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 KernelType _kernelType = KernelType.Pow2ModKernelScan | KernelType.OrderKernelScan | KernelType.IncrementalKernelScan | KernelType.IncrementalOrderKernelScan | KernelType.Pow2ModOrderKernelScan;
 	private static void CompileGpuKernelPoolKernels(Accelerator accelerator)
 	{
-        _ = accelerator;
-        var lease = GpuKernelPool.GetKernel(useGpuOrder: false);
+
+		GpuPrimeWorkLimiter.Acquire();
+		int acceleratorIndex = AcceleratorPool.Shared.Rent();
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
 		try
 		{
-            _ = lease.OrderKernel;
-            _ = lease.IncrementalKernel;
-            _ = lease.Pow2ModKernel;
-            _ = lease.IncrementalOrderKernel;
-            _ = lease.Pow2ModOrderKernel;
+			var kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, _kernelType);
+			_ = kernels.Order;
+			_ = kernels.Incremental;
+			_ = kernels.Pow2Mod;
+			_ = kernels.IncrementalOrder;
+			_ = kernels.Pow2ModOrder;
 		}
 		finally
 		{
-            lease.Dispose();
+			stream.Synchronize();
+			AcceleratorStreamPool.Return(acceleratorIndex, stream);
+			// GpuPrimeWorkLimiter.Release();
 		}
 
-        var orderLease = GpuKernelPool.GetKernel(useGpuOrder: true);
+		stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		GpuPrimeWorkLimiter.Acquire();
 		try
 		{
-            _ = orderLease.OrderKernel;
-            _ = orderLease.IncrementalKernel;
-            _ = orderLease.Pow2ModKernel;
-            _ = orderLease.IncrementalOrderKernel;
-            _ = orderLease.Pow2ModOrderKernel;
+			var kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, _kernelType);
+			_ = kernels.Order;
+			_ = kernels.Incremental;
+			_ = kernels.Pow2Mod;
+			_ = kernels.IncrementalOrder;
+			_ = kernels.Pow2ModOrder;
 		}
 		finally
 		{
-            orderLease.Dispose();
+			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

@@ -150,10 +150,8 @@ public class NttGpuMathTests
 		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();
@@ -168,8 +166,11 @@ public class NttGpuMathTests
 		int count = 7; // length 8 -> 7 twiddles
 		var forward = new GpuUInt128[count];
 		var inverse = new GpuUInt128[count];
-            twiddlesMont.CopyToCPU(forward);
-            twiddlesInvMont.CopyToCPU(inverse);
+		twiddlesMont.CopyToCPU(stream, forward);
+		twiddlesInvMont.CopyToCPU(stream, inverse);
+		stream.Synchronize();
+		stream.Dispose();
+		accelerator.Dispose();
 
 		for (int i = 0; i < count; i++)
 		{
@@ -177,11 +178,6 @@ public class NttGpuMathTests
 			prod.Should().Be(r2);
 		}
 	}
-        finally
-        {
-            gpu.Dispose();
-        }
-    }
 
 	[Fact]
 	[Trait("Category", "Fast")]
@@ -345,24 +341,15 @@ public class NttGpuMathTests
 		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
-            {
+		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();
-            }
-        }
-        finally
-        {
-            gpu.Dispose();
-        }
+		accelerator.Dispose();
 
 		// Verify that the square cache now holds entries for the accelerator.
 		var field = typeof(NttGpuMath).GetField("SquareCache", BindingFlags.NonPublic | BindingFlags.Static)!;
@@ -370,8 +357,6 @@ public class NttGpuMathTests
 		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);
 	}

PerfectNumbers.Core.Tests/Gpu/NttMontgomeryTests.cs

@@ -33,24 +33,21 @@ public class NttMontgomeryTests
 		NttGpuMath.Square(cpu, modulus, primitiveRoot);
 
 		// Device execution
-        GpuContextPool.GpuContextLease context = GpuContextPool.Rent();
-        try
-        {
-            var accelerator = context.Accelerator;
+		var accelerator = SharedGpuContext.CreateAccelerator();
+		var stream = accelerator.CreateStream();
 		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);
+		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();
-        }
+
 	}
 
 	[Fact]

PerfectNumbers.Core.Tests/Gpu/PrimeTesterGpuTests.cs

@@ -1,5 +1,4 @@
 using FluentAssertions;
-using PerfectNumbers.Core.Gpu;
 using Xunit;
 
 namespace PerfectNumbers.Core.Tests.Gpu;
@@ -9,37 +8,23 @@ 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();
+			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();
+			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

@@ -1,7 +1,6 @@
-using System.Collections.Generic;
 using System.Runtime.CompilerServices;
-using PerfectNumbers.Core;
 using PerfectNumbers.Core.Gpu;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core.Cpu;
 
@@ -99,8 +98,6 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 		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)
@@ -131,7 +128,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 
 		// 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;
+		_ = DivisorCycleCache.Shared;
 
 		ulong stepHigh = step.High;
 		ulong stepLow = step.Low;
@@ -201,6 +198,8 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 		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;
@@ -217,6 +216,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 				// else
 				{
 					if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
+							gpu,
 							candidate,
 							prime,
 							divisorData,
@@ -241,6 +241,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 					// 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;
 				}
 
@@ -257,6 +258,8 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 			remainder7 = AddMod7(remainder7, step7);
 			remainder11 = AddMod11(remainder11, step11);
 		}
+
+		PrimeOrderCalculatorAccelerator.Return(gpu);
 		processedAll = true;
 		return false;
 	}
@@ -282,6 +285,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 		processedAll = false;
 
 		bool canAdvance = step <= limit;
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
 		if (!canAdvance)
 		{
 			if (divisor <= limit)
@@ -291,6 +295,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 					MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
 					ulong divisorCycle;
 					if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
+							gpu,
 							divisor,
 							prime,
 							divisorData,
@@ -310,6 +315,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 					if (divisorCycle == prime)
 					{
 						processedAll = true;
+						PrimeOrderCalculatorAccelerator.Return(gpu);
 						return true;
 					}
 
@@ -340,6 +346,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 				MontgomeryDivisorData divisorData = MontgomeryDivisorData.FromModulus(divisor);
 				ulong divisorCycle;
 				if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(
+						gpu,
 						divisor,
 						prime,
 						divisorData,
@@ -359,6 +366,7 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 				if (divisorCycle == prime)
 				{
 					processedAll = true;
+					PrimeOrderCalculatorAccelerator.Return(gpu);
 					return true;
 				}
 

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -1,11 +1,8 @@
-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;
 
@@ -14,8 +11,9 @@ public sealed class DivisorCycleCache
 	private const byte ByteZero = 0;
 	private const byte ByteOne = 1;
 	private const int StackBufferThreshold = 128;
-
-    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 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;
@@ -36,9 +34,16 @@ public sealed class DivisorCycleCache
 
 	public int PreferredBatchSize => _divisorCyclesBatchSize;
 
+	// private static readonly Accelerator[] _accelerators;
+
 	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();
 	}
 
@@ -79,7 +84,6 @@ public sealed class DivisorCycleCache
 				throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
 			}
 
-            // ObserveCycleCacheDivisor(divisor);
 			return cached;
 		}
 
@@ -94,7 +98,7 @@ public sealed class DivisorCycleCache
 		int[]? rentedMissing = null;
 		Span<int> missingBuffer = length <= StackBufferThreshold
 			? stackalloc int[length]
-            : new Span<int>(rentedMissing = ArrayPool<int>.Shared.Rent(length), 0, length);
+			: new Span<int>(rentedMissing = ThreadStaticPools.IntPool.Rent(length), 0, length);
 
 		int missingCount = 0;
 
@@ -114,7 +118,6 @@ public sealed class DivisorCycleCache
 					throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
 				}
 
-                // ObserveCycleCacheDivisor(divisor);
 				cycles[i] = cached;
 			}
 			else
@@ -138,7 +141,7 @@ public sealed class DivisorCycleCache
 
 		if (rentedMissing is not null)
 		{
-            ArrayPool<int>.Shared.Return(rentedMissing, clearArray: false);
+			ThreadStaticPools.IntPool.Return(rentedMissing, clearArray: false);
 		}
 	}
 
@@ -197,12 +200,12 @@ public sealed class DivisorCycleCache
 		ulong[]? rentedDivisors = null;
 		Span<ulong> workDivisors = count <= StackBufferThreshold
 			? stackalloc ulong[count]
-            : new Span<ulong>(rentedDivisors = ArrayPool<ulong>.Shared.Rent(count), 0, 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);
+			: new Span<ulong>(rentedResults = ThreadStaticPools.UlongPool.Rent(count), 0, count);
 
 		for (int i = 0; i < count; i++)
 		{
@@ -218,47 +221,80 @@ public sealed class DivisorCycleCache
 
 		if (rentedDivisors is not null)
 		{
-            ArrayPool<ulong>.Shared.Return(rentedDivisors, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedResults!, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedDivisors, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedResults!, clearArray: false);
+		}
+	}
+
+	private const int DefaultCapacity = 128;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? divisorBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? powBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? orderBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? resultBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<byte, Stride1D.Dense>? statusBuffer;
+
+	private void EnsureCapacity(int requiredCapacity)
+	{
+		if (requiredCapacity <= (divisorBuffer?.Length ?? 0))
+		{
+			return;
+		}
+
+		if (divisorBuffer != null)
+		{
+			divisorBuffer.Dispose();
+			powBuffer!.Dispose();
+			orderBuffer!.Dispose();
+			resultBuffer!.Dispose();
+			statusBuffer!.Dispose();
+		}
+
+		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);
 		}
 	}
 
 	private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
 	{
 		int length = divisors.Length;
-        var gpuLease = GpuKernelPool.GetKernel(useGpuOrder: true);
 
-        Accelerator accelerator = gpuLease.Accelerator;
-        var kernel = _gpuKernelCache.GetOrAdd(accelerator, LoadKernel);
-
-        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);
+		EnsureCapacity(length);
 
 		ulong[]? rentedPow = null;
 		Span<ulong> powSpan = length <= StackBufferThreshold
 			? stackalloc ulong[length]
-            : new Span<ulong>(rentedPow = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+			: new Span<ulong>(rentedPow = ThreadStaticPools.UlongPool.Rent(length), 0, length);
 
 		ulong[]? rentedOrder = null;
 		Span<ulong> orderSpan = length <= StackBufferThreshold
 			? stackalloc ulong[length]
-            : new Span<ulong>(rentedOrder = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+			: new Span<ulong>(rentedOrder = ThreadStaticPools.UlongPool.Rent(length), 0, length);
 
 		ulong[]? rentedResult = null;
 		Span<ulong> resultSpan = length <= StackBufferThreshold
 			? stackalloc ulong[length]
-            : new Span<ulong>(rentedResult = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+			: new Span<ulong>(rentedResult = ThreadStaticPools.UlongPool.Rent(length), 0, length);
 
 		byte[]? rentedStatus = null;
 		Span<byte> statusSpan = length <= StackBufferThreshold
 			? stackalloc byte[length]
-            : new Span<byte>(rentedStatus = ArrayPool<byte>.Shared.Rent(length), 0, length);
-
-        ref ulong divisorRef = ref MemoryMarshal.GetReference(divisors);
-        divisorBuffer.View.CopyFromCPU(ref divisorRef, length);
+			: new Span<byte>(rentedStatus = ThreadStaticPools.BytePool.Rent(length), 0, length);
 
 		for (int i = 0; i < length; i++)
 		{
@@ -275,18 +311,25 @@ public sealed class DivisorCycleCache
 			statusSpan[i] = ByteZero;
 		}
 
-        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);
-
+		// GpuPrimeWorkLimiter.Acquire();
 		int pending;
+		var kernel = _gpuKernel;
+		var acceleratorIndex = _acceleratorIndex;
+		var stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+
+		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(length, _divisorCyclesBatchSize, divisorBuffer.View, powBuffer.View, orderBuffer.View, resultBuffer.View, statusBuffer.View);
-            accelerator.Synchronize();
+			kernel(stream, length, _divisorCyclesBatchSize, divisorBuffer.View, powBuffer.View, orderBuffer.View, resultBuffer.View, statusBuffer.View);
+
+			statusBuffer.View.CopyToCPU(stream, statusSpan);
+			stream.Synchronize();
 
-            statusBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(statusSpan), length);
 			pending = 0;
 			for (int i = 0; i < length; i++)
 			{
@@ -298,22 +341,27 @@ public sealed class DivisorCycleCache
 		}
 		while (pending > 0);
 
-        resultBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(resultSpan), length);
+		resultBuffer.View.CopyToCPU(stream, resultSpan);
+		stream.Synchronize();
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
+
 		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);
+			ThreadStaticPools.UlongPool.Return(rentedPow, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedOrder!, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(rentedResult!, clearArray: false);
+			ThreadStaticPools.BytePool.Return(rentedStatus!, clearArray: false);
 		}
 
-        gpuLease.Dispose();
+		// GpuPrimeWorkLimiter.Release();
 	}
 
-    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)
+	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)
 	{
-        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);
+		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>

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;
 
@@ -25,10 +24,7 @@ internal struct ExponentRemainderStepperGpu
     }
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public bool MatchesDivisor(in GpuDivisorPartialData divisor)
-    {
-        return _modulus == divisor.Modulus;
-    }
+	public readonly bool MatchesDivisor(in GpuDivisorPartialData divisor) => _modulus == divisor.Modulus;
 
 	/// <summary>
 	/// Clears the cached residue so the stepper can be reused for another divisor.

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

@@ -1,188 +1,188 @@
-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)
+	public static T InitOnce<T>(ref T? target, Func<T> valueFactory) where T : class
 	{
-        return KernelCache.GetOrAdd(accelerator, _ => new KernelContainer());
+		if (target is { } value)
+		{
+			return value;
 		}
 
+		var newValue = valueFactory();
+		return Interlocked.CompareExchange(ref target, newValue, null) ?? newValue;
+	}
+
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
+
+	internal static KernelContainer GetOrAddKernels(int acceleratorIndex, AcceleratorStream stream, KernelType kernelType)
+	{
+		var pool = _pool ??= [];
+		if (pool[acceleratorIndex] is not { } kernels)
+		{
+			kernels = new();
+		}
+
+		Accelerator accelerator = _accelerators[acceleratorIndex];
+		if (kernelType.HasFlag(KernelType.OrderKernelScan))
+		{
+			InitOrderKernelScan(accelerator, kernels);
+		}
+
+		if (kernelType.HasFlag(KernelType.Pow2ModKernelScan))
+		{
+			InitPow2ModKernelScan(accelerator, kernels);
+		}
+
+		if (kernelType.HasFlag(KernelType.IncrementalKernelScan))
+		{
+			InitIncrementalKernelScan(accelerator, kernels);
+		}
+
+		if (kernelType.HasFlag(KernelType.IncrementalOrderKernelScan))
+		{
+			IncrementalOrderKernelScan(accelerator, kernels);
+		}
+
+		if (kernelType.HasFlag(KernelType.Pow2ModOrderKernelScan))
+		{
+			InitPow2ModOrderKernelScan(accelerator, kernels);
+		}
+
+		PreloadStaticTables(accelerator, kernels, stream, kernelType);
+
+		pool[acceleratorIndex] = kernels;
+		return kernels;
+
+		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 kernel = KernelUtil.GetKernel(loaded);
+
+				return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>>();
+			});
+		}
+
+		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);
+
+				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>>>();
+			});
+		}
+
+		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);
+
+				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>>>();
+			});
+		}
+
+		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);
+
+				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).
-    public static ArrayView1D<ulong, Stride1D.Dense> EnsureSmallCyclesOnDevice(Accelerator accelerator)
+	private static void EnsureSmallCyclesOnDevice(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream)
 	{
-        var kernels = GetKernels(accelerator);
-        if (kernels.SmallCycles is { } buffer)
+		if (kernels.SmallCycles is { })
 		{
-            return buffer.View;
+			return;
 		}
 
-        // 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.
+		var host = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
+		MemoryBuffer1D<ulong, Stride1D.Dense>? device;
+
+		lock (accelerator)
 		{
-            if (kernels.SmallCycles is { } existing)
-            {
-                return existing.View;
+			device = accelerator.Allocate1D<ulong>(host.Length);
 		}
 
-            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);
+		device.View.CopyFromCPU(stream, host);
 		kernels.SmallCycles = device;
-            return device.View;
-        }
 	}
 
-    public static ResiduePrimeViews EnsureSmallPrimesOnDevice(Accelerator accelerator)
+	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)
 	{
-        var kernels = GetKernels(accelerator);
-        if (kernels.SmallPrimesLastOne is { } lastOne &&
-            kernels.SmallPrimesLastSeven is { } lastSeven &&
-            kernels.SmallPrimesPow2LastOne is { } lastOnePow2 &&
-            kernels.SmallPrimesPow2LastSeven is { } lastSevenPow2)
+		if (kernels.SmallPrimesLastOne is { } &&
+			kernels.SmallPrimesLastSeven is { } &&
+			kernels.SmallPrimesPow2LastOne is { } &&
+			kernels.SmallPrimesPow2LastSeven is { })
 		{
-            return new ResiduePrimeViews(lastOne.View, lastSeven.View, lastOnePow2.View, lastSevenPow2.View);
+			return;
 		}
 
-        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 { } &&
+			kernels.SmallPrimesLastSeven is { } &&
+			kernels.SmallPrimesPow2LastOne is { } &&
+			kernels.SmallPrimesPow2LastSeven is { })
 		{
-            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);
+			return;
 		}
 
 		var hostLastOne = PrimesGenerator.SmallPrimesLastOne;
@@ -190,100 +190,27 @@ public class GpuKernelPool
 		var hostLastOnePow2 = PrimesGenerator.SmallPrimesPow2LastOne;
 		var hostLastSevenPow2 = PrimesGenerator.SmallPrimesPow2LastSeven;
 
-            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);
+		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;
-
-            return new ResiduePrimeViews(deviceLastOne.View, deviceLastSeven.View, deviceLastOnePow2.View, deviceLastSevenPow2.View);
-        }
-    }
-
-    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);
-        }
-
-        lock (kernels)
-        {
-            if (kernels.SmallPrimeFactorsPrimes is { } existingPrimes && kernels.SmallPrimeFactorsSquares is { } existingSquares)
-            {
-                return new SmallPrimeFactorTables(existingPrimes, existingSquares, (int)existingPrimes.Length);
-            }
-
-            var hostPrimes = PrimesGenerator.SmallPrimes;
-            var hostSquares = PrimesGenerator.SmallPrimesPow2;
-
-            var devicePrimes = accelerator.Allocate1D<uint>(hostPrimes.Length);
-            devicePrimes.View.CopyFromCPU(hostPrimes);
-            var deviceSquares = accelerator.Allocate1D<ulong>(hostSquares.Length);
-            deviceSquares.View.CopyFromCPU(hostSquares);
-
-            kernels.SmallPrimeFactorsPrimes = devicePrimes;
-            kernels.SmallPrimeFactorsSquares = deviceSquares;
-
-            return new SmallPrimeFactorTables(devicePrimes, deviceSquares, hostPrimes.Length);
-        }
-    }
-
-    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;
-            }
-
-            MemoryBuffer1D<int, Stride1D.Dense>? exponentSlots = kernels.SmallPrimeFactorExponentSlots;
-            if (exponentSlots is null || exponentSlots.Length < slotCount)
-            {
-                exponentSlots?.Dispose();
-                exponentSlots = accelerator.Allocate1D<int>(slotCount);
-                kernels.SmallPrimeFactorExponentSlots = exponentSlots;
-            }
-
-            MemoryBuffer1D<int, Stride1D.Dense>? countSlot = kernels.SmallPrimeFactorCountSlot;
-            if (countSlot is null)
-            {
-                countSlot = accelerator.Allocate1D<int>(1);
-                kernels.SmallPrimeFactorCountSlot = countSlot;
-            }
-
-            MemoryBuffer1D<ulong, Stride1D.Dense>? remainingSlot = kernels.SmallPrimeFactorRemainingSlot;
-            if (remainingSlot is null)
-            {
-                remainingSlot = accelerator.Allocate1D<ulong>(1);
-                kernels.SmallPrimeFactorRemainingSlot = remainingSlot;
-            }
-
-            return new SmallPrimeFactorScratch(primeSlots!, exponentSlots!, countSlot!, remainingSlot!);
-        }
-    }
-
-    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);
 	}
 
 	/// <summary>
@@ -292,12 +219,14 @@ public class GpuKernelPool
 	/// <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();
+		// GpuPrimeWorkLimiter.Acquire();
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
+		var stream = AcceleratorStreamPool.Rent(acceleratorIndex);
 		action(accelerator, stream);
+		stream.Synchronize();
 		stream.Dispose();
-        lease.Dispose();
-
+		accelerator.Dispose();
+		// GpuPrimeWorkLimiter.Release();
 	}
 }

PerfectNumbers.Core/Gpu/GpuPrimeWorkLimiter.cs

@@ -1,6 +1,3 @@
-using System;
-using System.Threading;
-
 namespace PerfectNumbers.Core.Gpu;
 
 public static class GpuPrimeWorkLimiter
@@ -8,14 +5,8 @@ public static class GpuPrimeWorkLimiter
 	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)
     {
@@ -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

@@ -1,18 +1,26 @@
+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
 {
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	private static GpuUInt128 AdvancePolynomialGpu(GpuUInt128 x, ulong c, ulong modulus)
+	{
+		x.Pow2();
+		x.AddMod(c, modulus);
+		return x;
+	}
+
 	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]);
+		results[0] = ULongExtensions.Pow2MontgomeryModWindowedGpuKeepMontgomery(divisor, exponents[0]);
 	}
 
 	public static void Pow2MontgomeryKernelConvertToStandard(
@@ -21,6 +29,52 @@ internal static class Pow2MontgomeryKernels
 		MontgomeryDivisorData divisor,
 		ArrayView1D<ulong, Stride1D.Dense> results)
 	{
-        results[index] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponents[index]);
+		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/PrimeOrderKernels.cs

@@ -1,15 +1,51 @@
-using System;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using ILGPU;
-using ILGPU.Algorithms.Random;
 using ILGPU.Runtime;
-using PerfectNumbers.Core;
 
 namespace PerfectNumbers.Core.Gpu;
 
 internal static partial class PrimeOrderGpuHeuristics
 {
+	internal static void CheckFactorsKernel(
+		int count,
+		ulong phi,
+		ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense> toTest,
+		MontgomeryDivisorData divisorData,
+		ArrayView1D<ulong, Stride1D.Dense> result
+	)
+	{
+		ulong order = phi;
+		int entryCount = count;
+
+		for (int i = 0; i < entryCount; i++)
+		{
+			(ulong primeFactor, int exponent) = toTest[i];
+
+			for (int iteration = 0; iteration < exponent; iteration++)
+			{
+				ulong candidate = order / primeFactor;
+				ulong remainder = order - (candidate * primeFactor);
+				// ulong remainder = order % primeFactor;
+				if (remainder != 0UL)
+				{
+					break;
+				}
+
+				// ulong candidate = order / primeFactor;
+				if (ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisorData, candidate) == 1UL)
+				{
+					order = candidate;
+					continue;
+				}
+
+				break;
+			}
+		}
+
+		result[0] = order;
+	}
+
 	/// This kernel always sets the result and operates within the allowed / required bounds, never accessing elements outside of them.
 	/// It sets the value only when it needs to. The callers don't need to clear the output / input buffers.
 	internal static void PartialFactorKernel(
@@ -1170,16 +1206,6 @@ internal static partial class PrimeOrderGpuHeuristics
 		remainders[index] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponent);
 	}
 
-    private static Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>> GetPow2ModWideKernel(Accelerator accelerator)
-    {
-        return Pow2ModKernelWideCache.GetOrAdd(accelerator, static accel =>
-        {
-            var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(Pow2ModKernelWide);
-            var kernel = KernelUtil.GetKernel(loaded);
-            return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
-        });
-    }
-
 	/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.
 	internal static void Pow2ModKernelWide(Index1D index, ArrayView1D<GpuUInt128, Stride1D.Dense> exponents, GpuUInt128 modulus, ArrayView1D<GpuUInt128, Stride1D.Dense> remainders)
 	{
@@ -1367,5 +1393,4 @@ internal static partial class PrimeOrderGpuHeuristics
 		ulong directMask = (1UL << windowBitCount) - 1UL;
 		return exponent.Low & directMask;
 	}
-
 }

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

@@ -2,16 +2,16 @@ 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
@@ -20,12 +20,10 @@ internal static class PrimeTesterKernels
     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,
@@ -35,8 +33,8 @@ internal static class PrimeTesterKernels
         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)
         {
@@ -56,6 +54,8 @@ internal static class PrimeTesterKernels
                 primes = smallPrimesLastNine;
                 primeSquares = smallPrimesPow2LastNine;
 				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;
             }
         }
+    }
+
+    [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/MersenneNumberDivisorByDivisorGpuTester.cs

@@ -1,12 +1,10 @@
-using System;
 using System.Buffers;
 using System.Collections.Concurrent;
 using System.Runtime.CompilerServices;
 using System.Runtime.InteropServices;
-using System.Threading;
-using PerfectNumbers.Core;
 using ILGPU;
 using ILGPU.Runtime;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core.Gpu;
 
@@ -23,15 +21,26 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 	private ulong _divisorLimit;
 	// private bool _isConfigured;
 
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>> _kernelCache = new(AcceleratorReferenceComparer.Instance);
-    private readonly ConcurrentBag<BatchResources> _resourcePool = new();
-    private readonly ConcurrentBag<GpuContextPool.GpuContextLease> _acceleratorPool = new();
-    private readonly ConcurrentBag<DivisorScanSession> _sessionPool = new();
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>>? _kernel;
 
-    private Action<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> GetKernel(Accelerator accelerator) =>
-        _kernelCache.GetOrAdd(accelerator, acc =>
-            acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>(DivisorByDivisorKernels.CheckKernel));
+	private readonly ConcurrentBag<BatchResources> _resourcePool = [];
+	private readonly ConcurrentBag<DivisorScanSession> _sessionPool = [];
 
+	private static Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> GetKernel(Accelerator accelerator)
+	{
+		var pool = _kernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>(DivisorByDivisorKernels.CheckKernel);
+		var kernel = KernelUtil.GetKernel(loaded);
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
 
 	public int GpuBatchSize
 	{
@@ -59,6 +68,8 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		// _isConfigured = true;
 	}
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
 	public bool IsPrime(ulong prime, out bool divisorsExhausted)
 	{
 		ulong allowedMax;
@@ -91,8 +102,13 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		ulong processedCount;
 		ulong lastProcessed;
 
-        var gpuLease = RentAccelerator();
-        var accelerator = gpuLease.Accelerator;
+		// GpuPrimeWorkLimiter();
+		// int acceleratorIndex = AcceleratorPool.Shared.Rent();
+		// var accelerator = _accelerators[acceleratorIndex];
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(batchCapacity);
+		var acceleratorIndex = gpu.AcceleratorIndex;
+		var accelerator = gpu.Accelerator;
+		var stream = AcceleratorStreamPool.Rent(acceleratorIndex);
 
 		// Monitor.Enter(gpuLease.ExecutionLock);
 
@@ -100,9 +116,11 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		var resources = RentBatchResources(accelerator, batchCapacity);
 
 		composite = CheckDivisors(
+			gpu,
 			prime,
 			allowedMax,
 			kernel,
+			stream,
 			resources.DivisorDataBuffer,
 			resources.OffsetBuffer,
 			resources.CountBuffer,
@@ -120,9 +138,11 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 			out coveredRange,
 			out processedCount);
 
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		ReturnBatchResources(resources);
+		PrimeOrderCalculatorAccelerator.Return(gpu);
 		// Monitor.Exit(gpuLease.ExecutionLock);
-        ReturnAccelerator(gpuLease);
+		// GpuPrimeWorkLimiter.Release();
 
 		if (composite)
 		{
@@ -164,9 +184,11 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 	}
 
 	private static bool CheckDivisors(
+		PrimeOrderCalculatorAccelerator gpu,
 		ulong prime,
 		ulong allowedMax,
-        Action<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> kernel,
+		Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> kernel,
+		AcceleratorStream stream,
 		MemoryBuffer1D<GpuDivisorPartialData, Stride1D.Dense> divisorDataBuffer,
 		MemoryBuffer1D<int, Stride1D.Dense> offsetBuffer,
 		MemoryBuffer1D<int, Stride1D.Dense> countBuffer,
@@ -186,7 +208,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 	{
 		int batchCapacity = (int)divisorDataBuffer.Length;
 		bool composite = false;
-        bool processedAll = false;
+		bool processedAll;
 		processedCount = 0UL;
 		lastProcessed = 0UL;
 
@@ -284,7 +306,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 				{
 					ulong divisorValue = (ulong)nextDivisor128;
 					MontgomeryDivisorData montgomeryData = MontgomeryDivisorData.FromModulus(divisorValue);
-                    ulong divisorCycle = ResolveDivisorCycle(divisorValue, prime, in montgomeryData);
+					ulong divisorCycle = ResolveDivisorCycle(gpu, divisorValue, prime, in montgomeryData);
 					if (divisorCycle == prime)
 					{
 						processedCount += (ulong)(i + 1);
@@ -328,19 +350,20 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 				continue;
 			}
 
-            divisorDataView.CopyFromCPU(ref divisorDataRef, admissibleCount);
-            offsetView.CopyFromCPU(ref offsetRef, admissibleCount);
-            countView.CopyFromCPU(ref countRef, admissibleCount);
-            cycleView.CopyFromCPU(ref cycleRef, admissibleCount);
-            exponentViewDevice.CopyFromCPU(ref exponentRef, admissibleCount);
+			divisorDataView.CopyFromCPU(stream, ref divisorDataRef, admissibleCount);
+			offsetView.CopyFromCPU(stream, ref offsetRef, admissibleCount);
+			countView.CopyFromCPU(stream, ref countRef, admissibleCount);
+			cycleView.CopyFromCPU(stream, ref cycleRef, admissibleCount);
+			exponentViewDevice.CopyFromCPU(stream, ref exponentRef, admissibleCount);
 
 			int sentinel = int.MaxValue;
-            hitIndexView.CopyFromCPU(ref sentinel, 1);
+			hitIndexView.CopyFromCPU(stream, ref sentinel, 1);
 
-            kernel(admissibleCount, divisorDataView, offsetView, countView, exponentViewDevice, cycleView, hitsView, hitIndexView);
+			kernel(stream, new Index1D(admissibleCount), divisorDataView, offsetView, countView, exponentViewDevice, cycleView, hitsView, hitIndexView);
 
 			int firstHit = sentinel;
-            hitIndexView.CopyToCPU(ref firstHit, 1);
+			hitIndexView.CopyToCPU(stream, ref firstHit, 1);
+			stream.Synchronize();
 			int hitIndex = firstHit >= admissibleCount ? -1 : firstHit;
 			bool hitFound = hitIndex >= 0;
 			composite = hitFound;
@@ -355,9 +378,9 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static ulong ResolveDivisorCycle(ulong divisor, ulong prime, in MontgomeryDivisorData divisorData)
+	private static ulong ResolveDivisorCycle(PrimeOrderCalculatorAccelerator gpu, ulong divisor, ulong prime, in MontgomeryDivisorData divisorData)
 	{
-        if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(divisor, prime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
+		if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(gpu, divisor, prime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
 		{
 			return MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
 		}
@@ -476,9 +499,10 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 	public sealed class DivisorScanSession : IMersenneNumberDivisorByDivisorTester.IDivisorScanSession
 	{
 		private readonly MersenneNumberDivisorByDivisorGpuTester _owner;
-        private readonly GpuContextPool.GpuContextLease _lease;
+		private readonly PrimeOrderCalculatorAccelerator _primeTesterAccelerator;
 		private readonly Accelerator _accelerator;
-        private Action<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> _kernel = null!;
+		private readonly AcceleratorStream _stream;
+		private Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> _kernel = null!;
 		private MemoryBuffer1D<GpuDivisorPartialData, Stride1D.Dense> _divisorBuffer = null!;
 		private MemoryBuffer1D<int, Stride1D.Dense> _offsetBuffer = null!;
 		private MemoryBuffer1D<int, Stride1D.Dense> _countBuffer = null!;
@@ -489,11 +513,15 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		private ulong[] _hostBuffer = null!;
 		private int _capacity;
 
+
 		internal DivisorScanSession(MersenneNumberDivisorByDivisorGpuTester owner)
 		{
 			_owner = owner;
-            _lease = GpuContextPool.Rent();
-            _accelerator = _lease.Accelerator;
+			// GpuPrimeWorkLimiter.Acquire();
+			_primeTesterAccelerator = PrimeOrderCalculatorAccelerator.Rent(1);
+			Accelerator accelerator = _primeTesterAccelerator.Accelerator;
+			_accelerator = accelerator;
+			_stream = accelerator.CreateStream();
 			_capacity = Math.Max(1, owner._gpuBatchSize);
 		}
 
@@ -503,18 +531,23 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 
 		private void EnsureExecutionResourcesLocked(int requiredCapacity)
 		{
+			Accelerator accelerator = _accelerator;
+
 			if (_kernel == null)
 			{
-                _kernel = _owner.GetKernel(_accelerator);
+				_kernel = GetKernel(_accelerator);
 			}
 
 			if (_divisorBuffer == null)
 			{
-                _divisorBuffer = _accelerator.Allocate1D<GpuDivisorPartialData>(1);
-                _offsetBuffer = _accelerator.Allocate1D<int>(1);
-                _countBuffer = _accelerator.Allocate1D<int>(1);
-                _cycleBuffer = _accelerator.Allocate1D<ulong>(1);
-                _firstHitBuffer = _accelerator.Allocate1D<int>(1);
+				lock(accelerator)
+				{
+					_divisorBuffer = accelerator.Allocate1D<GpuDivisorPartialData>(1);
+					_offsetBuffer = accelerator.Allocate1D<int>(1);
+					_countBuffer = accelerator.Allocate1D<int>(1);
+					_cycleBuffer = accelerator.Allocate1D<ulong>(1);
+					_firstHitBuffer = accelerator.Allocate1D<int>(1);
+				}
 			}
 
 			bool allocateBuffers = _exponentsBuffer == null;
@@ -537,9 +570,13 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 			{
 				int desiredCapacity = requiredCapacity > _capacity ? requiredCapacity : _capacity;
 				_capacity = desiredCapacity;
-                _exponentsBuffer = _accelerator.Allocate1D<ulong>(_capacity);
-                _hitBuffer = _accelerator.Allocate1D<byte>(_capacity);
-                _hostBuffer = ThreadStaticPools.UlongPool.Rent(_capacity);
+				lock (accelerator)
+				{
+					_exponentsBuffer = accelerator.Allocate1D<ulong>(desiredCapacity);
+					_hitBuffer = accelerator.Allocate1D<byte>(desiredCapacity);
+				}
+
+				_hostBuffer = ThreadStaticPools.UlongPool.Rent(desiredCapacity);
 			}
 		}
 
@@ -565,7 +602,8 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 
 			if (cycle == 0UL)
 			{
-                if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(divisor, firstPrime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
+				var gpu = _primeTesterAccelerator;
+				if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(gpu, divisor, firstPrime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
 				{
 					cycle = MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
 				}
@@ -592,39 +630,41 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 			ArrayView1D<ulong, Stride1D.Dense> exponentView = _exponentsBuffer.View;
 			ArrayView1D<byte, Stride1D.Dense> hitView = _hitBuffer.View;
 
+			AcceleratorStream stream = _stream;
 			GpuDivisorPartialData partialData = new GpuDivisorPartialData(divisor);
-            divisorView.CopyFromCPU(ref partialData, 1);
+			divisorView.CopyFromCPU(stream, ref partialData, 1);
 
 			int offsetValue = 0;
-            offsetView.CopyFromCPU(ref offsetValue, 1);
+			offsetView.CopyFromCPU(stream, ref offsetValue, 1);
 
 			int countValue = length;
-            countView.CopyFromCPU(ref countValue, 1);
+			countView.CopyFromCPU(stream, ref countValue, 1);
 
 			ulong cycleValue = cycle;
-            cycleView.CopyFromCPU(ref cycleValue, 1);
+			cycleView.CopyFromCPU(stream, ref cycleValue, 1);
 
 			Span<ulong> hostSpan = _hostBuffer.AsSpan(0, length);
 			primes.CopyTo(hostSpan);
 			ref ulong hostRef = ref MemoryMarshal.GetReference(hostSpan);
 
-            exponentView.CopyFromCPU(ref hostRef, length);
+			exponentView.CopyFromCPU(stream, ref hostRef, length);
 
 			int sentinel = int.MaxValue;
-            firstHitView.CopyFromCPU(ref sentinel, 1);
+			firstHitView.CopyFromCPU(stream, ref sentinel, 1);
 
 			var kernel = _kernel!;
 
-            kernel(1, divisorView, offsetView, countView, exponentView, cycleView, hitView, firstHitView);
+			kernel(stream, new Index1D(1), divisorView, offsetView, countView, exponentView, cycleView, hitView, firstHitView);
 
 			Span<byte> hitSlice = hits.Slice(0, length);
 			ref byte hitRef = ref MemoryMarshal.GetReference(hitSlice);
-            hitView.CopyToCPU(ref hitRef, length);
+			hitView.CopyToCPU(stream, ref hitRef, length);
+			stream.Synchronize();
 
 			// Monitor.Exit(_lease.ExecutionLock);
 		}
 
-        public void Dispose()
+		public void Return()
 		{
 			_owner._sessionPool.Add(this);
 		}
@@ -650,17 +690,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		}
 	}
 
-    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 BatchResources RentBatchResources(Accelerator accelerator, int capacity)
 	{
@@ -767,6 +797,8 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		}
 
 		private void AllocateDeviceBuffers(Accelerator accelerator, int capacity)
+		{
+			lock (accelerator)
 			{
 				_divisorDataBuffer = accelerator.Allocate1D<GpuDivisorPartialData>(capacity);
 				_offsetBuffer = accelerator.Allocate1D<int>(capacity);
@@ -775,6 +807,8 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 				_exponentBuffer = accelerator.Allocate1D<ulong>(capacity);
 				_hitsBuffer = accelerator.Allocate1D<byte>(capacity);
 				_hitIndexBuffer = accelerator.Allocate1D<int>(1);
+			}
+
 			_deviceCapacity = capacity;
 		}
 
@@ -900,7 +934,3 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 
 	IMersenneNumberDivisorByDivisorTester.IDivisorScanSession IMersenneNumberDivisorByDivisorTester.CreateDivisorSession() => CreateDivisorSession();
 }
-
-
-
-

PerfectNumbers.Core/Gpu/MersenneNumberDivisorGpuTester.cs

@@ -1,4 +1,3 @@
-using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
 
@@ -9,11 +8,29 @@ 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;
+		}
+
+		return bufferQueue;
+	}
+
+	private static Action<AcceleratorStream, Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>> GetDivisorKernel(Accelerator accelerator)
+	{
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>>(DivisorKernels.Kernel);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ReadOnlyGpuUInt128, ArrayView<byte>>>();
+	}
 
 	public static void BuildDivisorCandidates()
 	{
@@ -35,27 +52,39 @@ public sealed class MersenneNumberDivisorGpuTester
 		_divisorCandidates = count == 0 ? [] : list[..count];
 	}
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
 	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));
+		// 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);
+			}
+		}
+
 		// 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();
+		// 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)
 	{

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);
+		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);
+		stream.Dispose();
+        ulongPool.Return(orderArray);
 		orderBuffer.Dispose();
-        gpuLease.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)
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>? _addSmallKernel;
+
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>? _subSmallKernel;
+	
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>>>? _reduceKernel;
+
+	[ThreadStatic]
+    private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>>? _isZeroKernel;
+
+	[ThreadStatic]
+    private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<GpuUInt128>, ArrayView<GpuUInt128>>>? _lucasLehmerBatchKernel;
+
+    private static Action<AcceleratorStream, Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>> GetLucasLehmerKernel(Accelerator accelerator)
 	{
-        return KernelCache.GetOrAdd(accelerator, accel => accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>(LucasLehmerKernels.Kernel));
+		var pool = _lucasLehmerKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
 		}
 
-    private Action<Index1D, ulong, ArrayView<GpuUInt128>> GetAddSmallKernel(Accelerator accelerator) =>
-        AddSmallKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.AddSmallKernel));
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>(LucasLehmerKernels.Kernel);
 
-    private Action<Index1D, ulong, ArrayView<GpuUInt128>> GetSubSmallKernel(Accelerator accelerator) =>
-        SubSmallKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.SubtractSmallKernel));
+		var kernel = KernelUtil.GetKernel(loaded);
 
-    private Action<Index1D, ulong, ArrayView<GpuUInt128>> GetReduceKernel(Accelerator accelerator) =>
-        ReduceKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView<GpuUInt128>>(LucasLehmerKernels.ReduceModMersenneKernel));
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, ArrayView<GpuUInt128>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
 
-    private Action<Index1D, ArrayView<GpuUInt128>, ArrayView<byte>> GetIsZeroKernel(Accelerator accelerator) =>
-        IsZeroKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, ArrayView<byte>>(LucasLehmerKernels.IsZeroKernel));
+	private static Action<AcceleratorStream, Index1D, ulong, ArrayView<GpuUInt128>> GetAddSmallKernel(Accelerator accelerator)
+	{
+		var pool = _addSmallKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
 
-    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));
+		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 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 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 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);
-        zeroKernel(1, stateBuffer.View, resultBuffer.View);
-        accelerator.Synchronize();
-        byte[] result = new byte[1];
-        resultBuffer.View.CopyToCPU(result);
+        isZeroKernel(stream, 1, stateBuffer.View, resultBuffer.View);
+        Span<byte> result = stackalloc byte[1];
+        resultBuffer.View.CopyToCPU(stream, result);
+		stream.Synchronize();
+		
         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

@@ -7,34 +7,46 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 	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)
 	{
-                var gpuLease = GpuKernelPool.GetKernel(_useGpuOrder);
-                var accelerator = gpuLease.Accelerator;
-                var stream = gpuLease.Stream;
+		// 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
 
+		KernelContainer kernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, KernelType.Pow2ModOrderKernelScan | KernelType.IncrementalOrderKernelScan);
 		var kernel = _kernelType switch
 		{
-                        GpuKernelType.Pow2Mod => gpuLease.Pow2ModOrderKernel,
-                        _ => gpuLease.IncrementalOrderKernel,
+			GpuKernelType.Pow2Mod => kernels.Pow2ModOrder!,
+			_ => kernels.IncrementalOrder!,
 		};
 
-                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.
+		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.
+
+		}
 		exponent.Mod10_8_5_3Steps(out ulong step10, out ulong step8, out ulong step5, out ulong step3);
 
 		UInt128 remaining;
 		int currentSize;
 		int found = 0;
+
+		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();
+			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
@@ -44,12 +56,13 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 			var kernelArgs = new ResidueAutomatonArgs(q0m10, step10, q0m8, step8, q0m3, step3, q0m5, step5);
 
 			// 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);
+			kernel(
+				stream, currentSize, exponent, (GpuUInt128)twoP, (GpuUInt128)kStart,
+				last, divMul, kernelArgs, foundBuffer.View, smallCyclesView
+			);
 
+			foundBuffer.View.CopyToCPU(stream, ref found, 1);
 			stream.Synchronize();
-                        foundBuffer.View.CopyToCPU(ref found, 1);
 			if (found != 0)
 			{
 				Volatile.Write(ref isPrime, false);
@@ -59,7 +72,8 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 			kStart += (UInt128)currentSize;
 		}
 
+		stream.Dispose();
 		foundBuffer.Dispose();
-                gpuLease.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;
@@ -13,9 +8,12 @@ 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);
+
+	[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;
+
+	[ThreadStatic]
+	private static Queue<ResidueResources>? _resourcePools;
 
 	// 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
@@ -23,26 +21,30 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
 	// 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.
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
 	// 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;
+		// GpuPrimeWorkLimiter.Acquire();
+		var acceleratorIndex = AcceleratorPool.Shared.Rent();
+		var accelerator = _accelerators[acceleratorIndex];
 
 		// Monitor.Enter(gpuLease.ExecutionLock);
 
 		var stream = accelerator.CreateStream();
 		var kernel = GetPow2ModKernel(accelerator);
 		var resources = RentResources(accelerator, GpuConstants.ScanBatchSize);
+		var gpuKernels = GpuKernelPool.GetOrAddKernels(acceleratorIndex, stream, KernelType.SmallCycles | KernelType.SmallPrimes);
 
 		var orderBuffer = resources.OrderBuffer;
 		ulong[] orderArray = resources.OrderArray;
 		int batchSize = resources.Capacity;
 
 		// Ensure device has small cycles and primes tables for in-kernel lookup
-        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(accelerator);
-        ResiduePrimeViews primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(accelerator);
+		// TODO: Pass-in the stream down the path
+		var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice(gpuKernels);
+		ResiduePrimeViews primeViews = GpuKernelPool.GetSmallPrimesOnDevice(gpuKernels);
 
 		UInt128 kStart = 1UL;
 		UInt128 limit = maxK + UInt128.One;
@@ -68,8 +70,8 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
 			kernel(stream, currentSize, exponent, twoPGpu, (GpuUInt128)kStart, last, 0UL,
 				kernelArgs, orderBuffer.View, smallCyclesView, primeViews.LastOne, primeViews.LastSeven, primeViews.LastOnePow2, primeViews.LastSevenPow2);
 
+			orderBuffer.View.CopyToCPU(stream, ref ordersRef, currentSize);
 			stream.Synchronize();
-            orderBuffer.View.CopyToCPU(ref ordersRef, currentSize);
 			if (!Volatile.Read(ref isPrime))
 			{
 				break;
@@ -106,80 +108,99 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
 			}
 		}
 
-        ReturnResources(accelerator, resources);
+		ReturnResources(resources);
 		stream.Dispose();
 		// Monitor.Exit(gpuLease.ExecutionLock);
-        ReturnAccelerator(gpuLease);
-        limiter.Dispose();
+		gpuKernels.Dispose();
+		// GpuPrimeWorkLimiter.Release();
 	}
 
-    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 =>
+	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)
 	{
-                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(
+
+		var pool = _pow2ModKernel ?? [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		cached = CreatePow2ModKernel(accelerator);
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	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);
+
+		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);
-            });
-
-    private GpuContextPool.GpuContextLease RentAccelerator()
-    {
-        if (_acceleratorPool.TryTake(out var lease))
-        {
-            return lease;
+		return kernel;
 	}
 
-        return GpuContextPool.Rent();
-    }
 
-    private void ReturnAccelerator(GpuContextPool.GpuContextLease lease) => _acceleratorPool.Add(lease);
-
-    private ResidueResources RentResources(Accelerator accelerator, int capacity)
+	private static ResidueResources RentResources(Accelerator accelerator, int capacity)
 	{
-        var bag = _resourcePools.GetOrAdd(accelerator, static _ => new ConcurrentBag<ResidueResources>());
-        while (bag.TryTake(out var resources))
+		var resourcePools = _resourcePools ??= new();
+		if (resourcePools.TryDequeue(out var bag))
 		{
-            if (resources.Capacity >= capacity)
+			bag.EnsureCapacity(accelerator, capacity);
+			return bag;
+		}
+
+		return new ResidueResources(accelerator, capacity);
+	}
+
+	private static void ReturnResources(ResidueResources resources)
 	{
-				// The caller always overwrites the required elements. We don't need to worry about clearing the output buffer.
-                // resources.OrderBuffer.MemSetToZero();
-                return resources;
+		var resourcePools = _resourcePools ??= new();
+		resourcePools.Enqueue(resources);
 	}
 
-            resources.Dispose();
-        }
-
-        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;
-    }
-
-    private void ReturnResources(Accelerator accelerator, ResidueResources resources) =>
-        _resourcePools.GetOrAdd(accelerator, static _ => new ConcurrentBag<ResidueResources>()).Add(resources);
-
 	private sealed class ResidueResources
 	{
 		internal ResidueResources(Accelerator accelerator, int capacity)
 		{
 			Capacity = Math.Max(1, capacity);
+			// lock (accelerator)
+			{
 				OrderBuffer = accelerator.Allocate1D<ulong>(Capacity);
-            OrderArray = ArrayPool<ulong>.Shared.Rent(Capacity);
 			}
 
-        internal MemoryBuffer1D<ulong, Stride1D.Dense> OrderBuffer { get; }
+			OrderArray = ThreadStaticPools.UlongPool.Rent(Capacity);
+		}
 
-        internal ulong[] OrderArray { get; }
+		internal void EnsureCapacity(Accelerator accelerator, int capacity)
+		{
+			if (OrderBuffer.Length >= capacity)
+			{
+				return;
+			}
 
-        internal int Capacity { get; }
+			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();
-            ArrayPool<ulong>.Shared.Return(OrderArray, clearArray: false);
+			ThreadStaticPools.UlongPool.Return(OrderArray, clearArray: false);
 		}
 	}
 

PerfectNumbers.Core/Gpu/NttGpuMath.cs

@@ -1,4 +1,3 @@
-using System.Buffers;
 using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
@@ -29,23 +28,60 @@ public static class NttGpuMath
     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.
 
-    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.
+	
+	[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
     {
@@ -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);
@@ -330,99 +365,191 @@ public static class NttGpuMath
     {
         if (accelerator is null)
         {
-            foreach (var perAccel in SquareCache.Values)
-            {
-                foreach (var entry in perAccel.Values)
+            foreach (var entry in SquareCache.Values)
             {
 				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)
+		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;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuUInt128>, GpuUInt128, GpuUInt128>(NttPointwiseKernels.ScaleKernel);
+		
+		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
@@ -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);
+		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);
@@ -638,70 +766,70 @@ public static class NttGpuMath
 		{
 			nInv.ModInv(modulus);
 		}
-        kernel(n, inputBuffer.View, outputBuffer.View, n, modulus, rootInv, nInv);
-        accelerator.Synchronize();
-        outputBuffer.View.CopyToCPU(ref outputArray[0], n);
+
+		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)
+    private static void ForwardDevice(Accelerator accelerator, AcceleratorStream stream, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
 	{
-        var bitKernel = GetBitReverseKernel(accelerator);
-        bitKernel(length, data, cache.Bits);
-        accelerator.Synchronize();
+		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();
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesMont!.View, cache.ModulusLow, cache.MontNPrime64);
 			}
-            var fromMont = GetFromMont64Kernel(accelerator);
-            fromMont(length, data, 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
@@ -723,39 +850,36 @@ 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)
+    private static void InverseDevice(Accelerator accelerator, AcceleratorStream stream, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
 	{
-        var bitKernel = GetBitReverseKernel(accelerator);
-        bitKernel(length, data, cache.Bits);
-        accelerator.Synchronize();
+		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();
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInvMont!.View, cache.ModulusLow, cache.MontNPrime64);
 			}
-            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);
+
+			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)
         {
@@ -765,12 +889,11 @@ public static class NttGpuMath
 			{
 				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();
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
 			}
-            var scaleKernel = GetScaleBarrett128Kernel(accelerator);
-            scaleKernel(length, data, cache.NInv, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
-            accelerator.Synchronize();
+
+			var scaleBarrett128Kernel = GetScaleBarrett128Kernel(accelerator);
+            scaleBarrett128Kernel(stream, length, data, cache.NInv, cache.ModulusHigh, cache.ModulusLow, cache.BarrettMuHigh, cache.BarrettMuLow);
         }
         else
         {
@@ -780,15 +903,14 @@ public static class NttGpuMath
 			{
 				int len = 1 << (s + 1);
 				int half = len >> 1;
-                stageKernel(butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, modulus);
-                accelerator.Synchronize();
+				stageKernel(stream, butterflies, data, len, half, cache.StageOffsets[s], cache.TwiddlesInv.View, modulus);
 			}
+
 			var scaleKernel = GetScaleKernel(accelerator);
-            scaleKernel(length, data, cache.NInv, modulus);
-            accelerator.Synchronize();
+            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/PrimeOrderGpuHeuristics.cs

@@ -1,12 +1,10 @@
 using System.Buffers;
 using System.Collections.Concurrent;
+using System.Runtime.InteropServices;
 using ILGPU;
 using ILGPU.Runtime;
-using System.Numerics;
-using PerfectNumbers.Core;
-using System.Runtime.InteropServices;
-using System.Runtime.CompilerServices;
 using ILGPU.Runtime.OpenCL;
+using PerfectNumbers.Core.Gpu.Accelerators;
 namespace PerfectNumbers.Core.Gpu;
 
 internal enum GpuPow2ModStatus
@@ -18,13 +16,22 @@ internal enum GpuPow2ModStatus
 
 internal static partial class PrimeOrderGpuHeuristics
 {
+	internal static void PreloadStaticTables(AcceleratorStream stream)
+	{
+		// _ = GetSmallPrimeDeviceCache(stream);
+	}
+
 	private static readonly ConcurrentDictionary<ulong, byte> OverflowedPrimes = new();
 	private static readonly ConcurrentDictionary<UInt128, byte> OverflowedPrimesWide = new();
-    private static readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>> Pow2ModKernelCache = new();
-    private static readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>> Pow2ModKernelWideCache = new();
-    private static readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>> PartialFactorKernelCache = new();
-    private static readonly ConcurrentDictionary<Accelerator, OrderKernelLauncher> OrderKernelCache = new();
-    private static readonly ConcurrentDictionary<Accelerator, SmallPrimeDeviceCache> SmallPrimeDeviceCaches = new();
+
+	[ThreadStatic]
+	private static Kernel[]? _pow2ModKernel;
+
+	[ThreadStatic]
+	private static Kernel[]? _partialFactorKernel;
+
+	[ThreadStatic]
+	private static Kernel[]? _orderKernel;
 
 	public readonly struct OrderKernelConfig(ulong previousOrder, byte hasPreviousOrder, uint smallFactorLimit, int maxPowChecks, int mode)
 	{
@@ -59,25 +66,7 @@ internal static partial class PrimeOrderGpuHeuristics
 		public readonly ArrayView1D<byte, Stride1D.Dense> Status = status;
 	}
 
-    private delegate void OrderKernelLauncher(
-        AcceleratorStream stream,
-        Index1D extent,
-        ulong prime,
-        OrderKernelConfig config,
-        MontgomeryDivisorData divisor,
-        ArrayView1D<uint, Stride1D.Dense> primes,
-        ArrayView1D<ulong, Stride1D.Dense> squares,
-        int primeCount,
-        OrderKernelBuffers buffers);
-
-    private sealed class SmallPrimeDeviceCache
-    {
-        public MemoryBuffer1D<uint, Stride1D.Dense>? Primes;
-        public MemoryBuffer1D<ulong, Stride1D.Dense>? Squares;
-        public int Count;
-    }
-
-    private const int WideStackThreshold = 8;
+	private const int WideStackThreshold = 12;
 	private static PrimeOrderGpuCapability s_capability = PrimeOrderGpuCapability.Default;
 
 	private const int Pow2WindowSizeBits = 8;
@@ -101,47 +90,33 @@ internal static partial class PrimeOrderGpuHeuristics
 		s_capability = PrimeOrderGpuCapability.Default;
 	}
 
-    private static SmallPrimeDeviceCache GetSmallPrimeDeviceCache(Accelerator accelerator)
+	private static Kernel GetPartialFactorKernel(int acceleratorIndex)
 	{
-        return SmallPrimeDeviceCaches.GetOrAdd(accelerator, static acc =>
+		var pool = _partialFactorKernel ??= [];
+		if (pool[acceleratorIndex] is {} cached)
 		{
-            uint[] primes = PrimesGenerator.SmallPrimes;
-            ulong[] squares = PrimesGenerator.SmallPrimesPow2;
-            var primeBuffer = acc.Allocate1D<uint>(primes.Length);
-            primeBuffer.View.CopyFromCPU(primes);
-            var squareBuffer = acc.Allocate1D<ulong>(squares.Length);
-            squareBuffer.View.CopyFromCPU(squares);
-            return new SmallPrimeDeviceCache
-            {
-                Primes = primeBuffer,
-                Squares = squareBuffer,
-                Count = primes.Length,
-            };
-        });
+			return cached;
 		}
 
-    private static Action<AcceleratorStream, Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> GetPartialFactorKernel(Accelerator accelerator)
-    {
-        return PartialFactorKernelCache.GetOrAdd(accelerator, static accel =>
-        {
-            var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>(PartialFactorKernel);
-            var kernel = KernelUtil.GetKernel(loaded);
-            return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>>();
-        });
+		var accelerator = _accelerators[acceleratorIndex];
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>(PartialFactorKernel);
+
+		cached = KernelUtil.GetKernel(loaded);
+		pool[acceleratorIndex] = cached;
+		return cached;
 	}
 
 	public static bool TryPartialFactor(
+		PrimeOrderCalculatorAccelerator gpu,
 		ulong value,
 		uint limit,
 		Span<ulong> primeTargets,
 		Span<int> exponentTargets,
 		out int factorCount,
-        out ulong remaining,
-        out bool fullyFactored)
+		out ulong remaining)
 	{
 		factorCount = 0;
 		remaining = value;
-        fullyFactored = false;
 
 		if (primeTargets.Length == 0 || exponentTargets.Length == 0)
 		{
@@ -152,13 +127,13 @@ internal static partial class PrimeOrderGpuHeuristics
 		exponentTargets.Clear();
 
 		if (!TryLaunchPartialFactorKernel(
+				gpu,
 				value,
 				limit,
 				primeTargets,
 				exponentTargets,
 				out int extracted,
-                out ulong leftover,
-                out bool kernelFullyFactored))
+				out ulong leftover))
 		{
 			return false;
 		}
@@ -168,58 +143,61 @@ internal static partial class PrimeOrderGpuHeuristics
 		{
 			factorCount = 0;
 			remaining = value;
-            fullyFactored = false;
 			return false;
 		}
 
 		factorCount = extracted;
 		remaining = leftover;
-        fullyFactored = kernelFullyFactored && leftover == 1UL;
 		return true;
 	}
 
+	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
+
 	private static bool TryLaunchPartialFactorKernel(
+		PrimeOrderCalculatorAccelerator gpu,
 		ulong value,
 		uint limit,
 		Span<ulong> primeTargets,
 		Span<int> exponentTargets,
 		out int factorCount,
-        out ulong remaining,
-        out bool fullyFactored)
+		out ulong remaining)
 	{
 		factorCount = 0;
 		remaining = value;
-        fullyFactored = false;
 
 		try
 		{
-            var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-            Accelerator accelerator = lease.Accelerator;
-            AcceleratorStream stream = lease.Stream;
+			// GpuPrimeWorkLimiter.Acquire();
+			int acceleratorIndex = AcceleratorPool.Shared.Rent();
+			Accelerator accelerator = _accelerators[acceleratorIndex];
+			AcceleratorStream stream = accelerator.CreateStream();
 
-            var kernel = GetPartialFactorKernel(accelerator);
-            SmallPrimeDeviceCache cache = GetSmallPrimeDeviceCache(accelerator);
+			var smallPrimesView = gpu.SmallPrimeFactorPrimes;
+			var smallSquaresView = gpu.SmallPrimeFactorSquares;
 
 			// TODO: We should create / reallocate these buffer only once or if the new required length is bigger than capacity.
-            var factorBuffer = accelerator.Allocate1D<ulong>(primeTargets.Length);
-            var exponentBuffer = accelerator.Allocate1D<int>(exponentTargets.Length);
-            var countBuffer = accelerator.Allocate1D<int>(1);
-            var remainingBuffer = accelerator.Allocate1D<ulong>(1);
-            var fullyFactoredBuffer = accelerator.Allocate1D<byte>(1);
+			MemoryBuffer1D<ulong, Stride1D.Dense>? factorBuffer = accelerator.Allocate1D<ulong>(primeTargets.Length);
+			MemoryBuffer1D<int, Stride1D.Dense>? exponentBuffer = accelerator.Allocate1D<int>(exponentTargets.Length);
+			MemoryBuffer1D<int, Stride1D.Dense>? countBuffer = accelerator.Allocate1D<int>(1);
+			MemoryBuffer1D<ulong, Stride1D.Dense>? remainingBuffer = accelerator.Allocate1D<ulong>(1);
+			MemoryBuffer1D<byte, Stride1D.Dense>? fullyFactoredBuffer = accelerator.Allocate1D<byte>(1);
 
-            // TODO: There is no need to clear these buffers because the kernel will always assign values within the required bounds.
-			// factorBuffer.MemSetToZero();
-            // exponentBuffer.MemSetToZero();
-            // countBuffer.MemSetToZero();
-            // remainingBuffer.MemSetToZero();
-            // fullyFactoredBuffer.MemSetToZero();
+			// There is no need to clear these buffers because the kernel will always assign values within the required bounds. Keep it commented out.
+			// factorBuffer.MemSetToZero(stream);
+			// exponentBuffer.MemSetToZero(stream);
+			// countBuffer.MemSetToZero(stream);
+			// remainingBuffer.MemSetToZero(stream);
+			// fullyFactoredBuffer.MemSetToZero(stream);
 
-            kernel(
+			var partialFactorKernel = GetPartialFactorKernel(acceleratorIndex);
+			var kernelLauncher = partialFactorKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>>();
+
+			kernelLauncher(
 				stream,
 				1,
-                cache.Primes!.View,
-                cache.Squares!.View,
-                cache.Count,
+				smallPrimesView,
+				smallSquaresView,
+				(int)smallPrimesView.Length,
 				primeTargets.Length,
 				value,
 				limit,
@@ -229,17 +207,18 @@ internal static partial class PrimeOrderGpuHeuristics
 				remainingBuffer.View,
 				fullyFactoredBuffer.View);
 
-            stream.Synchronize();
 
-            countBuffer.View.CopyToCPU(ref factorCount, 1);
-            factorCount = Math.Min(factorCount, primeTargets.Length);
-            factorBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(primeTargets), primeTargets.Length);
-            exponentBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(exponentTargets), exponentTargets.Length);
-            remainingBuffer.View.CopyToCPU(ref remaining, 1);
+			countBuffer.View.CopyToCPU(stream, ref factorCount, 1);
+			factorBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(primeTargets), primeTargets.Length);
+			exponentBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(exponentTargets), exponentTargets.Length);
+			remainingBuffer.View.CopyToCPU(stream, ref remaining, 1);
 
 			byte fullyFactoredFlag = 0;
-            fullyFactoredBuffer.View.CopyToCPU(ref fullyFactoredFlag, 1);
-            fullyFactored = fullyFactoredFlag != 0;
+			fullyFactoredBuffer.View.CopyToCPU(stream, ref fullyFactoredFlag, 1);
+			stream.Synchronize();
+			stream.Dispose();
+
+			factorCount = Math.Min(factorCount, primeTargets.Length);
 
 			// TODO: These buffers shouldn't be disposed but rather reused accross call with the assigned accelerator.
 			factorBuffer.Dispose();
@@ -247,24 +226,18 @@ internal static partial class PrimeOrderGpuHeuristics
 			countBuffer.Dispose();
 			remainingBuffer.Dispose();
 			fullyFactoredBuffer.Dispose();
-            lease.Dispose();
+			// GpuPrimeWorkLimiter.Release();
 			return true;
 		}
 		catch (CLException ex)
 		{
 			Console.WriteLine($"GPU ERROR ({ex.Error}): {ex.Message}");
-            factorCount = 0;
-            remaining = value;
-            fullyFactored = false;
-            return false;
+			throw;
 		}
 		catch (Exception ex) when (ex is AcceleratorException or InternalCompilerException or NotSupportedException or InvalidOperationException or AggregateException)
 		{
 			Console.WriteLine($"GPU ERROR: {ex.Message}");
-            factorCount = 0;
-            remaining = value;
-            fullyFactored = false;
-            return false;
+			throw;
 		}
 	}
 
@@ -351,6 +324,7 @@ internal static partial class PrimeOrderGpuHeuristics
 	}
 
 	internal static bool TryCalculateOrder(
+		PrimeOrderCalculatorAccelerator gpu,
 		ulong prime,
 		ulong? previousOrder,
 		PrimeOrderCalculator.PrimeOrderSearchConfig config,
@@ -359,38 +333,52 @@ internal static partial class PrimeOrderGpuHeuristics
 	{
 		order = 0UL;
 
-        var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-        Accelerator accelerator = lease.Accelerator;
-        AcceleratorStream stream = lease.Stream;
-
-        var kernel = GetOrderKernel(accelerator);
-            SmallPrimeDeviceCache cache = GetSmallPrimeDeviceCache(accelerator);
+		// GpuPrimeWorkLimiter.Acquire();
+		int acceleratorIndex = AcceleratorPool.Shared.Rent();
+		Accelerator accelerator = _accelerators[acceleratorIndex];
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		var smallPrimesView = gpu.SmallPrimeFactorPrimes;
+		var smallSquaresView = gpu.SmallPrimeFactorSquares;
 
 		// TODO: These buffers should be created reallocated once and assigned to an accelerator. Callers should use their own
 		// thread static cache to prevent other threads from using taking accelerators with pre-allocated buffers if they don't
 		// use them.
-        var phiFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
-        var phiExponentBuffer = accelerator.Allocate1D<int>(GpuSmallPrimeFactorSlots);
-        var workFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
-        var workExponentBuffer = accelerator.Allocate1D<int>(GpuSmallPrimeFactorSlots);
-        var candidateBuffer = accelerator.Allocate1D<ulong>(HeuristicCandidateLimit);
-        var stackIndexBuffer = accelerator.Allocate1D<int>(HeuristicStackCapacity);
-        var stackExponentBuffer = accelerator.Allocate1D<int>(HeuristicStackCapacity);
-        var stackProductBuffer = accelerator.Allocate1D<ulong>(HeuristicStackCapacity);
-        var resultBuffer = accelerator.Allocate1D<ulong>(1);
-        var statusBuffer = accelerator.Allocate1D<byte>(1);
+		MemoryBuffer1D<ulong, Stride1D.Dense>? phiFactorBuffer;
+		MemoryBuffer1D<int, Stride1D.Dense>? phiExponentBuffer;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? workFactorBuffer;
+		MemoryBuffer1D<int, Stride1D.Dense>? workExponentBuffer;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? candidateBuffer;
+		MemoryBuffer1D<int, Stride1D.Dense>? stackIndexBuffer;
+		MemoryBuffer1D<int, Stride1D.Dense>? stackExponentBuffer;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? stackProductBuffer;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? resultBuffer;
+		MemoryBuffer1D<byte, Stride1D.Dense>? statusBuffer;
+
+		lock(accelerator)
+		{
+			phiFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
+			phiExponentBuffer = accelerator.Allocate1D<int>(GpuSmallPrimeFactorSlots);
+			workFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
+			workExponentBuffer = accelerator.Allocate1D<int>(GpuSmallPrimeFactorSlots);
+			candidateBuffer = accelerator.Allocate1D<ulong>(HeuristicCandidateLimit);
+			stackIndexBuffer = accelerator.Allocate1D<int>(HeuristicStackCapacity);
+			stackExponentBuffer = accelerator.Allocate1D<int>(HeuristicStackCapacity);
+			stackProductBuffer = accelerator.Allocate1D<ulong>(HeuristicStackCapacity);
+			resultBuffer = accelerator.Allocate1D<ulong>(1);
+			statusBuffer = accelerator.Allocate1D<byte>(1);
+		}
 
 		// TODO: Remove the cleaning after the order kernel is modified to always set the result.
-		phiFactorBuffer.MemSetToZero();
-        phiExponentBuffer.MemSetToZero();
-        workFactorBuffer.MemSetToZero();
-        workExponentBuffer.MemSetToZero();
-        candidateBuffer.MemSetToZero();
-        stackIndexBuffer.MemSetToZero();
-        stackExponentBuffer.MemSetToZero();
-        stackProductBuffer.MemSetToZero();
-        resultBuffer.MemSetToZero();
-        statusBuffer.MemSetToZero();
+		phiFactorBuffer.MemSetToZero(stream);
+		phiExponentBuffer.MemSetToZero(stream);
+		workFactorBuffer.MemSetToZero(stream);
+		workExponentBuffer.MemSetToZero(stream);
+		candidateBuffer.MemSetToZero(stream);
+		stackIndexBuffer.MemSetToZero(stream);
+		stackExponentBuffer.MemSetToZero(stream);
+		stackProductBuffer.MemSetToZero(stream);
+		resultBuffer.MemSetToZero(stream);
+		statusBuffer.MemSetToZero(stream);
 
 		uint limit = config.SmallFactorLimit == 0 ? uint.MaxValue : config.SmallFactorLimit;
 		byte hasPrevious = previousOrder.HasValue ? (byte)1 : (byte)0;
@@ -409,51 +397,27 @@ internal static partial class PrimeOrderGpuHeuristics
 			resultBuffer.View,
 			statusBuffer.View);
 
-        kernel(
+		var orderKernel = GetOrderKernel(acceleratorIndex);
+		var kernelLauncher = orderKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, OrderKernelConfig, MontgomeryDivisorData, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>>();
+
+		kernelLauncher(
 			stream,
 			1,
 			prime,
 			kernelConfig,
 			divisorData,
-            cache.Primes!.View,
-            cache.Squares!.View,
-            cache.Count,
+			smallPrimesView,
+			smallSquaresView,
+			(int)smallPrimesView.Length,
 			buffers);
 
-        stream.Synchronize();
 
 		byte status = 0;
-        statusBuffer.View.CopyToCPU(ref status, 1);
+		statusBuffer.View.CopyToCPU(stream, ref status, 1);
+		resultBuffer.View.CopyToCPU(stream, ref order, 1);
+		stream.Synchronize();
 
-        PrimeOrderKernelStatus kernelStatus = (PrimeOrderKernelStatus)status;
-        if (kernelStatus == PrimeOrderKernelStatus.Fallback)
-        {
-            DisposeResources();
-            lease.Dispose();
-            return false;
-        }
-
-        if (kernelStatus == PrimeOrderKernelStatus.PollardOverflow)
-        {
-            DisposeResources();
-            lease.Dispose();
-            throw new InvalidOperationException("GPU Pollard Rho stack overflow; increase HeuristicStackCapacity.");
-        }
-
-        resultBuffer.View.CopyToCPU(ref order, 1);
-
-        if (kernelStatus == PrimeOrderKernelStatus.FactoringFailure)
-        {
-            order = 0UL;
-        }
-
-        DisposeResources();
-        lease.Dispose();
-
-        return order != 0UL;
-
-        void DisposeResources()
-        {
+		stream.Dispose();
 		phiFactorBuffer.Dispose();
 		phiExponentBuffer.Dispose();
 		workFactorBuffer.Dispose();
@@ -464,17 +428,41 @@ internal static partial class PrimeOrderGpuHeuristics
 		stackProductBuffer.Dispose();
 		resultBuffer.Dispose();
 		statusBuffer.Dispose();
-        }
+		// GpuPrimeWorkLimiter.Release();
+
+		PrimeOrderKernelStatus kernelStatus = (PrimeOrderKernelStatus)status;
+		if (kernelStatus == PrimeOrderKernelStatus.Fallback)
+		{
+			return false;
 		}
 
-    private static OrderKernelLauncher GetOrderKernel(Accelerator accelerator)
+		if (kernelStatus == PrimeOrderKernelStatus.PollardOverflow)
 		{
-        return OrderKernelCache.GetOrAdd(accelerator, static accel =>
+			throw new InvalidOperationException("GPU Pollard Rho stack overflow; increase HeuristicStackCapacity.");
+		}
+
+		if (kernelStatus == PrimeOrderKernelStatus.FactoringFailure)
 		{
-            var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, OrderKernelConfig, MontgomeryDivisorData, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>(CalculateOrderKernel);
+			order = 0UL;
+		}
+
+		return order != 0UL;
+	}
+
+	private static Kernel GetOrderKernel(int acceleratorIndex)
+	{
+		var pool = _orderKernel ??= [];
+		if (pool[acceleratorIndex] is {} cached)
+		{
+			return cached;
+		}
+
+		var accelerator = _accelerators[acceleratorIndex];
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, OrderKernelConfig, MontgomeryDivisorData, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>(CalculateOrderKernel);
+
 		var kernel = KernelUtil.GetKernel(loaded);
-            return kernel.CreateLauncherDelegate<OrderKernelLauncher>();
-        });
+		pool[acceleratorIndex] = kernel;
+		return kernel;
 	}
 
 	private static GpuPow2ModStatus TryPow2ModBatchInternal(ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> remainders)
@@ -489,7 +477,7 @@ internal static partial class PrimeOrderGpuHeuristics
 			throw new ArgumentException("Remainder span is shorter than the exponent span.", nameof(remainders));
 		}
 
-        Span<UInt128> target = remainders.Slice(0, exponents.Length);
+		Span<UInt128> target = remainders[..exponents.Length];
 		target.Clear();
 
 		ConcurrentDictionary<UInt128, byte> overflowRegistryWide = OverflowedPrimesWide;
@@ -515,45 +503,57 @@ internal static partial class PrimeOrderGpuHeuristics
 			}
 		}
 
-        bool computed = TryComputeOnGpuWide(exponents, prime, target);
-        return computed ? GpuPow2ModStatus.Success : GpuPow2ModStatus.Unavailable;
+		ComputeOnGpuWide(exponents, prime, target);
+		return GpuPow2ModStatus.Success;
 	}
 
 	private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
 	{
-        var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-        Accelerator accelerator = lease.Accelerator;
-        AcceleratorStream stream = lease.Stream;
-        var kernel = GetPow2ModKernel(accelerator);
+		// GpuPrimeWorkLimiter.Acquire();
+		int acceleratorIndex = AcceleratorPool.Shared.Rent();
+		Accelerator accelerator = _accelerators[acceleratorIndex];
+		AcceleratorStream? stream = accelerator.CreateStream();
 		// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
 		// Modify the callers to use their own pool of buffers per accelerator, so that other threads don't use the accelerators
 		// with out preallocated buffers. Share the pool with Pow2ModWide kernel.
-        var exponentBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
-        var remainderBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
+		MemoryBuffer1D<ulong, Stride1D.Dense>? exponentBuffer;
+		MemoryBuffer1D<ulong, Stride1D.Dense>? remainderBuffer;
+		lock(accelerator)
+		{
+			exponentBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
+			remainderBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
+		}
 
-        exponentBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(exponents), exponents.Length);
+		exponentBuffer.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(exponents), exponents.Length);
 
 		// Pow2Mod kernel always assigns the required output elements so we don't need to worry about clearing these.
-		// remainderBuffer.MemSetToZero();
+		// remainderBuffer.MemSetToZero(stream);
+
+		var pow2ModKernel = GetPow2ModKernel(acceleratorIndex);
+		var kernelLauncher = pow2ModKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
 
 		try
 		{
-			kernel(stream, exponents.Length, exponentBuffer.View, divisorData, remainderBuffer.View);
+			kernelLauncher(stream, exponents.Length, exponentBuffer.View, divisorData, remainderBuffer.View);
+			remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(results), exponents.Length);
 			stream.Synchronize();
+
+			stream.Dispose();
+			stream = null;
+			exponentBuffer.Dispose();
+			remainderBuffer.Dispose();
+			// GpuPrimeWorkLimiter.Release();
 		}
 		catch (Exception)
 		{
 			Console.WriteLine($"Exception for {prime} and exponents: {string.Join(",", exponents.ToArray())}.");
+			stream?.Dispose();
 			exponentBuffer.Dispose();
 			remainderBuffer.Dispose();
-			lease.Dispose();
+			// GpuPrimeWorkLimiter.Release();
 			throw;
 		}
 
-        remainderBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(results), exponents.Length);
-        exponentBuffer.Dispose();
-        remainderBuffer.Dispose();
-        lease.Dispose();
 		return true;
 	}
 
@@ -566,80 +566,64 @@ internal static partial class PrimeOrderGpuHeuristics
 		}
 	}
 
-    private static bool TryComputeOnGpuWide(ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> results)
+	private static void ComputeOnGpuWide(ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> results)
 	{
 		int length = exponents.Length;
-        if (length == 0)
-        {
-            return true;
-        }
 
 		GpuUInt128[]? rentedExponents = null;
 		GpuUInt128[]? rentedResults = null;
-        var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-
-        try
-        {
-            Accelerator accelerator = lease.Accelerator;
-            AcceleratorStream stream = lease.Stream;
-            var kernel = GetPow2ModWideKernel(accelerator);
-			// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
-			// Modify the callers to use their own pool of buffers per accelerator, so that other threads don't use the accelerators
-			// with out preallocated buffers. Share the pool with Pow2Mod kernel.
-            var exponentBuffer = accelerator.Allocate1D<GpuUInt128>(length);
-            var remainderBuffer = accelerator.Allocate1D<GpuUInt128>(length);
-
-            Span<GpuUInt128> exponentSpan = length <= WideStackThreshold
+		ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
+		PrimeOrderCalculatorAccelerator? gpu = null;
+		bool poolingRequired = length > WideStackThreshold;
+		Span<GpuUInt128> exponentSpan = poolingRequired
 			? stackalloc GpuUInt128[length]
-                : new Span<GpuUInt128>(rentedExponents = ArrayPool<GpuUInt128>.Shared.Rent(length), 0, length);
+			: new Span<GpuUInt128>(rentedExponents = gpuUInt128Pool.Rent(length), 0, length);
+		Span<GpuUInt128> resultSpan = length <= WideStackThreshold
+			? stackalloc GpuUInt128[length]
+			: new Span<GpuUInt128>(rentedResults = gpuUInt128Pool.Rent(length), 0, length);
+
 
 		for (int i = 0; i < length; i++)
 		{
 			exponentSpan[i] = (GpuUInt128)exponents[i];
 		}
 
-            exponentBuffer.View.CopyFromCPU(ref MemoryMarshal.GetReference(exponentSpan), length);
-			// Pow2ModWide kernel always assigns the required output elements so we don't need to worry about clearing these.
-            // remainderBuffer.MemSetToZero();
-
 		GpuUInt128 modulus = (GpuUInt128)prime;
-            kernel(stream, length, exponentBuffer.View, modulus, remainderBuffer.View);
+		gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+		int acceleratorIndex = gpu.AcceleratorIndex;
+		Accelerator accelerator = gpu.Accelerator;
 
+		AcceleratorStream? stream = AcceleratorStreamPool.Rent(acceleratorIndex);
+		MemoryBuffer1D<GpuUInt128, Stride1D.Dense> exponentBuffer;
+		MemoryBuffer1D<GpuUInt128, Stride1D.Dense> remainderBuffer;
+
+		lock(accelerator)
+		{
+			exponentBuffer = accelerator.Allocate1D<GpuUInt128>(length);
+			remainderBuffer = accelerator.Allocate1D<GpuUInt128>(length);
+		}
+
+		exponentBuffer.View.CopyFromCPU(stream, exponentSpan);
+		var kernelLauncher = gpu.Pow2ModWideKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
+
+		kernelLauncher(stream, length, exponentBuffer.View, modulus, remainderBuffer.View);
+		remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
 		stream.Synchronize();
 
-            Span<GpuUInt128> resultSpan = length <= WideStackThreshold
-                ? stackalloc GpuUInt128[length]
-                : new Span<GpuUInt128>(rentedResults = ArrayPool<GpuUInt128>.Shared.Rent(length), 0, length);
-
-            remainderBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(resultSpan), length);
-
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		for (int i = 0; i < length; i++)
 		{
 			results[i] = (UInt128)resultSpan[i];
 		}
 
-            exponentBuffer.Dispose();
 		remainderBuffer.Dispose();
-            return true;
-        }
-        catch (Exception)
-        {
-            Console.WriteLine($"Exception when computing {prime} with exponents: {string.Join(", ", exponents.ToArray())}");
-            throw;
-        }
-        finally
-        {
+		exponentBuffer.Dispose();
+		PrimeOrderCalculatorAccelerator.Return(gpu);
+
 		if (rentedExponents is not null)
 		{
-                ArrayPool<GpuUInt128>.Shared.Return(rentedExponents, clearArray: false);
-            }
-
-            if (rentedResults is not null)
-            {
-                ArrayPool<GpuUInt128>.Shared.Return(rentedResults, clearArray: false);
-            }
-
-            lease.Dispose();
+			gpuUInt128Pool.Return(rentedExponents, clearArray: false);
+			gpuUInt128Pool.Return(rentedResults!, clearArray: false);
 		}
 	}
 
@@ -662,18 +646,20 @@ internal static partial class PrimeOrderGpuHeuristics
 		return ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisorData, exponent);
 	}
 
-    private static Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> GetPow2ModKernel(Accelerator accelerator)
+	private static Kernel GetPow2ModKernel(int acceleratorIndex)
 	{
-        return Pow2ModKernelCache.GetOrAdd(accelerator, static accel =>
+		var pool = _pow2ModKernel ??= [];
+		if (pool[acceleratorIndex] is {} cached)
 		{
-            var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernel);
-            var kernel = KernelUtil.GetKernel(loaded);
-            return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
-        });
+			return cached;
 		}
 
+		var accelerator = _accelerators[acceleratorIndex];
+		var loaded = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernel));
 
-
+		pool[acceleratorIndex] = loaded;
+		return loaded;
+	}
 
 	internal readonly record struct PrimeOrderGpuCapability(int ModulusBits, int ExponentBits)
 	{

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);
+}