert + 1] = candidates[insert];
+                insert--;
+            }
+
+            candidates[insert + 1] = current;
+        }
+
+        var divisorPartial = new GpuDivisorPartialData(divisor.Modulus);
+        var stepper = new ExponentRemainderStepperGpu(divisorPartial);
+        if (stepper.InitializeIsUnityGpu(candidates[0]))
+        {
+			Atomic.Exchange(ref resultOut[0], 0);
+            return;
+        }
+
+        for (int i = 1; i < actual; i++)
+        {
+            if (stepper.ComputeNextIsUnityGpu(candidates[i]))
+            {
+				Atomic.Exchange(ref resultOut[0], 0);
+                return;
+            }
+        }
+
+        Atomic.Exchange(ref resultOut[0], 1);
+    }
+}

PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs

@@ -2,41 +2,39 @@ using System.Runtime.CompilerServices;
 using ILGPU;
 using ILGPU.Algorithms;
 using ILGPU.Runtime;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core.Gpu;
 
-internal enum HeuristicGpuDivisorTableKind : byte
+public enum HeuristicGpuDivisorTableKind : byte
 {
     GroupA = 0,
     GroupBEnding1 = 1,
     GroupBEnding7 = 7,
     GroupBEnding9 = 9,
-    Combined = 255,
 }
 
 internal static class PrimeTesterKernels
 {
     // GPU kernel: small-prime sieve only. Returns 1 if passes sieve (probable prime), 0 otherwise.
     public static void SmallPrimeSieveKernel(
-        Index1D index,
-        ArrayView<ulong> numbers,
-        ArrayView<uint> smallPrimes,
-        ArrayView<uint> smallPrimesLastOne,
-        ArrayView<uint> smallPrimesLastSeven,
-        ArrayView<uint> smallPrimesLastThree,
-        ArrayView<uint> smallPrimesLastNine,
-        ArrayView<ulong> smallPrimesPow2,
-        ArrayView<ulong> smallPrimesPow2LastOne,
-        ArrayView<ulong> smallPrimesPow2LastSeven,
-        ArrayView<ulong> smallPrimesPow2LastThree,
-        ArrayView<ulong> smallPrimesPow2LastNine,
-        ArrayView<byte> results)
+		Index1D index,
+		ArrayView<ulong> numbers,
+		ArrayView<uint> smallPrimesLastOne,
+		ArrayView<uint> smallPrimesLastSeven,
+		ArrayView<uint> smallPrimesLastThree,
+		ArrayView<uint> smallPrimesLastNine,
+		ArrayView<ulong> smallPrimesPow2LastOne,
+		ArrayView<ulong> smallPrimesPow2LastSeven,
+		ArrayView<ulong> smallPrimesPow2LastThree,
+		ArrayView<ulong> smallPrimesPow2LastNine,
+		ArrayView<byte> results)
     {
         ulong n = numbers[index];
         byte result = 1;
 
-        ArrayView<uint> primes = smallPrimes;
-        ArrayView<ulong> primeSquares = smallPrimesPow2;
+        ArrayView<uint> primes;
+        ArrayView<ulong> primeSquares;
         ulong lastDigit = n % 10UL;
         switch (lastDigit)
         {
@@ -55,7 +53,9 @@ internal static class PrimeTesterKernels
             case 9UL:
                 primes = smallPrimesLastNine;
                 primeSquares = smallPrimesPow2LastNine;
-                break;
+				break;
+			default:
+				return;
         }
 
         int length = (int)primes.Length;
@@ -78,17 +78,49 @@ internal static class PrimeTesterKernels
         results[index] = result;
     }
 
-    public static void HeuristicTrialDivisionKernel(
+    public static void HeuristicTrialCombinedDivisionKernel(
+        Index1D index,
+		byte nMod10,
+        ArrayView<int> resultFlag,
+        ulong n,
+        ulong maxDivisorSquare,
+        HeuristicCombinedGpuViews tables)
+    {
+        // byte nMod10 = (byte)(n % 10UL);
+        ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(nMod10);
+        ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(nMod10);
+
+        int divisorLength = (int)divisorSquares.Length;
+        int threadIndex = index;
+        if (threadIndex >= divisorLength)
+        {
+            return;
+        }
+
+        ulong divisorSquare = divisorSquares[threadIndex];
+        if (divisorSquare > maxDivisorSquare)
+        {
+            return;
+        }
+
+        ulong divisor = divisors[threadIndex];
+        divisor = n % divisor;
+        if (divisor == 0UL)
+        {
+            resultFlag[0] = 1;
+        }
+    }
+
+    public static void HeuristicTrialGroupABDivisionKernel(
         Index1D index,
         ArrayView<int> resultFlag,
         ulong n,
         ulong maxDivisorSquare,
         HeuristicGpuDivisorTableKind tableKind,
-        HeuristicGpuDivisorTables tables)
+        HeuristicGroupABGpuViews tables)
     {
-        byte nMod10 = (byte)(n % 10UL);
-        ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(tableKind, nMod10);
-        ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(tableKind, nMod10);
+        ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(tableKind);
+        ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(tableKind);
 
         int divisorLength = (int)divisorSquares.Length;
         int threadIndex = index;
@@ -121,35 +153,37 @@ internal static class PrimeTesterKernels
 
     private static ulong BinaryGcdGpu(ulong u, ulong v)
     {
-        // TODO: Replace this inline GPU binary GCD with the kernel extracted from
-        // GpuUInt128BinaryGcdBenchmarks via GpuKernelPool so device callers reuse the
-        // fully unrolled ladder instead of this branchy fallback.
         bool uIsZero = u == 0UL;
         bool vIsZero = v == 0UL;
         bool eitherZero = uIsZero | vIsZero;
         ulong zeroResult = uIsZero ? v : u;
 
-        ulong combined = u | v | 1UL;
+        ulong combined = (u | v) | 1UL;
         int shift = XMath.TrailingZeroCount(combined);
+        ulong currentU = u >> XMath.TrailingZeroCount(u | 1UL);
+        ulong currentV = v >> XMath.TrailingZeroCount(v | 1UL);
 
-        ulong normalizedU = u >> XMath.TrailingZeroCount(u | 1UL);
-        ulong normalizedV = v >> XMath.TrailingZeroCount(v | 1UL);
-
-        ulong currentU = normalizedU;
-        ulong currentV = normalizedV;
-        bool loopCondition = currentV != 0UL;
-
-        while (loopCondition)
+        while (true)
         {
             currentV >>= XMath.TrailingZeroCount(currentV);
-            ulong minValue = XMath.Min(currentU, currentV);
-            ulong maxValue = XMath.Max(currentU, currentV);
-            currentU = minValue;
+            bool swap = currentU > currentV;
+            ulong minValue = ConditionalSelect(currentU, currentV, swap);
+            ulong maxValue = ConditionalSelect(currentV, currentU, swap);
             currentV = maxValue - minValue;
-            loopCondition = currentV != 0UL;
+            currentU = minValue;
+            if (currentV == 0UL)
+            {
+                ulong gcd = currentU << shift;
+                return eitherZero ? zeroResult : gcd;
+            }
         }
-
-        ulong gcd = currentU << shift;
-        return eitherZero ? zeroResult : gcd;
     }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static ulong ConditionalSelect(ulong left, ulong right, bool useRight)
+    {
+        ulong mask = useRight ? ulong.MaxValue : 0UL;
+        return left ^ ((left ^ right) & mask);
+    }
+
 }

PerfectNumbers.Core/Gpu/MersenneNumberDivisorGpuTester.cs

@@ -1,4 +1,3 @@
-using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
 
@@ -9,127 +8,157 @@ namespace PerfectNumbers.Core.Gpu;
 /// </summary>
 public sealed class MersenneNumberDivisorGpuTester
 {
-	private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>>> _kernelCache = new();
-	private readonly ConcurrentDictionary<Accelerator, MemoryBuffer1D<byte, Stride1D.Dense>> _resultBuffers = new();
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Queue<MemoryBuffer1D<byte, Stride1D.Dense>>>? _resultBuffers;
 
-	private Action<Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>> GetKernel(Accelerator accelerator) =>
-			_kernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>>(DivisorKernels.Kernel));
+	private static Queue<MemoryBuffer1D<byte, Stride1D.Dense>> GetQueue(Accelerator accelerator)
+	{
+		var pool = _resultBuffers ??= [];
+		if (!pool.TryGetValue(accelerator, out var bufferQueue))
+		{
+			bufferQueue = [];
+			pool[accelerator] = bufferQueue;
+		}
 
-        public static void BuildDivisorCandidates()
-        {
-                ulong[] snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
-                (ulong divisor, uint cycle)[] list = new (ulong divisor, uint cycle)[snapshot.Length / 2];
-                ulong cycle;
-                int count = 0, i, snapshotLength = snapshot.Length;
-                for (i = 3; i < snapshotLength; i += 2)
-                {
-                        cycle = snapshot[i];
-                        if (cycle == 0U)
-                        {
-                                continue;
-                        }
+		return bufferQueue;
+	}
 
-                        list[count++] = ((ulong)i, (uint)cycle);
-                }
+	private static Action<AcceleratorStream, Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>> GetDivisorKernel(Accelerator accelerator)
+	{
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>>(DivisorKernels.Kernel);
 
-                _divisorCandidates = count == 0 ? [] : list[..count];
-        }
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>>>();
+	}
+
+	public static void BuildDivisorCandidates()
+	{
+		ulong[] snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
+		(ulong divisor, uint cycle)[] list = new (ulong divisor, uint cycle)[snapshot.Length / 2];
+		ulong cycle;
+		int count = 0, i, snapshotLength = snapshot.Length;
+		for (i = 3; i < snapshotLength; i += 2)
+		{
+			cycle = snapshot[i];
+			if (cycle == 0U)
+			{
+				continue;
+			}
+
+			list[count++] = ((ulong)i, (uint)cycle);
+		}
+
+		_divisorCandidates = count == 0 ? [] : list[..count];
+	}
+
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
+	public bool IsDivisible(ulong exponent, in ReadOnlyGpuUInt128 divisor)
+	{
+		// var accelerator = SharedGpuContext.CreateAccelerator();
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+		var stream = accelerator.CreateStream();
+		Queue<MemoryBuffer1D<byte, Stride1D.Dense>> resultBuffers = GetQueue(accelerator);
+		if (!resultBuffers.TryDequeue(out var resultBuffer))
+		{
+			// lock(accelerator)
+			{
+				resultBuffer = accelerator.Allocate1D<byte>(1);
+			}
+		}
 
-    public bool IsDivisible(ulong exponent, in ReadOnlyGpuUInt128 divisor)
-    {
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var kernel = GetKernel(accelerator);
-        var resultBuffer = _resultBuffers.GetOrAdd(accelerator, acc => acc.Allocate1D<byte>(1));
 		// There is no point in clearing this buffer. We always override item [0] and never use it beyond item [0]
-        // resultBuffer.MemSetToZero();
-        kernel(1, exponent, divisor, resultBuffer.View);
-        accelerator.Synchronize();
-        Span<byte> result = stackalloc byte[1];
-        resultBuffer.View.CopyToCPU(ref result[0], 1);
-        bool divisible = result[0] != 0;
-        result[0] = 0;
-        resultBuffer.View.CopyFromCPU(ref result[0], 1);
-        accelerator.Synchronize();
-        gpu.Dispose();
-        return divisible;
-    }
+		// resultBuffer.MemSetToZero(stream);
+		var divisorKernel = GetDivisorKernel(accelerator);
+		divisorKernel(stream, 1, exponent, divisor, resultBuffer.View);
+		byte result = 0;
+		resultBuffer.View.CopyToCPU(stream, ref result, 1);
+		stream.Synchronize();
+		bool divisible = result != 0;
+		resultBuffers.Enqueue(resultBuffer);
+		stream.Dispose();
+		// AcceleratorPool.Shared.Return(accelerator);
+		// accelerator.Dispose();
+		return divisible;
+	}
 
-        private static (ulong divisor, uint cycle)[]? _divisorCandidates = Array.Empty<(ulong divisor, uint cycle)>();
+	private static (ulong divisor, uint cycle)[]? _divisorCandidates = [];
 
-    public bool IsPrime(ulong p, UInt128 d, ulong divisorCyclesSearchLimit, out bool divisorsExhausted)
-    {
-        ReadOnlyGpuUInt128 readOnlyDivisor;
+	public bool IsPrime(ulong p, UInt128 d, ulong divisorCyclesSearchLimit, out bool divisorsExhausted)
+	{
+		ReadOnlyGpuUInt128 readOnlyDivisor;
 
-        if (d != UInt128.Zero)
-        {
-            readOnlyDivisor = new ReadOnlyGpuUInt128(d);
-            if (IsDivisible(p, in readOnlyDivisor))
-            {
-                divisorsExhausted = true;
-                return false;
-            }
+		if (d != UInt128.Zero)
+		{
+			readOnlyDivisor = new ReadOnlyGpuUInt128(d);
+			if (IsDivisible(p, in readOnlyDivisor))
+			{
+				divisorsExhausted = true;
+				return false;
+			}
 
-            divisorsExhausted = false;
-            return true;
-        }
+			divisorsExhausted = false;
+			return true;
+		}
 
-        if (_divisorCandidates is { Length: > 0 } candidates)
-        {
-            int candidateCount = candidates.Length;
-            for (int index = 0; index < candidateCount; index++)
-            {
-                (ulong candidateDivisor, uint cycle) = candidates[index];
-                if (p % cycle != 0UL)
-                {
-                    continue;
-                }
+		if (_divisorCandidates is { Length: > 0 } candidates)
+		{
+			int candidateCount = candidates.Length;
+			for (int index = 0; index < candidateCount; index++)
+			{
+				(ulong candidateDivisor, uint cycle) = candidates[index];
+				if (p % cycle != 0UL)
+				{
+					continue;
+				}
 
-                readOnlyDivisor = new ReadOnlyGpuUInt128(candidateDivisor); // Reusing readOnlyDivisor for candidate divisors.
-                if (IsDivisible(p, in readOnlyDivisor))
-                {
-                    divisorsExhausted = true;
-                    return false;
-                }
-            }
-        }
+				readOnlyDivisor = new ReadOnlyGpuUInt128(candidateDivisor); // Reusing readOnlyDivisor for candidate divisors.
+				if (IsDivisible(p, in readOnlyDivisor))
+				{
+					divisorsExhausted = true;
+					return false;
+				}
+			}
+		}
 
-        UInt128 kMul2;
-        var divisorCycles = MersenneDivisorCycles.Shared;
-        UInt128 maxK2 = UInt128.MaxValue / ((UInt128)p << 1);
-        ulong limit = divisorCyclesSearchLimit;
-        if ((UInt128)limit > maxK2)
-        {
-            limit = (ulong)maxK2;
-        }
+		UInt128 kMul2;
+		var divisorCycles = MersenneDivisorCycles.Shared;
+		UInt128 maxK2 = UInt128.MaxValue / ((UInt128)p << 1);
+		ulong limit = divisorCyclesSearchLimit;
+		if ((UInt128)limit > maxK2)
+		{
+			limit = (ulong)maxK2;
+		}
 
-        for (ulong k2 = 1UL; k2 <= limit; k2++)
-        {
-            kMul2 = (UInt128)k2 << 1;
-            UInt128 candidate = checked(kMul2 * p);
-            d = checked(candidate + UInt128.One);
-            if (p < 64UL && d == ((UInt128)1 << (int)p) - UInt128.One)
-            {
-                continue;
-            }
+		for (ulong k2 = 1UL; k2 <= limit; k2++)
+		{
+			kMul2 = (UInt128)k2 << 1;
+			UInt128 candidate = checked(kMul2 * p);
+			d = checked(candidate + UInt128.One);
+			if (p < 64UL && d == ((UInt128)1 << (int)p) - UInt128.One)
+			{
+				continue;
+			}
 
-            UInt128 cycle128 = MersenneDivisorCycles.GetCycle(d);
-            if ((UInt128)p % cycle128 != UInt128.Zero)
-            {
-                continue;
-            }
+			UInt128 cycle128 = MersenneDivisorCycles.GetCycle(d);
+			if ((UInt128)p % cycle128 != UInt128.Zero)
+			{
+				continue;
+			}
 
-            readOnlyDivisor = new ReadOnlyGpuUInt128(d); // Reusing readOnlyDivisor for generated divisors.
-            if (IsDivisible(p, in readOnlyDivisor))
-            {
-                divisorsExhausted = true;
-                return false;
-            }
-        }
+			readOnlyDivisor = new ReadOnlyGpuUInt128(d); // Reusing readOnlyDivisor for generated divisors.
+			if (IsDivisible(p, in readOnlyDivisor))
+			{
+				divisorsExhausted = true;
+				return false;
+			}
+		}
 
-        divisorsExhausted = (UInt128)divisorCyclesSearchLimit >= maxK2;
-        return true;
-    }
+		divisorsExhausted = (UInt128)divisorCyclesSearchLimit >= maxK2;
+		return true;
+	}
 
 }
 

PerfectNumbers.Core/Gpu/MersenneNumberIncrementalGpuTester.cs

@@ -12,36 +12,41 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
     private readonly GpuKernelType _kernelType = kernelType;
     private readonly bool _useGpuOrder = useGpuOrder;
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
     public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
     {
         throw new NotImplementedException($"GPU incremental scanning requires the device cycle heuristics implementation (kernel: {_kernelType}, GPU order: {_useGpuOrder}).");
 
-        var gpuLease = GpuKernelPool.GetKernel(_useGpuOrder);
-        var accelerator = gpuLease.Accelerator;
-        var stream = gpuLease.Stream;
+        // GpuPrimeWorkLimiter.Acquire();
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
         int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy
         UInt128 kStart = 1UL;
         ulong divMul = (ulong)((((UInt128)1 << 64) - UInt128.One) / exponent) + 1UL;
         byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0; // ILGPU kernels do not support bool parameters
 
-        var pow2Kernel = gpuLease.Pow2ModKernel;
-        var incKernel = gpuLease.IncrementalKernel;
+		KernelContainer kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, KernelType.Pow2ModKernelScan | KernelType.IncrementalKernelScan);
+		var pow2Kernel = kernels.Pow2Mod!;
+        var incKernel = kernels.Incremental!;
         ulong step10 = (exponent.Mod10() << 1).Mod10();
         ulong step8 = ((exponent & 7UL) << 1) & 7UL;
         ulong step3 = ((exponent % 3UL) << 1) % 3UL;
         ulong step5 = ((exponent % 5UL) << 1) % 5UL;
         GpuUInt128 twoPGpu = (GpuUInt128)twoP;
-        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(accelerator);
+        var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice(kernels);
         ResiduePrimeViews primeViews = default;
         if (_kernelType == GpuKernelType.Pow2Mod)
         {
-            primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(accelerator);
+            primeViews = GpuKernelPool.GetSmallPrimesOnDevice(kernels);
         }
 
         var orderBuffer = accelerator.Allocate1D<ulong>(batchSize);
-        // Avoid giant stack allocations that can trigger StackOverflow when batchSize is large.
-        // Rent a reusable array from the shared pool instead.
-        ulong[] orderArray = ArrayPool<ulong>.Shared.Rent(batchSize);
+		// Avoid giant stack allocations that can trigger StackOverflow when batchSize is large.
+		// Rent a reusable array from the shared pool instead.
+		ArrayPool<ulong> ulongPool = ThreadStaticPools.UlongPool;
+		ulong[] orderArray = ulongPool.Rent(batchSize);
         UInt128 remaining;
         int currentSize;
         int i;
@@ -92,8 +97,8 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
                     smallCyclesView);
             }
 
+            orderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(orders), currentSize);
             stream.Synchronize();
-            orderBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(orders), currentSize);
             if (_kernelType == GpuKernelType.Pow2Mod)
             {
                 for (i = 0; i < currentSize; i++)
@@ -120,8 +125,9 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
             kStart += (UInt128)currentSize;
         }
 
-        ArrayPool<ulong>.Shared.Return(orderArray);
-        orderBuffer.Dispose();
-        gpuLease.Dispose();
+		stream.Dispose();
+        ulongPool.Return(orderArray);
+		orderBuffer.Dispose();
+        // GpuPrimeWorkLimiter.Release();
     }
 }

PerfectNumbers.Core/Gpu/MersenneNumberLucasLehmerGpuTester.cs

@@ -8,32 +8,125 @@ namespace PerfectNumbers.Core.Gpu;
 
 public class MersenneNumberLucasLehmerGpuTester
 {
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>> KernelCache = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ulong, ArrayView<GpuUInt128>>> AddSmallKernelCache = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ulong, ArrayView<GpuUInt128>>> SubSmallKernelCache = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ulong, ArrayView<GpuUInt128>>> ReduceKernelCache = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>> IsZeroKernelCache = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>>> BatchKernelCache = new();
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>>? _lucasLehmerKernel;
 
-    private Action<Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>> GetKernel(Accelerator accelerator)
-    {
-        return KernelCache.GetOrAdd(accelerator, accel => accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>(LucasLehmerKernels.Kernel));
-    }
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>? _addSmallKernel;
 
-    private Action<Index1D, ulong, ArrayView<GpuUInt128>> GetAddSmallKernel(Accelerator accelerator) =>
-        AddSmallKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.AddSmallKernel));
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>? _subSmallKernel;
 	
-    private Action<Index1D, ulong, ArrayView<GpuUInt128>> GetSubSmallKernel(Accelerator accelerator) =>
-        SubSmallKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.SubtractSmallKernel));
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>? _reduceKernel;
 
-    private Action<Index1D, ulong, ArrayView<GpuUInt128>> GetReduceKernel(Accelerator accelerator) =>
-        ReduceKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.ReduceModMersenneKernel));
+	[ThreadStatic]
+    private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>>? _isZeroKernel;
 
-    private Action<Index1D, ArrayView<GpuUInt128>, ArrayView<byte>> GetIsZeroKernel(Accelerator accelerator) =>
-        IsZeroKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>(LucasLehmerKernels.IsZeroKernel));
+	[ThreadStatic]
+    private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>>>? _lucasLehmerBatchKernel;
 
-    private Action<Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>> GetBatchKernel(Accelerator accelerator) =>
-        BatchKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>>(LucasLehmerKernels.KernelBatch));
+    private static Action<AcceleratorStream, Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>> GetLucasLehmerKernel(Accelerator accelerator)
+	{
+		var pool = _lucasLehmerKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>(LucasLehmerKernels.Kernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>> GetAddSmallKernel(Accelerator accelerator)
+	{
+		var pool = _addSmallKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.AddSmallKernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>> GetSubSmallKernel(Accelerator accelerator)
+	{
+		var pool = _subSmallKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.SubtractSmallKernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>> GetReduceKernel(Accelerator accelerator)
+	{
+		var pool = _reduceKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.ReduceModMersenneKernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<byte>> GetIsZeroKernel(Accelerator accelerator)
+	{
+		var pool = _isZeroKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>(LucasLehmerKernels.IsZeroKernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>> GetLucasLehmerBatchKernel(Accelerator accelerator)
+	{
+		var pool = _lucasLehmerBatchKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+		
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>>(LucasLehmerKernels.KernelBatch);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
 
     // Configurable LL slice size to keep kernels short. Default 32.
     public int SliceSize = 32;
@@ -54,13 +147,11 @@ public class MersenneNumberLucasLehmerGpuTester
             return false;
         }
 
-        var limiter = GpuPrimeWorkLimiter.Acquire();
         bool result;
         if (exponent >= 128UL)
         {
             if (!TryGetNttParameters(exponent, out var nttMod, out var primitiveRoot))
             {
-                limiter.Dispose();
                 throw new NotSupportedException("NTT parameters for the given exponent are not available.");
             }
 
@@ -68,19 +159,22 @@ public class MersenneNumberLucasLehmerGpuTester
         }
         else
         {
-            var gpu = GpuContextPool.Rent();
-            var accelerator = gpu.Accelerator;
-            var kernel = GetKernel(accelerator);
+        	// GpuPrimeWorkLimiter.Acquire();
+            var acceleratorIndex = AcceleratorPool.Shared.Rent();
+			var accelerator = _accelerators[acceleratorIndex];
+            var stream = accelerator.CreateStream();
             var modulus = new GpuUInt128(((UInt128)1 << (int)exponent) - 1UL); // TODO: Cache these Mersenne moduli per exponent so LL GPU runs skip rebuilding them every launch.
-            var buffer = accelerator.Allocate1D<GpuUInt128>(1);
-            kernel(1, exponent, modulus, buffer.View);
-            accelerator.Synchronize();
+			var buffer = accelerator.Allocate1D<GpuUInt128>(1);
+			var lucasLehmerKernel = GetLucasLehmerKernel(accelerator);
+            lucasLehmerKernel(stream, 1, exponent, modulus, buffer.View);
+			stream.Synchronize();
+			// TODO: Replace this with .CopyToCPU with pre-allocated buffer using stream. Synchronize on the stream after copy.
             result = buffer.GetAsArray1D()[0].IsZero;
             buffer.Dispose();
-            gpu.Dispose();
+            stream.Dispose();
         }
 
-        limiter.Dispose();
+		// GpuPrimeWorkLimiter.Release();
         return result;
     }
 
@@ -100,16 +194,18 @@ public class MersenneNumberLucasLehmerGpuTester
             }
         }
 
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var kernel = GetBatchKernel(accelerator);
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
 
         var expBuffer = accelerator.Allocate1D<ulong>(count);
-        ulong[] expArray = ArrayPool<ulong>.Shared.Rent(count);
+		ArrayPool<ulong> ulongPool = ThreadStaticPools.UlongPool;
+		ulong[] expArray = ulongPool.Rent(count);
         exponents.CopyTo(expArray);
-        expBuffer.View.CopyFromCPU(ref expArray[0], count);
+        expBuffer.View.CopyFromCPU(stream, ref expArray[0], count);
 
-        GpuUInt128[] modulusArray = ArrayPool<GpuUInt128>.Shared.Rent(count);
+		ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
+		GpuUInt128[] modulusArray = gpuUInt128Pool.Rent(count);
         for (int i = 0; i < count; i++)
         {
             // TODO: Replace this per-exponent shift with a small shared table (one entry per supported exponent < 128)
@@ -118,19 +214,20 @@ public class MersenneNumberLucasLehmerGpuTester
         }
 
         var modBuffer = accelerator.Allocate1D<GpuUInt128>(count);
-        modBuffer.View.CopyFromCPU(ref modulusArray[0], count);
+        modBuffer.View.CopyFromCPU(stream, ref modulusArray[0], count);
 
-        var stateBuffer = accelerator.Allocate1D<GpuUInt128>(count);
-        kernel(count, expBuffer.View, modBuffer.View, stateBuffer.View);
-        accelerator.Synchronize();
-        stateBuffer.View.CopyToCPU(ref residues[0], count);
+		var stateBuffer = accelerator.Allocate1D<GpuUInt128>(count);
+		var lucasLehmerBatchKernel = GetLucasLehmerBatchKernel(accelerator);
+        lucasLehmerBatchKernel(stream, count, expBuffer.View, modBuffer.View, stateBuffer.View);
+        stateBuffer.View.CopyToCPU(stream, ref residues[0], count);
+        stream.Synchronize();
 
         stateBuffer.Dispose();
         modBuffer.Dispose();
         expBuffer.Dispose();
-        ArrayPool<GpuUInt128>.Shared.Return(modulusArray);
-        ArrayPool<ulong>.Shared.Return(expArray);
-        gpu.Dispose();
+        gpuUInt128Pool.Return(modulusArray);
+        ulongPool.Return(expArray);
+        stream.Dispose();
     }
 
     public void IsMersennePrimeBatch(ReadOnlySpan<ulong> exponents, Span<bool> results)
@@ -141,23 +238,27 @@ public class MersenneNumberLucasLehmerGpuTester
             throw new ArgumentException("Result span too small", nameof(results));
         }
 
-        GpuUInt128[] buffer = ArrayPool<GpuUInt128>.Shared.Rent(count);
+		ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
+		GpuUInt128[] buffer = gpuUInt128Pool.Rent(count);
         ComputeResidues(exponents, buffer.AsSpan(0, count));
         for (int i = 0; i < count; i++)
         {
             results[i] = buffer[i].IsZero;
         }
-        ArrayPool<GpuUInt128>.Shared.Return(buffer);
+        gpuUInt128Pool.Return(buffer);
     }
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
     private bool IsMersennePrimeNtt(ulong exponent, GpuUInt128 nttMod, GpuUInt128 primitiveRoot, bool runOnGpu)
     {
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var addKernel = GetAddSmallKernel(accelerator);
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
+
         var subKernel = GetSubSmallKernel(accelerator);
-        var reduceKernel = GetReduceKernel(accelerator);
-        var zeroKernel = GetIsZeroKernel(accelerator);
+		var reduceKernel = GetReduceKernel(accelerator);
+		var isZeroKernel = GetIsZeroKernel(accelerator);
 
         int limbCount = (int)((exponent + 127UL) / 128UL);
         int target = limbCount * 2;
@@ -167,8 +268,9 @@ public class MersenneNumberLucasLehmerGpuTester
             transformLength <<= 1;
         }
         var stateBuffer = accelerator.Allocate1D<GpuUInt128>(transformLength);
-        stateBuffer.MemSetToZero();
-        addKernel(1, 4UL, stateBuffer.View);
+		stateBuffer.MemSetToZero(stream);
+		var addSmallKernel = GetAddSmallKernel(accelerator);
+		addSmallKernel(stream, 1, 4UL, stateBuffer.View);
 
         // TODO(LL-SLICE): Slice Lucas–Lehmer iterations into short batches to
         // avoid long-running kernels and TDR. Example: process 8–64 iterations
@@ -183,27 +285,26 @@ public class MersenneNumberLucasLehmerGpuTester
             for (int s = 0; s < iter; s++)
             {
                 // Square in NTT domain with staged kernels (short runtime per kernel).
-                NttGpuMath.SquareDevice(accelerator, stateBuffer.View, nttMod, primitiveRoot);
-                subKernel(1, 2UL, stateBuffer.View);
-                reduceKernel(1, exponent, stateBuffer.View);
+                NttGpuMath.SquareDevice(accelerator, stream, stateBuffer.View, nttMod, primitiveRoot);
+                subKernel(stream, 1, 2UL, stateBuffer.View);
+                reduceKernel(stream, 1, exponent, stateBuffer.View);
             }
 
             // Synchronize between slices to yield to scheduler and avoid TDR.
-            accelerator.Synchronize();
             i += (ulong)iter;
         }
 
-        reduceKernel(1, exponent, stateBuffer.View);
+        reduceKernel(stream, 1, exponent, stateBuffer.View);
+		var resultBuffer = accelerator.Allocate1D<byte>(1);
+        isZeroKernel(stream, 1, stateBuffer.View, resultBuffer.View);
+        Span<byte> result = stackalloc byte[1];
+        resultBuffer.View.CopyToCPU(stream, result);
+		stream.Synchronize();
 		
-        var resultBuffer = accelerator.Allocate1D<byte>(1);
-        zeroKernel(1, stateBuffer.View, resultBuffer.View);
-        accelerator.Synchronize();
-        byte[] result = new byte[1];
-        resultBuffer.View.CopyToCPU(result);
         bool isPrime = result[0] != 0;
         resultBuffer.Dispose();
         stateBuffer.Dispose();
-        gpu.Dispose();
+        stream.Dispose();
         return isPrime;
     }
 
@@ -472,6 +573,4 @@ public class MersenneNumberLucasLehmerGpuTester
 
         return result;
     }
-
-
 }

PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs

@@ -4,62 +4,76 @@ namespace PerfectNumbers.Core.Gpu;
 
 public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuOrder)
 {
-        private readonly GpuKernelType _kernelType = kernelType;
-        private readonly bool _useGpuOrder = useGpuOrder;
+	private readonly GpuKernelType _kernelType = kernelType;
+	private readonly bool _useGpuOrder = useGpuOrder;
 
-        public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
-        {
-                var gpuLease = GpuKernelPool.GetKernel(_useGpuOrder);
-                var accelerator = gpuLease.Accelerator;
-                var stream = gpuLease.Stream;
-                int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy and avoids TDR
-                UInt128 kStart = UInt128.One;
-                ulong divMul = (ulong)((((UInt128)1 << 64) - UInt128.One) / exponent) + 1UL;
-                byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0; // ILGPU kernels do not support bool parameters
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
 
-                var kernel = _kernelType switch
-                {
-                        GpuKernelType.Pow2Mod => gpuLease.Pow2ModOrderKernel,
-                        _ => gpuLease.IncrementalOrderKernel,
-                };
+	public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
+	{
+		// GpuPrimeWorkLimiter.Acquire();
+		// var accelerator = SharedGpuContext.Accelerator;
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+		var stream = accelerator.CreateStream();
+		int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy and avoids TDR
+		UInt128 kStart = UInt128.One;
+		ulong divMul = (ulong)((((UInt128)1 << 64) - UInt128.One) / exponent) + 1UL;
+		byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0; // ILGPU kernels do not support bool parameters
 
-                var foundBuffer = accelerator.Allocate1D<int>(1); // TODO: Replace this allocation with the pooled device buffer from GpuOrderKernelBenchmarks so repeated scans reuse the pinned staging memory instead of allocating per run.
-                exponent.Mod10_8_5_3Steps(out ulong step10, out ulong step8, out ulong step5, out ulong step3);
+		KernelContainer kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, KernelType.Pow2ModOrderKernelScan | KernelType.IncrementalOrderKernelScan);
+		var kernel = _kernelType switch
+		{
+			GpuKernelType.Pow2Mod => kernels.Pow2ModOrder!,
+			_ => kernels.IncrementalOrder!,
+		};
 
-                UInt128 remaining;
-                int currentSize;
-                int found = 0;
-                while (kStart <= maxK && Volatile.Read(ref isPrime))
-                {
-                        remaining = maxK - kStart + 1UL;
-                        currentSize = remaining > (UInt128)batchSize ? batchSize : (int)remaining;
-                        // This clear is required because the result may not always be set. Remove it after the kernels are modified to always set the result.
-                        foundBuffer.MemSetToZero();
-                        // Precompute residue automaton bases for this batch
-                        UInt128 q0 = twoP * kStart + 1UL;
-                        // TODO: Replace these modulo recomputations with the divisor-cycle stepping deltas captured in
-                        // MersenneDivisorCycleLengthGpuBenchmarks so each batch advances via cached residues from the
-                        // single snapshot block instead of recalculating mod 10/8/5/3 in every iteration.
-                        q0.Mod10_8_5_3(out ulong q0m10, out ulong q0m8, out ulong q0m5, out ulong q0m3);
-                        var kernelArgs = new ResidueAutomatonArgs(q0m10, step10, q0m8, step8, q0m3, step3, q0m5, step5);
+		MemoryBuffer1D<int, ILGPU.Stride1D.Dense>? foundBuffer;
+		// lock (accelerator)
+		{
+			foundBuffer = accelerator.Allocate1D<int>(1); // TODO: Replace this allocation with the pooled device buffer from GpuOrderKernelBenchmarks so repeated scans reuse the pinned staging memory instead of allocating per run.
 
-                        // Ensure device has small cycles table for early rejections in order kernels
-                        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(accelerator);
-                        kernel(stream, currentSize, exponent, (GpuUInt128)twoP, (GpuUInt128)kStart, last, divMul,
-                                        kernelArgs, foundBuffer.View, smallCyclesView);
+		}
+		exponent.Mod10_8_5_3Steps(out ulong step10, out ulong step8, out ulong step5, out ulong step3);
 
-                        stream.Synchronize();
-                        foundBuffer.View.CopyToCPU(ref found, 1);
-                        if (found != 0)
-                        {
-                                Volatile.Write(ref isPrime, false);
-                                break;
-                        }
+		UInt128 remaining;
+		int currentSize;
+		int found = 0;
 
-                        kStart += (UInt128)currentSize;
-                }
+		var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice(kernels);
+		while (kStart <= maxK && Volatile.Read(ref isPrime))
+		{
+			remaining = maxK - kStart + 1UL;
+			currentSize = remaining > (UInt128)batchSize ? batchSize : (int)remaining;
+			// This clear is required because the result may not always be set. Remove it after the kernels are modified to always set the result.
+			foundBuffer.MemSetToZero(stream);
+			// Precompute residue automaton bases for this batch
+			UInt128 q0 = twoP * kStart + 1UL;
+			// TODO: Replace these modulo recomputations with the divisor-cycle stepping deltas captured in
+			// MersenneDivisorCycleLengthGpuBenchmarks so each batch advances via cached residues from the
+			// single snapshot block instead of recalculating mod 10/8/5/3 in every iteration.
+			q0.Mod10_8_5_3(out ulong q0m10, out ulong q0m8, out ulong q0m5, out ulong q0m3);
+			var kernelArgs = new ResidueAutomatonArgs(q0m10, step10, q0m8, step8, q0m3, step3, q0m5, step5);
 
-                foundBuffer.Dispose();
-                gpuLease.Dispose();
-        }
+			// Ensure device has small cycles table for early rejections in order kernels
+			kernel(
+				stream, currentSize, exponent, (GpuUInt128)twoP, (GpuUInt128)kStart,
+				last, divMul, kernelArgs, foundBuffer.View, smallCyclesView
+			);
+
+			foundBuffer.View.CopyToCPU(stream, ref found, 1);
+			stream.Synchronize();
+			if (found != 0)
+			{
+				Volatile.Write(ref isPrime, false);
+				break;
+			}
+
+			kStart += (UInt128)currentSize;
+		}
+
+		stream.Dispose();
+		foundBuffer.Dispose();
+		// GpuPrimeWorkLimiter.Release();
+	}
 }

PerfectNumbers.Core/Gpu/MersenneNumberResidueGpuTester.cs

@@ -1,9 +1,4 @@
-using System;
-using System.Buffers;
-using System.Collections.Concurrent;
-using System.Collections.Generic;
 using System.Runtime.InteropServices;
-using System.Threading;
 using System.Runtime.CompilerServices;
 using ILGPU;
 using ILGPU.Runtime;
@@ -12,185 +7,211 @@ namespace PerfectNumbers.Core.Gpu;
 
 public class MersenneNumberResidueGpuTester(bool useGpuOrder)
 {
-    private readonly bool _useGpuOrder = useGpuOrder;
-    private readonly ConcurrentBag<GpuContextPool.GpuContextLease> _acceleratorPool = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>> _pow2ModKernelCache = new(AcceleratorReferenceComparer.Instance);
-    private readonly ConcurrentDictionary<Accelerator, ConcurrentBag<ResidueResources>> _resourcePools = new(AcceleratorReferenceComparer.Instance);
+	private readonly bool _useGpuOrder = useGpuOrder;
 
-    // TODO: Integrate MersenneDivisorCycles.Shared to consult cycle lengths for small q (<= 4M) and fast-reject
-    // candidates without launching heavy kernels. Consider a device-visible constant buffer per-accelerator
-    // via GpuKernelPool to avoid host round-trips.
-    // TODO: When cycles are missing for larger q values, compute only the single required cycle on the device
-    // requested by the caller, skip queuing additional block generation, and keep the snapshot cache untouched.
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>? _pow2ModKernel;
 
-    // GPU residue variant: check 2^p % q == 1 for q = 2*p*k + 1.
-    public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
-    {
-        var limiter = GpuPrimeWorkLimiter.Acquire();
-        var gpuLease = RentAccelerator();
-        var accelerator = gpuLease.Accelerator;
+	[ThreadStatic]
+	private static Queue<ResidueResources>? _resourcePools;
 
-        // Monitor.Enter(gpuLease.ExecutionLock);
+	// TODO: Integrate MersenneDivisorCycles.Shared to consult cycle lengths for small q (<= 4M) and fast-reject
+	// candidates without launching heavy kernels. Consider a device-visible constant buffer per-accelerator
+	// via GpuKernelPool to avoid host round-trips.
+	// TODO: When cycles are missing for larger q values, compute only the single required cycle on the device
+	// requested by the caller, skip queuing additional block generation, and keep the snapshot cache untouched.
 
-        var stream = accelerator.CreateStream();
-        var kernel = GetPow2ModKernel(accelerator);
-        var resources = RentResources(accelerator, GpuConstants.ScanBatchSize);
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
 
-        var orderBuffer = resources.OrderBuffer;
-        ulong[] orderArray = resources.OrderArray;
-        int batchSize = resources.Capacity;
+	// GPU residue variant: check 2^p % q == 1 for q = 2*p*k + 1.
+	public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
+	{
+		// GpuPrimeWorkLimiter.Acquire();
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
 
-        // Ensure device has small cycles and primes tables for in-kernel lookup
-        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(accelerator);
-        ResiduePrimeViews primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(accelerator);
+		// Monitor.Enter(gpuLease.ExecutionLock);
 
-        UInt128 kStart = 1UL;
-        UInt128 limit = maxK + UInt128.One;
-        byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0;
-        twoP.Mod10_8_5_3(out ulong step10, out ulong step8, out ulong step5, out ulong step3);
-        step10 = step10.Mod10();
+		var stream = accelerator.CreateStream();
+		var kernel = GetPow2ModKernel(accelerator);
+		var resources = RentResources(accelerator, GpuConstants.ScanBatchSize);
+		var gpuKernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, KernelType.SmallCycles | KernelType.SmallPrimes);
 
-        GpuUInt128 twoPGpu = (GpuUInt128)twoP;
-        UInt128 batchSize128 = (UInt128)batchSize;
-        UInt128 fullBatchStep = twoP * batchSize128;
-        UInt128 q = twoP * kStart + UInt128.One;
+		var orderBuffer = resources.OrderBuffer;
+		ulong[] orderArray = resources.OrderArray;
+		int batchSize = resources.Capacity;
 
-        while (kStart < limit && Volatile.Read(ref isPrime))
-        {
-            UInt128 remaining = limit - kStart;
-            int currentSize = remaining > batchSize128 ? batchSize : (int)remaining;
-            Span<ulong> orders = orderArray.AsSpan(0, currentSize);
-            ref ulong ordersRef = ref MemoryMarshal.GetReference(orders);
+		// Ensure device has small cycles and primes tables for in-kernel lookup
+		// TODO: Pass-in the stream down the path
+		var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice(gpuKernels);
+		ResiduePrimeViews primeViews = GpuKernelPool.GetSmallPrimesOnDevice(gpuKernels);
 
-            q.Mod10_8_5_3(out ulong q0m10, out ulong q0m8, out ulong q0m5, out ulong q0m3);
-            var kernelArgs = new ResidueAutomatonArgs(q0m10, step10, q0m8, step8, q0m3, step3, q0m5, step5);
+		UInt128 kStart = 1UL;
+		UInt128 limit = maxK + UInt128.One;
+		byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0;
+		twoP.Mod10_8_5_3(out ulong step10, out ulong step8, out ulong step5, out ulong step3);
+		step10 = step10.Mod10();
 
-            kernel(stream, currentSize, exponent, twoPGpu, (GpuUInt128)kStart, last, 0UL,
-                kernelArgs, orderBuffer.View, smallCyclesView, primeViews.LastOne, primeViews.LastSeven, primeViews.LastOnePow2, primeViews.LastSevenPow2);
+		GpuUInt128 twoPGpu = (GpuUInt128)twoP;
+		UInt128 batchSize128 = (UInt128)batchSize;
+		UInt128 fullBatchStep = twoP * batchSize128;
+		UInt128 q = twoP * kStart + UInt128.One;
 
-            stream.Synchronize();
-            orderBuffer.View.CopyToCPU(ref ordersRef, currentSize);
-            if (!Volatile.Read(ref isPrime))
-            {
-                break;
-            }
+		while (kStart < limit && Volatile.Read(ref isPrime))
+		{
+			UInt128 remaining = limit - kStart;
+			int currentSize = remaining > batchSize128 ? batchSize : (int)remaining;
+			Span<ulong> orders = orderArray.AsSpan(0, currentSize);
+			ref ulong ordersRef = ref MemoryMarshal.GetReference(orders);
 
-            bool compositeFound = false;
-            for (int i = 0; i < currentSize; i++)
-            {
-                if (orders[i] != 0UL)
-                {
-                    Volatile.Write(ref isPrime, false);
-                    compositeFound = true;
-                    break;
-                }
-            }
+			q.Mod10_8_5_3(out ulong q0m10, out ulong q0m8, out ulong q0m5, out ulong q0m3);
+			var kernelArgs = new ResidueAutomatonArgs(q0m10, step10, q0m8, step8, q0m3, step3, q0m5, step5);
 
-            if (compositeFound)
-            {
-                break;
-            }
+			kernel(stream, currentSize, exponent, twoPGpu, (GpuUInt128)kStart, last, 0UL,
+				kernelArgs, orderBuffer.View, smallCyclesView, primeViews.LastOne, primeViews.LastSeven, primeViews.LastOnePow2, primeViews.LastSevenPow2);
 
-            UInt128 processed = (UInt128)currentSize;
-            kStart += processed;
-            if (kStart < limit)
-            {
-                if (processed == batchSize128)
-                {
-                    q += fullBatchStep;
-                }
-                else
-                {
-                    q += twoP * processed; // TODO: Swap this multiplication for the shared cycle stepping helper once residue cycles are mandatory so we can reuse the cached remainder ladder instead of performing UInt128 multiply operations.
-                }
-            }
-        }
-
-        ReturnResources(accelerator, resources);
-        stream.Dispose();
-        // Monitor.Exit(gpuLease.ExecutionLock);
-        ReturnAccelerator(gpuLease);
-        limiter.Dispose();
-    }
-
-    private Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> GetPow2ModKernel(Accelerator accelerator) =>
-        _pow2ModKernelCache.GetOrAdd(
-            accelerator,
-            static acc =>
-            {
-                var loaded = acc.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModKernelScan);
-                return (Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>)Delegate.CreateDelegate(
-                    typeof(Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>),
-                    loaded.Target,
-                    loaded.Method);
-            });
-
-    private GpuContextPool.GpuContextLease RentAccelerator()
-    {
-        if (_acceleratorPool.TryTake(out var lease))
-        {
-            return lease;
-        }
-
-        return GpuContextPool.Rent();
-    }
-
-    private void ReturnAccelerator(GpuContextPool.GpuContextLease lease) => _acceleratorPool.Add(lease);
-
-    private ResidueResources RentResources(Accelerator accelerator, int capacity)
-    {
-        var bag = _resourcePools.GetOrAdd(accelerator, static _ => new ConcurrentBag<ResidueResources>());
-        while (bag.TryTake(out var resources))
-        {
-            if (resources.Capacity >= capacity)
+			orderBuffer.View.CopyToCPU(stream, ref ordersRef, currentSize);
+			stream.Synchronize();
+			if (!Volatile.Read(ref isPrime))
 			{
-				// The caller always overwrites the required elements. We don't need to worry about clearing the output buffer.
-                // resources.OrderBuffer.MemSetToZero();
-                return resources;
-            }
+				break;
+			}
 
-            resources.Dispose();
-        }
+			bool compositeFound = false;
+			for (int i = 0; i < currentSize; i++)
+			{
+				if (orders[i] != 0UL)
+				{
+					Volatile.Write(ref isPrime, false);
+					compositeFound = true;
+					break;
+				}
+			}
 
-        var created = new ResidueResources(accelerator, capacity);
-		// The caller always overwrites the required elements. We don't need to worry about clearing the output buffer.
-        // created.OrderBuffer.MemSetToZero();
-        return created;
-    }
+			if (compositeFound)
+			{
+				break;
+			}
 
-    private void ReturnResources(Accelerator accelerator, ResidueResources resources) =>
-        _resourcePools.GetOrAdd(accelerator, static _ => new ConcurrentBag<ResidueResources>()).Add(resources);
+			UInt128 processed = (UInt128)currentSize;
+			kStart += processed;
+			if (kStart < limit)
+			{
+				if (processed == batchSize128)
+				{
+					q += fullBatchStep;
+				}
+				else
+				{
+					q += twoP * processed; // TODO: Swap this multiplication for the shared cycle stepping helper once residue cycles are mandatory so we can reuse the cached remainder ladder instead of performing UInt128 multiply operations.
+				}
+			}
+		}
 
-    private sealed class ResidueResources
-    {
-        internal ResidueResources(Accelerator accelerator, int capacity)
-        {
-            Capacity = Math.Max(1, capacity);
-            OrderBuffer = accelerator.Allocate1D<ulong>(Capacity);
-            OrderArray = ArrayPool<ulong>.Shared.Rent(Capacity);
-        }
+		ReturnResources(resources);
+		stream.Dispose();
+		// Monitor.Exit(gpuLease.ExecutionLock);
+		gpuKernels.Dispose();
+		// GpuPrimeWorkLimiter.Release();
+	}
 
-        internal MemoryBuffer1D<ulong, Stride1D.Dense> OrderBuffer { get; }
+	private static Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> GetPow2ModKernel(Accelerator accelerator)
+	{
 
-        internal ulong[] OrderArray { get; }
+		var pool = _pow2ModKernel ?? [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
 
-        internal int Capacity { get; }
+		cached = CreatePow2ModKernel(accelerator);
+		pool[accelerator] = cached;
+		return cached;
+	}
 
-        public void Dispose()
-        {
-            OrderBuffer.Dispose();
-            ArrayPool<ulong>.Shared.Return(OrderArray, clearArray: false);
-        }
-    }
+	private static Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> CreatePow2ModKernel(Accelerator accelerator)
+	{
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModKernelScan);
 
-    private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
-    {
-        public static AcceleratorReferenceComparer Instance { get; } = new();
+		var kernel = (Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>)Delegate.CreateDelegate(
+			typeof(Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>),
+			loaded.Target,
+			loaded.Method);
+		return kernel;
+	}
 
-        public bool Equals(Accelerator? x, Accelerator? y) => ReferenceEquals(x, y);
 
-        public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
-    }
+	private static ResidueResources RentResources(Accelerator accelerator, int capacity)
+	{
+		var resourcePools = _resourcePools ??= new();
+		if (resourcePools.TryDequeue(out var bag))
+		{
+			bag.EnsureCapacity(accelerator, capacity);
+			return bag;
+		}
+
+		return new ResidueResources(accelerator, capacity);
+	}
+
+	private static void ReturnResources(ResidueResources resources)
+	{
+		var resourcePools = _resourcePools ??= new();
+		resourcePools.Enqueue(resources);
+	}
+
+	private sealed class ResidueResources
+	{
+		internal ResidueResources(Accelerator accelerator, int capacity)
+		{
+			Capacity = Math.Max(1, capacity);
+			// lock (accelerator)
+			{
+				OrderBuffer = accelerator.Allocate1D<ulong>(Capacity);
+			}
+
+			OrderArray = ThreadStaticPools.UlongPool.Rent(Capacity);
+		}
+
+		internal void EnsureCapacity(Accelerator accelerator, int capacity)
+		{
+			if (OrderBuffer.Length >= capacity)
+			{
+				return;
+			}
+
+			OrderBuffer.Dispose();
+			ThreadStaticPools.UlongPool.Return(OrderArray);
+
+			// lock (accelerator)
+			{
+				OrderBuffer = accelerator.Allocate1D<ulong>(capacity);
+			}
+
+			OrderArray = ThreadStaticPools.UlongPool.Rent(capacity);
+			Capacity = capacity;
+		}
+
+		internal MemoryBuffer1D<ulong, Stride1D.Dense> OrderBuffer { get; private set; }
+
+		internal ulong[] OrderArray { get; private set; }
+
+		internal int Capacity { get; private set; }
+
+		public void Dispose()
+		{
+			OrderBuffer.Dispose();
+			ThreadStaticPools.UlongPool.Return(OrderArray, clearArray: false);
+		}
+	}
+
+	private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
+	{
+		public static AcceleratorReferenceComparer Instance { get; } = new();
+
+		public bool Equals(Accelerator? x, Accelerator? y) => ReferenceEquals(x, y);
+
+		public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
+	}
 
 
 }

PerfectNumbers.Core/Gpu/NttGpuMath.cs

@@ -1,4 +1,3 @@
-using System.Buffers;
 using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
@@ -27,27 +26,64 @@ public static class NttGpuMath
     // Global backend setting for GPU transforms. Used by ForwardGpu/InverseGpu.
     // TODO(NTT-OPT): Consider passing backend explicitly to avoid global state.
     public static NttBackend GpuTransformBackend { get; set; } = NttBackend.Reference;
-    // Controls modular reduction strategy inside staged NTT paths.
-    public static ModReductionMode ReductionMode { get; set; } = ModReductionMode.Auto;
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128>> MulKernelCache = new(); // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128>> StageKernelCache = new(); // TODO: Same as above – materialize staged kernels during startup so we can drop ConcurrentDictionary usage entirely per the benchmark guidance.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128>> ScaleKernelCache = new(); // TODO: Promote to the static kernel table initialized alongside the GpuModularArithmeticBenchmarks fast path.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong>> StageMontKernelCache = new(); // TODO: Inline the Montgomery kernels into the startup table instead of using a concurrent cache now that we no longer mutate state at runtime.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>> StageBarrett128KernelCache = new(); // TODO: Same plan – reuse the precomputed kernels measured fastest in MontgomeryMultiplyBenchmarks rather than looking them up dynamically.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong>> ScaleBarrett128KernelCache = new(); // TODO: Collapse into the startup kernel array once Barrett128 constants are preloaded.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>> SquareBarrett128KernelCache = new(); // TODO: Fold these kernels into the same startup table to avoid concurrent access overhead highlighted in GpuModularArithmeticBenchmarks.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>> ToMont64KernelCache = new(); // TODO: Prebind and reuse the Montgomery conversions from MontgomeryMultiplyBenchmarks instead of storing them in a concurrent cache.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong>> FromMont64KernelCache = new(); // TODO: As above – drop ConcurrentDictionary once startup prewarms every kernel.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong>> SquareMont64KernelCache = new(); // TODO: Inline into the static kernel table established during initialization.
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>> ScaleMont64KernelCache = new(); // TODO: Move to the startup kernel table so Montgomery-scaled launches skip concurrent lookups.
+	// Controls modular reduction strategy inside staged NTT paths.
+	public static ModReductionMode ReductionMode { get; set; } = ModReductionMode.Auto;
 
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128>> ForwardKernelCache = new(); // TODO: Warm these forward kernels during initialization and store them in a plain array so the LucasLehmerGpuBenchmarks launch costs disappear.
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128>>? _mulKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
 	
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128>> InverseKernelCache = new(); // TODO: Same as above – replace with startup-prepared tables per LucasLehmerGpuBenchmarks guidance.
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream,Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128>>? _stageKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
 
-    private static readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuUInt128>, int>> BitReverseKernelCache = new(); // TODO: Precompute bit-reversal kernels during boot so we can remove the concurrent lookup overhead flagged in the benchmarks.
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128>>? _scaleKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
 	
-    private sealed class SquareCacheEntry
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong>>? _stageMontKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>>? _stageBarrett128Kernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>>? _toMont64Kernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong>>? _fromMont64Kernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong>>? _scaleBarrett128Kernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static readonly Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128>>? _squareMont64Kernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>>? _scaleMont64Kernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128>>? _forwardKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128>>? _inverseKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int>>? _bitReverseKernel; // TODO: Replace this concurrent cache with the prewarmed accelerator-indexed tables from GpuModularArithmeticBenchmarks so kernel launches avoid dictionary lookups once the kernels are baked during startup.
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128> GetStageKernel(Accelerator accelerator)
+	{
+		var pool = _stageKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+		
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128>(NttButterflyKernels.StageKernel);
+		var kernel = KernelUtil.GetKernel(loaded);
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private sealed class SquareCacheEntry
     {
         public MemoryBuffer1D<GpuUInt128, Stride1D.Dense> Buffer { get; }
         public GpuUInt128 Root { get; }
@@ -72,7 +108,7 @@ public static class NttGpuMath
         public ulong BarrettMuHigh { get; }
         public ulong BarrettMuLow { get; }
 
-        public SquareCacheEntry(Accelerator accelerator, int length, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
+        public SquareCacheEntry(Accelerator accelerator, AcceleratorStream stream, int length, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
         {
             Buffer = accelerator.Allocate1D<GpuUInt128>(length);
             Root = new GpuUInt128(primitiveRoot);
@@ -99,7 +135,7 @@ public static class NttGpuMath
             // Precompute stage twiddle factors (forward and inverse) and upload to GPU.
             Bits = (int)Math.Log2(length);
             int twiddleCount = length - 1;
-            var pool = ArrayPool<GpuUInt128>.Shared;
+            var pool = ThreadStaticPools.GpuUInt128Pool;
             var forward = pool.Rent(twiddleCount);
             var inverse = pool.Rent(twiddleCount);
             StageOffsets = new int[Bits];
@@ -136,9 +172,9 @@ public static class NttGpuMath
             }
 
             Twiddles = accelerator.Allocate1D<GpuUInt128>(twiddleCount);
-            Twiddles.View.CopyFromCPU(ref forward[0], twiddleCount);
+            Twiddles.View.CopyFromCPU(stream, ref forward[0], twiddleCount);
             TwiddlesInv = accelerator.Allocate1D<GpuUInt128>(twiddleCount);
-            TwiddlesInv.View.CopyFromCPU(ref inverse[0], twiddleCount);
+            TwiddlesInv.View.CopyFromCPU(stream, ref inverse[0], twiddleCount);
             
             // Prepare Montgomery constants for 64-bit moduli and twiddles in Montgomery domain.
             if (modulus.High == 0UL && (ReductionMode == ModReductionMode.Auto || ReductionMode == ModReductionMode.Mont64))
@@ -166,9 +202,9 @@ public static class NttGpuMath
                 }
 
                 TwiddlesMont = accelerator.Allocate1D<GpuUInt128>(twiddleCount);
-                TwiddlesMont.View.CopyFromCPU(ref forwardMont[0], twiddleCount);
+                TwiddlesMont.View.CopyFromCPU(stream, ref forwardMont[0], twiddleCount);
                 TwiddlesInvMont = accelerator.Allocate1D<GpuUInt128>(twiddleCount);
-                TwiddlesInvMont.View.CopyFromCPU(ref inverseMont[0], twiddleCount);
+				TwiddlesInvMont.View.CopyFromCPU(stream, ref inverseMont[0], twiddleCount);
 
                 pool.Return(forwardMont, clearArray: true);
                 pool.Return(inverseMont, clearArray: true);
@@ -250,7 +286,7 @@ public static class NttGpuMath
 
     private readonly record struct SquareCacheKey(int Length, GpuUInt128 Modulus, GpuUInt128 PrimitiveRoot);
 
-    private static readonly ConcurrentDictionary<Accelerator, ConcurrentDictionary<SquareCacheKey, SquareCacheEntry>> SquareCache = new();
+    private static readonly ConcurrentDictionary<SquareCacheKey, SquareCacheEntry> SquareCache = [];
 
     // TODO(NTT-OPT): Add a TwiddleCache for stage-wise Cooley–Tukey NTT.
     // Design:
@@ -258,11 +294,10 @@ public static class NttGpuMath
     // - Store one GPU buffer with all stage twiddles, or per-stage slices.
     // - Expose accessors to retrieve per-stage views used by butterfly kernels.
 
-    private static SquareCacheEntry GetSquareCache(Accelerator accelerator, int length, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
+    private static SquareCacheEntry GetSquareCache(Accelerator accelerator, AcceleratorStream stream, int length, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
     {
-        var perAccel = SquareCache.GetOrAdd(accelerator, _ => new ConcurrentDictionary<SquareCacheKey, SquareCacheEntry>());
         var key = new SquareCacheKey(length, modulus, primitiveRoot);
-        return perAccel.GetOrAdd(key, _ => new SquareCacheEntry(accelerator, length, modulus, primitiveRoot));
+        return SquareCache.GetOrAdd(key, _ => new SquareCacheEntry(accelerator, stream, length, modulus, primitiveRoot));
     }
 
     private static readonly ulong[] SmallPrimes = GenerateSmallPrimes(1000);
@@ -330,105 +365,197 @@ public static class NttGpuMath
     {
         if (accelerator is null)
         {
-            foreach (var perAccel in SquareCache.Values)
+            foreach (var entry in SquareCache.Values)
             {
-                foreach (var entry in perAccel.Values)
-                {
-                    entry.Dispose();
-                }
+				entry.Dispose();
             }
 
             SquareCache.Clear();
-            MulKernelCache.Clear();
-            StageKernelCache.Clear();
-            ScaleKernelCache.Clear();
-            StageMontKernelCache.Clear();
-            StageBarrett128KernelCache.Clear();
-            ToMont64KernelCache.Clear();
-            FromMont64KernelCache.Clear();
-            SquareMont64KernelCache.Clear();
-            ScaleMont64KernelCache.Clear();
-            ForwardKernelCache.Clear();
-            InverseKernelCache.Clear();
             // TODO(NTT-OPT): Clear TwiddleCache once introduced.
             return;
         }
 
-        if (SquareCache.TryRemove(accelerator, out var cache))
-        {
-            foreach (var entry in cache.Values)
-            {
-                entry.Dispose();
-            }
-        }
+		var squareCache = SquareCache; 
+		foreach (var entry in squareCache.Values)
+		{
+			entry.Dispose();
+		}
 
-        MulKernelCache.TryRemove(accelerator, out _);
-        StageKernelCache.TryRemove(accelerator, out _);
-        ScaleKernelCache.TryRemove(accelerator, out _);
-        StageMontKernelCache.TryRemove(accelerator, out _);
-        StageBarrett128KernelCache.TryRemove(accelerator, out _);
-        ToMont64KernelCache.TryRemove(accelerator, out _);
-        FromMont64KernelCache.TryRemove(accelerator, out _);
-        SquareMont64KernelCache.TryRemove(accelerator, out _);
-        ScaleMont64KernelCache.TryRemove(accelerator, out _);
-        ForwardKernelCache.TryRemove(accelerator, out _);
-        InverseKernelCache.TryRemove(accelerator, out _);
         // TODO(NTT-OPT): Remove per-accelerator twiddle buffers once added.
     }
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, int> GetBitReverseKernel(Accelerator accelerator) =>
-        BitReverseKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int>(NttTransformKernels.BitReverseKernel));
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int> GetBitReverseKernel(Accelerator accelerator)
+	{
+		var pool = _bitReverseKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128> GetMulKernel(Accelerator accelerator) =>
-        MulKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128>(NttPointwiseKernels.MulKernel));
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int>(NttTransformKernels.BitReverseKernel);
 		
-    private static Action<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128> GetStageKernel(Accelerator accelerator) =>
-        StageKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, GpuUInt128>(NttButterflyKernels.StageKernel));
+		var kernel = KernelUtil.GetKernel(loaded);
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128> GetScaleKernel(Accelerator accelerator) =>
-        ScaleKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128>(NttPointwiseKernels.ScaleKernel));
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong> GetStageMontKernel(Accelerator accelerator) =>
-        StageMontKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong>(NttMontgomeryKernels.StageMontKernel));
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128> GetForwardKernel(Accelerator accelerator)
+	{
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128>(NttTransformKernels.ForwardKernel);
+		var kernel = KernelUtil.GetKernel(loaded);
+		return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128>>();
+	}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong> GetStageBarrett128Kernel(Accelerator accelerator) =>
-        StageBarrett128KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>(NttBarrettKernels.StageBarrett128Kernel));
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong> GetFromMont64Kernel(Accelerator accelerator)
+	{
+		var pool = _fromMont64Kernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong> GetScaleBarrett128Kernel(Accelerator accelerator) =>
-        ScaleBarrett128KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong>(NttBarrettKernels.ScaleBarrett128Kernel));
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong>(NttMontgomeryKernels.FromMont64Kernel);
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong> GetSquareBarrett128Kernel(Accelerator accelerator) =>
-        SquareBarrett128KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>(NttBarrettKernels.SquareBarrett128Kernel));
+		var kernel = KernelUtil.GetKernel(loaded);
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong> GetToMont64Kernel(Accelerator accelerator) =>
-        ToMont64KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>(NttMontgomeryKernels.ToMont64Kernel));
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong> GetFromMont64Kernel(Accelerator accelerator) =>
-        FromMont64KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong>(NttMontgomeryKernels.FromMont64Kernel));
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128> GetInverseKernel(Accelerator accelerator)
+	{
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128>(NttTransformKernels.InverseKernel);
+		var kernel = KernelUtil.GetKernel(loaded);
+		return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128>>();
+	}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong> GetSquareMont64Kernel(Accelerator accelerator) =>
-        SquareMont64KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong>(NttMontgomeryKernels.SquareMont64Kernel));
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128> GetMulKernel(Accelerator accelerator)
+	{
+		var pool = _mulKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong> GetScaleMont64Kernel(Accelerator accelerator) =>
-        ScaleMont64KernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>(NttMontgomeryKernels.ScaleMont64Kernel));
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128>(NttPointwiseKernels.MulKernel);
 		
-    private static Action<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128> GetForwardKernel(Accelerator accelerator) =>
-        ForwardKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128>(NttTransformKernels.ForwardKernel));
+		var kernel = KernelUtil.GetKernel(loaded);
 
-    // TODO(NTT-OPT): Introduce stage-wise Cooley–Tukey kernels and a
-    // per-(accelerator,length,modulus,root) twiddle cache. The plan:
-    // - Precompute twiddle factors for all stages once and cache on GPU.
-    // - Replace the current O(n^2) ForwardKernel with O(n log n) butterflies.
-    // - Provide accessors like GetStageKernel(...) and GetTwiddleBuffer(...).
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, GpuUInt128>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
 
-    private static Action<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128> GetInverseKernel(Accelerator accelerator) =>
-        InverseKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>, int, GpuUInt128, GpuUInt128, GpuUInt128>(NttTransformKernels.InverseKernel));
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128> GetScaleKernel(Accelerator accelerator)
+	{
+		var pool = _scaleKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
 
-    // TODO(NTT-OPT): Mirror the forward stage-wise design for the inverse
-    // transform. Precompute inverse twiddles and a normalization factor
-    // (n^-1 mod m). Launch one short kernel per stage to avoid TDR.
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128>(NttPointwiseKernels.ScaleKernel);
 		
-    public static void BitReverse(Span<GpuUInt128> values)
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong> GetScaleBarrett128Kernel(Accelerator accelerator)
+	{
+		var pool = _scaleBarrett128Kernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong>(NttBarrettKernels.ScaleBarrett128Kernel);
+		
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, GpuUInt128, ulong, ulong, ulong, ulong>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong> GetScaleMont64Kernel(Accelerator accelerator)
+	{
+		var pool = _scaleMont64Kernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>(NttMontgomeryKernels.ScaleMont64Kernel);
+		
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong> GetStageBarrett128Kernel(Accelerator accelerator)
+	{
+		var pool = _stageBarrett128Kernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>(NttBarrettKernels.StageBarrett128Kernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+		
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong, ulong, ulong>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong> GetStageMontKernel(Accelerator accelerator)
+	{
+		var pool = _stageMontKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong>(NttMontgomeryKernels.StageMontKernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, int, int, int, ArrayView<GpuUInt128>, ulong, ulong>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong> GetToMont64Kernel(Accelerator accelerator)
+	{
+		var pool = _toMont64Kernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>(NttMontgomeryKernels.ToMont64Kernel);
+		
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ulong, ulong, ulong>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	// TODO(NTT-OPT): Mirror the forward stage-wise design for the inverse
+	// transform. Precompute inverse twiddles and a normalization factor
+	// (n^-1 mod m). Launch one short kernel per stage to avoid TDR.
+
+	public static void BitReverse(Span<GpuUInt128> values)
     {
         int n = values.Length;
         int bits = (int)Math.Log2(n);
@@ -446,21 +573,22 @@ public static class NttGpuMath
     {
         int n = values.Length;
         int bits = (int)Math.Log2(n);
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var kernel = GetBitReverseKernel(accelerator);
-        var pool = ArrayPool<GpuUInt128>.Shared;
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var pool = ThreadStaticPools.GpuUInt128Pool;
         var array = pool.Rent(n);
         values.CopyTo(array);
         var buffer = accelerator.Allocate1D<GpuUInt128>(n);
-        buffer.View.CopyFromCPU(ref array[0], n);
-        kernel(n, buffer.View, bits);
-        accelerator.Synchronize();
-        buffer.View.CopyToCPU(ref array[0], n);
+        var stream = accelerator.CreateStream();
+		buffer.View.CopyFromCPU(stream, ref array[0], n);
+		var bitReverseKernel = GetBitReverseKernel(accelerator);
+        bitReverseKernel(stream, n, buffer.View, bits);
+        buffer.View.CopyToCPU(stream, ref array[0], n);
+        stream.Synchronize();
         array.AsSpan(0, n).CopyTo(values);
         pool.Return(array, clearArray: true);
         buffer.Dispose();
-        gpu.Dispose();
+        stream.Dispose();
 
     }
 
@@ -508,30 +636,30 @@ public static class NttGpuMath
     private static void ForwardGpuReference(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
     {
         int n = (int)values.Length;
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var kernel = GetForwardKernel(accelerator);
-        var pool = ArrayPool<GpuUInt128>.Shared;
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
+        var pool = ThreadStaticPools.GpuUInt128Pool;
         var inputArray = pool.Rent(n);
         var outputArray = pool.Rent(n);
         values.CopyTo(inputArray);
         var inputBuffer = accelerator.Allocate1D<GpuUInt128>(n);
         var outputBuffer = accelerator.Allocate1D<GpuUInt128>(n);
-        inputBuffer.View.CopyFromCPU(ref inputArray[0], n);
+        inputBuffer.View.CopyFromCPU(stream, ref inputArray[0], n);
         UInt128 modValue = modulus;
         UInt128 expBase = (modValue - 1UL) / (ulong)n;
         var root = new GpuUInt128(primitiveRoot);
-        root.ModPow((ulong)expBase, modulus);
-        kernel(n, inputBuffer.View, outputBuffer.View, n, modulus, root);
-        accelerator.Synchronize();
-        outputBuffer.View.CopyToCPU(ref outputArray[0], n);
+		root.ModPow((ulong)expBase, modulus);
+		var forwardKernel = GetForwardKernel(accelerator);
+        forwardKernel(stream, n, inputBuffer.View, outputBuffer.View, n, modulus, root);
+        outputBuffer.View.CopyToCPU(stream, ref outputArray[0], n);
+        stream.Synchronize();
         outputArray.AsSpan(0, n).CopyTo(values);
         pool.Return(inputArray, clearArray: true);
         pool.Return(outputArray, clearArray: true);
         outputBuffer.Dispose();
         inputBuffer.Dispose();
-        gpu.Dispose();
-
+        stream.Dispose();
     }
 
     public static void Inverse(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
@@ -605,16 +733,16 @@ public static class NttGpuMath
     private static void InverseGpuReference(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
     {
         int n = values.Length;
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var kernel = GetInverseKernel(accelerator);
-        var pool = ArrayPool<GpuUInt128>.Shared;
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
+        var pool = ThreadStaticPools.GpuUInt128Pool;
         var inputArray = pool.Rent(n);
         var outputArray = pool.Rent(n);
         values.CopyTo(inputArray);
         var inputBuffer = accelerator.Allocate1D<GpuUInt128>(n);
         var outputBuffer = accelerator.Allocate1D<GpuUInt128>(n);
-        inputBuffer.View.CopyFromCPU(ref inputArray[0], n);
+        inputBuffer.View.CopyFromCPU(stream, ref inputArray[0], n);
         UInt128 modValue = modulus;
         UInt128 expBase = (modValue - 1UL) / (ulong)n;
         var root = new GpuUInt128(primitiveRoot);
@@ -630,78 +758,78 @@ public static class NttGpuMath
         }
 
         var nInv = new GpuUInt128((UInt128)n);
-        if (modulus.High == 0UL)
-        {
-            nInv.ModInv(modulus.Low);
-        }
-        else
-        {
-            nInv.ModInv(modulus);
-        }
-        kernel(n, inputBuffer.View, outputBuffer.View, n, modulus, rootInv, nInv);
-        accelerator.Synchronize();
-        outputBuffer.View.CopyToCPU(ref outputArray[0], n);
+		if (modulus.High == 0UL)
+		{
+			nInv.ModInv(modulus.Low);
+		}
+		else
+		{
+			nInv.ModInv(modulus);
+		}
+
+		var inverseKernel = GetInverseKernel(accelerator);
+        inverseKernel(stream, n, inputBuffer.View, outputBuffer.View, n, modulus, rootInv, nInv);
+        outputBuffer.View.CopyToCPU(stream, ref outputArray[0], n);
+        stream.Synchronize();
         outputArray.AsSpan(0, n).CopyTo(values);
         pool.Return(inputArray, clearArray: true);
         pool.Return(outputArray, clearArray: true);
         outputBuffer.Dispose();
         inputBuffer.Dispose();
-        gpu.Dispose();
-
+        stream.Dispose();
     }
 
     private static void ForwardGpuStaged(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
     {
         int n = values.Length;
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
         var buffer = accelerator.Allocate1D<GpuUInt128>(n);
-        buffer.View.CopyFromCPU(ref values[0], n);
-        var cache = GetSquareCache(accelerator, n, modulus, primitiveRoot);
-        ForwardDevice(accelerator, buffer.View, n, modulus, cache);
-        buffer.View.CopyToCPU(ref values[0], n);
+        buffer.View.CopyFromCPU(stream, ref values[0], n);
+        var cache = GetSquareCache(accelerator, stream, n, modulus, primitiveRoot);
+        ForwardDevice(accelerator, stream, buffer.View, n, modulus, cache);
+		buffer.View.CopyToCPU(stream, ref values[0], n);
+		stream.Synchronize();
+        stream.Dispose();
         buffer.Dispose();
-        gpu.Dispose();
-
     }
 
     private static void InverseGpuStaged(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
     {
         int n = values.Length;
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
         var buffer = accelerator.Allocate1D<GpuUInt128>(n);
-        buffer.View.CopyFromCPU(ref values[0], n);
-        var cache = GetSquareCache(accelerator, n, modulus, primitiveRoot);
-        InverseDevice(accelerator, buffer.View, n, modulus, cache);
-        buffer.View.CopyToCPU(ref values[0], n);
+        buffer.View.CopyFromCPU(stream, ref values[0], n);
+        var cache = GetSquareCache(accelerator, stream, n, modulus, primitiveRoot);
+        InverseDevice(accelerator, stream, buffer.View, n, modulus, cache);
+        buffer.View.CopyToCPU(stream, ref values[0], n);
+		stream.Synchronize();
+        stream.Dispose();
         buffer.Dispose();
-        gpu.Dispose();
-
     }
 
-    private static void ForwardDevice(Accelerator accelerator, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
-    {
-        var bitKernel = GetBitReverseKernel(accelerator);
-        bitKernel(length, data, cache.Bits);
-        accelerator.Synchronize();
+    private static void ForwardDevice(Accelerator accelerator, AcceleratorStream stream, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
+	{
+		var bitReverseKernel = GetBitReverseKernel(accelerator);
+        bitReverseKernel(stream, length, data, cache.Bits);
 
         if (cache.UseMontgomery64)
-        {
-            var toMont = GetToMont64Kernel(accelerator);
-            toMont(length, data, cache.ModulusLow, cache.MontNPrime64, cache.MontR2Mod64);
-            accelerator.Synchronize();
+		{
+			var toMont64Kernel = GetToMont64Kernel(accelerator);
+            toMont64Kernel(stream, length, data, cache.ModulusLow, cache.MontNPrime64, cache.MontR2Mod64);
             var stageKernel = GetStageMontKernel(accelerator);
-            int butterflies = length >> 1;
-            for (int s = 0; s < cache.Bits; s++)
-            {
-                int len = 1 << (s + 1);
-                int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesMont!.View, cache.ModulusLow, cache.MontNPrime64);
-                accelerator.Synchronize();
-            }
-            var fromMont = GetFromMont64Kernel(accelerator);
-            fromMont(length, data, cache.ModulusLow, cache.MontNPrime64);
+			int butterflies = length >> 1;
+			for (int s = 0; s < cache.Bits; s++)
+			{
+				int len = 1 << (s + 1);
+				int half = len >> 1;
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesMont!.View, cache.ModulusLow, cache.MontNPrime64);
+			}
+			var fromMont64Kernel = GetFromMont64Kernel(accelerator);
+            fromMont64Kernel(stream, length, data, cache.ModulusLow, cache.MontNPrime64);
         }
         else if (cache.UseBarrett128)
         {
@@ -711,8 +839,7 @@ public static class NttGpuMath
             {
                 int len = 1 << (s + 1);
                 int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.Twiddles.View, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
-                accelerator.Synchronize();
+                stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.Twiddles.View, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
             }
         }
         else
@@ -723,72 +850,67 @@ public static class NttGpuMath
             {
                 int len = 1 << (s + 1);
                 int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.Twiddles.View, modulus);
-                accelerator.Synchronize();
+                stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.Twiddles.View, modulus);
             }
         }
 
-        accelerator.Synchronize();
+        stream.Synchronize();
     }
 
-    private static void InverseDevice(Accelerator accelerator, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
-    {
-        var bitKernel = GetBitReverseKernel(accelerator);
-        bitKernel(length, data, cache.Bits);
-        accelerator.Synchronize();
+    private static void InverseDevice(Accelerator accelerator, AcceleratorStream stream, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
+	{
+		var bitReverseKernel = GetBitReverseKernel(accelerator);
+        bitReverseKernel(stream, length, data, cache.Bits);
 
         if (cache.UseMontgomery64)
-        {
-            var toMont = GetToMont64Kernel(accelerator);
-            toMont(length, data, cache.ModulusLow, cache.MontNPrime64, cache.MontR2Mod64);
-            accelerator.Synchronize();
+		{
+			var toMont64Kernel = GetToMont64Kernel(accelerator);
+            toMont64Kernel(stream, length, data, cache.ModulusLow, cache.MontNPrime64, cache.MontR2Mod64);
             var stageKernel = GetStageMontKernel(accelerator);
             int butterflies = length >> 1;
-            for (int s = 0; s < cache.Bits; s++)
-            {
-                int len = 1 << (s + 1);
-                int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInvMont!.View, cache.ModulusLow, cache.MontNPrime64);
-                accelerator.Synchronize();
-            }
-            var scaleKernel = GetScaleMont64Kernel(accelerator);
-            scaleKernel(length, data, cache.ModulusLow, cache.MontNPrime64, cache.MontNInvR64);
-            accelerator.Synchronize();
-            var fromMont = GetFromMont64Kernel(accelerator);
-            fromMont(length, data, cache.ModulusLow, cache.MontNPrime64);
+			for (int s = 0; s < cache.Bits; s++)
+			{
+				int len = 1 << (s + 1);
+				int half = len >> 1;
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInvMont!.View, cache.ModulusLow, cache.MontNPrime64);
+			}
+
+			var scaleMont64Kernel = GetScaleMont64Kernel(accelerator);
+			scaleMont64Kernel(stream, length, data, cache.ModulusLow, cache.MontNPrime64, cache.MontNInvR64);
+
+			var fromMont64Kernel = GetFromMont64Kernel(accelerator);
+            fromMont64Kernel(stream, length, data, cache.ModulusLow, cache.MontNPrime64);
         }
         else if (cache.UseBarrett128)
         {
             var stageKernel = GetStageBarrett128Kernel(accelerator);
             int butterflies = length >> 1;
-            for (int s = 0; s < cache.Bits; s++)
-            {
-                int len = 1 << (s + 1);
-                int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
-                accelerator.Synchronize();
-            }
-            var scaleKernel = GetScaleBarrett128Kernel(accelerator);
-            scaleKernel(length, data, cache.NInv, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
-            accelerator.Synchronize();
+			for (int s = 0; s < cache.Bits; s++)
+			{
+				int len = 1 << (s + 1);
+				int half = len >> 1;
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
+			}
+
+			var scaleBarrett128Kernel = GetScaleBarrett128Kernel(accelerator);
+            scaleBarrett128Kernel(stream, length, data, cache.NInv, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
         }
         else
         {
             var stageKernel = GetStageKernel(accelerator);
             int butterflies = length >> 1;
-            for (int s = 0; s < cache.Bits; s++)
-            {
-                int len = 1 << (s + 1);
-                int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, modulus);
-                accelerator.Synchronize();
-            }
-            var scaleKernel = GetScaleKernel(accelerator);
-            scaleKernel(length, data, cache.NInv, modulus);
-            accelerator.Synchronize();
+			for (int s = 0; s < cache.Bits; s++)
+			{
+				int len = 1 << (s + 1);
+				int half = len >> 1;
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, modulus);
+			}
+
+			var scaleKernel = GetScaleKernel(accelerator);
+            scaleKernel(stream, length, data, cache.NInv, modulus);
         }
 
-        accelerator.Synchronize();
+        stream.Synchronize();
     }
 
     public static void Convolve(Span<GpuUInt128> left, Span<GpuUInt128> right, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
@@ -800,7 +922,7 @@ public static class NttGpuMath
             n <<= 1;
         }
 
-        var pool = ArrayPool<GpuUInt128>.Shared;
+        var pool = ThreadStaticPools.GpuUInt128Pool;
         var aBuffer = pool.Rent(n);
         var bBuffer = pool.Rent(n);
         var a = aBuffer.AsSpan(0, n);
@@ -837,7 +959,7 @@ public static class NttGpuMath
             n <<= 1;
         }
 
-        var pool = ArrayPool<GpuUInt128>.Shared;
+        var pool = ThreadStaticPools.GpuUInt128Pool;
         var aBuffer = pool.Rent(n);
         var bBuffer = pool.Rent(n);
         var a = aBuffer.AsSpan(0, n);
@@ -871,41 +993,43 @@ public static class NttGpuMath
     public static void SquareGpu(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot) =>
         ConvolveGpu(values, values, modulus, primitiveRoot);
 
-    public static void SquareDevice(Accelerator accelerator, ArrayView<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
+    public static void SquareDevice(Accelerator accelerator, AcceleratorStream stream, ArrayView<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
     {
-        var cache = GetSquareCache(accelerator, (int)values.Length, modulus, primitiveRoot);
-        ForwardDevice(accelerator, values, (int)values.Length, modulus, cache);
-        var mul = GetMulKernel(accelerator);
-        mul((int)values.Length, values, values, modulus);
-        InverseDevice(accelerator, values, (int)values.Length, modulus, cache);
+        var cache = GetSquareCache(accelerator, stream, (int)values.Length, modulus, primitiveRoot);
+		ForwardDevice(accelerator, stream, values, (int)values.Length, modulus, cache);
+		var mulKernel = GetMulKernel(accelerator);
+        mulKernel(stream, (int)values.Length, values, values, modulus);
+        InverseDevice(accelerator, stream, values, (int)values.Length, modulus, cache);
     }
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
     public static void PointwiseMultiply(Span<GpuUInt128> left, Span<GpuUInt128> right, GpuUInt128 modulus)
     {
-        var gpu = GpuContextPool.Rent();
-        var accelerator = gpu.Accelerator;
-        var kernel = GetMulKernel(accelerator);
-        var pool = ArrayPool<GpuUInt128>.Shared;
+        var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+        var stream = accelerator.CreateStream();
+        var pool = ThreadStaticPools.GpuUInt128Pool;
         var leftArray = pool.Rent(left.Length);
         var rightArray = pool.Rent(right.Length);
         left.CopyTo(leftArray);
         right.CopyTo(rightArray);
         var leftBuffer = accelerator.Allocate1D<GpuUInt128>(left.Length);
         var rightBuffer = accelerator.Allocate1D<GpuUInt128>(right.Length);
-        leftBuffer.View.CopyFromCPU(ref leftArray[0], left.Length);
-        rightBuffer.View.CopyFromCPU(ref rightArray[0], right.Length);
-        kernel(left.Length, leftBuffer.View, rightBuffer.View, modulus);
-        accelerator.Synchronize();
-        leftBuffer.View.CopyToCPU(ref leftArray[0], left.Length);
-        rightBuffer.View.CopyToCPU(ref rightArray[0], right.Length);
+        leftBuffer.View.CopyFromCPU(stream, ref leftArray[0], left.Length);
+		rightBuffer.View.CopyFromCPU(stream, ref rightArray[0], right.Length);
+		var mulKernel = GetMulKernel(accelerator);
+        mulKernel(stream, left.Length, leftBuffer.View, rightBuffer.View, modulus);
+        leftBuffer.View.CopyToCPU(stream, ref leftArray[0], left.Length);
+        rightBuffer.View.CopyToCPU(stream, ref rightArray[0], right.Length);
+        stream.Synchronize();
         leftArray.AsSpan(0, left.Length).CopyTo(left);
         rightArray.AsSpan(0, right.Length).CopyTo(right);
         pool.Return(leftArray, clearArray: true);
         pool.Return(rightArray, clearArray: true);
         rightBuffer.Dispose();
         leftBuffer.Dispose();
-        gpu.Dispose();
-
+        stream.Dispose();
     }
 
     private static ulong[] GenerateSmallPrimes(int limit)

PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs

@@ -0,0 +1,97 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+using PerfectNumbers.Core.Gpu.Accelerators;
+
+namespace PerfectNumbers.Core.Gpu;
+
+public static class Pow2MontgomeryGpuCalculator
+{
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static ulong CalculateKeep(ulong exponent, in MontgomeryDivisorData divisor)
+	{
+		// GpuPrimeWorkLimiter.Acquire();
+
+		ulong result = 0UL;
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+		int acceleratorIndex = gpu.AcceleratorIndex;
+		Accelerator? accelerator = gpu.Accelerator;
+		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
+		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
+		var kernel = gpu.KeepMontgomeryKernel;
+
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+
+		exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
+		// We don't need to worry about any left-overs here.
+		// resultBuffer.MemSetToZero(stream);
+
+		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
+
+		kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
+
+		resultBuffer.View.CopyToCPU(stream, ref result, 1);
+		stream.Synchronize();
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
+		// GpuPrimeWorkLimiter.Release();
+
+		PrimeOrderCalculatorAccelerator.Return(gpu);
+		return result;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static ulong CalculateConvert(ulong exponent, in MontgomeryDivisorData divisor)
+	{
+		// GpuPrimeWorkLimiter.Acquire();
+
+		ulong result = 0UL;
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
+		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
+
+		int acceleratorIndex = gpu.AcceleratorIndex;
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
+		// We don't need to worry about any left-overs here.
+		// resultBuffer.MemSetToZero(stream);
+
+		var kernelLauncher = gpu.ConvertToStandardKernelLauncher!;
+
+		kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);			
+
+		resultBuffer.View.CopyToCPU(stream, ref result, 1);
+		stream.Synchronize();
+
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
+		// GpuPrimeWorkLimiter.Release();
+		PrimeOrderCalculatorAccelerator.Return(gpu);
+
+		return result;
+	}
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static ulong CalculateConvert(PrimeOrderCalculatorAccelerator gpu, ulong exponent, in MontgomeryDivisorData divisor)
+	{
+		ulong result = 0UL;
+		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
+		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
+
+		int acceleratorIndex = gpu.AcceleratorIndex;
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
+		// We don't need to worry about any left-overs here.
+		// resultBuffer.MemSetToZero(stream);
+
+		
+		var accelerator = gpu.Accelerator;
+		var kernelLauncher = gpu.ConvertToStandardKernelLauncher;
+		kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);			
+
+		resultBuffer.View.CopyToCPU(stream, ref result, 1);
+		stream.Synchronize();
+		
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
+
+		return result;
+	}
+}

PerfectNumbers.Core/Gpu/ResidueAutomatonArgs.cs

@@ -0,0 +1,15 @@
+namespace PerfectNumbers.Core.Gpu;
+
+public readonly struct ResidueAutomatonArgs(
+	ulong q0m10, ulong step10, ulong q0m8, ulong step8, ulong q0m3, ulong step3, ulong q0m5, ulong step5
+)
+{
+	public readonly ulong Q0M10 = q0m10;
+	public readonly ulong Step10 = step10;
+	public readonly ulong Q0M8 = q0m8;
+	public readonly ulong Step8 = step8;
+	public readonly ulong Q0M3 = q0m3;
+	public readonly ulong Step3 = step3;
+	public readonly ulong Q0M5 = q0m5;
+	public readonly ulong Step5 = step5;
+}

PerfectNumbers.Core/Gpu/ResiduePrimeViews.cs

@@ -0,0 +1,16 @@
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu;
+
+public readonly struct ResiduePrimeViews(
+	ArrayView1D<uint, Stride1D.Dense> lastOne,
+	ArrayView1D<uint, Stride1D.Dense> lastSeven,
+	ArrayView1D<ulong, Stride1D.Dense> lastOnePow2,
+	ArrayView1D<ulong, Stride1D.Dense> lastSevenPow2)
+{
+	public readonly ArrayView1D<uint, Stride1D.Dense> LastOne = lastOne;
+	public readonly ArrayView1D<uint, Stride1D.Dense> LastSeven = lastSeven;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> LastOnePow2 = lastOnePow2;
+	public readonly ArrayView1D<ulong, Stride1D.Dense> LastSevenPow2 = lastSevenPow2;
+}

PerfectNumbers.Core/Gpu/SharedGpuContext.cs

@@ -0,0 +1,29 @@
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu;
+
+public static class SharedGpuContext
+{
+	public static readonly Context Context = Context.CreateDefault();
+	public static readonly Device Device = Context.GetPreferredDevice(false);
+
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static Accelerator CreateAccelerator() => Device.CreateAccelerator(Context);
+}
+
+public readonly struct GpuContext
+{
+	public readonly Context Context;
+	public readonly Device Device;
+
+	public GpuContext()
+	{
+		var context = Context.CreateDefault();
+		Context = context;
+		Device = context.GetPreferredDevice(false);
+	}
+
+	public Accelerator CreateAccelerator() => Device.CreateAccelerator(Context);
+}