Various improvements to make sure scratch buffers don't override each others memory

Add GPU lease warm-up and shared limiter

EvenPerfectBitScanner.Benchmarks/PrimeTesterBenchmarks.cs

@@ -1,6 +1,3 @@
-using System;
-using System.Collections.Generic;
-using System.Threading;
 using BenchmarkDotNet.Attributes;
 using BenchmarkDotNet.Jobs;
 using Open.Numeric.Primes;
@@ -12,7 +9,6 @@ namespace EvenPerfectBitScanner.Benchmarks;
 [MemoryDiagnoser]
 public class PrimeTesterBenchmarks
 {
-    private static readonly HeuristicCombinedPrimeTester _heuristicCombinedTester = new();
     private static readonly HeuristicPrimeTester _heuristicTester = new();
     private ulong[] _candidates = [];
 
@@ -34,7 +30,7 @@ public class PrimeTesterBenchmarks
         _ = Prime.Numbers.IsPrime(_candidates[0]);
     }
 
-    [Benchmark(Baseline = true)]
+    [Benchmark]
     public int HeuristicCpu()
     {
         HeuristicPrimeTester tester = _heuristicTester;
@@ -104,18 +100,15 @@ 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 +117,7 @@ public class PrimeTesterBenchmarks
         return primeCount;
     }
 
-    [Benchmark]
+    [Benchmark(Baseline = true)]
     public int NonHeuristicCpu()
     {
         ulong[] values = _candidates;

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;
@@ -58,14 +58,15 @@ internal static class Program
 			bool startPrimeProvided = _cliArguments.StartPrimeProvided;
 			int threadCount = Math.Max(1, _cliArguments.ThreadCount);
 			int gpuPrimeBatch = Math.Max(1, _cliArguments.GpuPrimeBatch);
+			int gpuPrimeThreads = Math.Max(1, _cliArguments.GpuPrimeThreads);
 			UnboundedTaskScheduler.ConfigureThreadCount(threadCount);
 			PrimeTester.GpuBatchSize = gpuPrimeBatch;
+			GpuPrimeWorkLimiter.SetLimit(gpuPrimeThreads);
 			GpuContextPool.WarmUpPool(threadCount);
-			PrimeTester.WarmUpGpuKernels(threadCount);
+			PrimeTester.WarmUpGpuKernels(gpuPrimeThreads);
 			Console.WriteLine("Starting up threads...");
 			_ = UnboundedTaskScheduler.Instance;
 			int blockSize = Math.Max(1, _cliArguments.BlockSize);
-			int gpuPrimeThreads = Math.Max(1, _cliArguments.GpuPrimeThreads);
 			GpuKernelType kernelType = _cliArguments.KernelType;
 			bool useBitTransform = _cliArguments.UseBitTransform;
 			bool useOrder = _cliArguments.UseOrder;
@@ -117,9 +118,6 @@ internal static class Program
 				return;
 			}
 
-			// Apply GPU prime sieve runtime configuration
-			GpuPrimeWorkLimiter.SetLimit(gpuPrimeThreads);
-			PrimeTester.GpuBatchSize = gpuPrimeBatch;
 
 			MersenneDivisorCycles mersenneDivisorCycles = new();
 			if (!File.Exists(cyclesPath) || continueCyclesGeneration)
@@ -178,7 +176,7 @@ 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)
@@ -628,7 +626,7 @@ internal static class Program
 		{
 			if (!(_runPrimesOnCpu
 					? PrimeTester.IsPrime(p)
-					: PrimeTesters.Value!.IsPrimeGpu(p)))
+					: PrimeTester.IsPrimeGpu(p)))
 			{
 				_lastCompositeP = true;
 				return false;

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/PrimeTesterGpuTests.cs

@@ -16,7 +16,7 @@ public class PrimeTesterGpuTests
             ulong[] primes = [31UL, 61UL, 89UL, 107UL, 127UL, 521UL];
             foreach (ulong prime in primes)
             {
-                tester.IsPrimeGpu(prime).Should().BeTrue();
+				PrimeTester.IsPrimeGpu(prime).Should().BeTrue();
             }
         }
         finally
@@ -34,7 +34,7 @@ public class PrimeTesterGpuTests
             ulong[] composites = [33UL, 39UL, 51UL, 77UL, 91UL];
             foreach (ulong composite in composites)
             {
-                tester.IsPrimeGpu(composite).Should().BeFalse();
+				PrimeTester.IsPrimeGpu(composite).Should().BeFalse();
             }
         }
         finally

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/Gpu/DivisorGeneratorGpuCache.cs

@@ -1,142 +0,0 @@
-using System;
-using System.Collections.Concurrent;
-using System.Runtime.CompilerServices;
-using ILGPU;
-using ILGPU.Runtime;
-
-namespace PerfectNumbers.Core.Gpu;
-
-internal static class DivisorGeneratorGpuCache
-{
-        private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
-        {
-                public static AcceleratorReferenceComparer Instance { get; } = new();
-
-                public bool Equals(Accelerator? x, Accelerator? y)
-                {
-                        return ReferenceEquals(x, y);
-                }
-
-                public int GetHashCode(Accelerator obj)
-                {
-                        return RuntimeHelpers.GetHashCode(obj);
-                }
-        }
-
-        private sealed class DivisorGeneratorGpuTables : IDisposable
-        {
-                internal DivisorGeneratorGpuTables(Accelerator accelerator)
-                {
-                        SmallPrimes = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimes);
-                        SmallPrimesPow2 = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2);
-                        SmallPrimesLastOne = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastOne);
-                        SmallPrimesPow2LastOne = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastOne);
-                        SmallPrimesLastSeven = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastSeven);
-                        SmallPrimesPow2LastSeven = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastSeven);
-                        SmallPrimesLastThree = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastThree);
-                        SmallPrimesPow2LastThree = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastThree);
-                        SmallPrimesLastNine = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastNine);
-                        SmallPrimesPow2LastNine = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastNine);
-                        SmallPrimesLastOneWithoutLastThree = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastOneWithoutLastThree);
-                        SmallPrimesPow2LastOneWithoutLastThree = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastOneWithoutLastThree);
-                        SmallPrimesLastSevenWithoutLastThree = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastSevenWithoutLastThree);
-                        SmallPrimesPow2LastSevenWithoutLastThree = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastSevenWithoutLastThree);
-                        SmallPrimesLastNineWithoutLastThree = AllocateAndCopyUInt(accelerator, DivisorGenerator.SmallPrimesLastNineWithoutLastThree);
-                        SmallPrimesPow2LastNineWithoutLastThree = AllocateAndCopyULong(accelerator, DivisorGenerator.SmallPrimesPow2LastNineWithoutLastThree);
-                }
-
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimes { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2 { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastOne { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastOne { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastSeven { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastSeven { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastThree { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastThree { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastNine { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastNine { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastOneWithoutLastThree { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastOneWithoutLastThree { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastSevenWithoutLastThree { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastSevenWithoutLastThree { get; }
-                internal MemoryBuffer1D<uint, Stride1D.Dense> SmallPrimesLastNineWithoutLastThree { get; }
-                internal MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimesPow2LastNineWithoutLastThree { get; }
-
-                public void Dispose()
-                {
-                        SmallPrimes.Dispose();
-                        SmallPrimesPow2.Dispose();
-                        SmallPrimesLastOne.Dispose();
-                        SmallPrimesPow2LastOne.Dispose();
-                        SmallPrimesLastSeven.Dispose();
-                        SmallPrimesPow2LastSeven.Dispose();
-                        SmallPrimesLastThree.Dispose();
-                        SmallPrimesPow2LastThree.Dispose();
-                        SmallPrimesLastNine.Dispose();
-                        SmallPrimesPow2LastNine.Dispose();
-                        SmallPrimesLastOneWithoutLastThree.Dispose();
-                        SmallPrimesPow2LastOneWithoutLastThree.Dispose();
-                        SmallPrimesLastSevenWithoutLastThree.Dispose();
-                        SmallPrimesPow2LastSevenWithoutLastThree.Dispose();
-                        SmallPrimesLastNineWithoutLastThree.Dispose();
-                        SmallPrimesPow2LastNineWithoutLastThree.Dispose();
-                }
-
-                private static MemoryBuffer1D<uint, Stride1D.Dense> AllocateAndCopyUInt(Accelerator accelerator, uint[] source)
-                {
-                        var buffer = accelerator.Allocate1D<uint>(source.Length);
-                        if (source.Length > 0)
-                        {
-                                buffer.View.CopyFromCPU(source);
-                        }
-
-                        return buffer;
-                }
-
-                private static MemoryBuffer1D<ulong, Stride1D.Dense> AllocateAndCopyULong(Accelerator accelerator, ulong[] source)
-                {
-                        var buffer = accelerator.Allocate1D<ulong>(source.Length);
-                        if (source.Length > 0)
-                        {
-                                buffer.View.CopyFromCPU(source);
-                        }
-
-                        return buffer;
-                }
-        }
-
-        private static readonly ConcurrentDictionary<Accelerator, DivisorGeneratorGpuTables> Cache = new(AcceleratorReferenceComparer.Instance);
-
-        public static void WarmUpByDivisorTables(Accelerator accelerator)
-        {
-                if (accelerator is null || accelerator.AcceleratorType == AcceleratorType.CPU)
-                {
-                        return;
-                }
-
-                Cache.GetOrAdd(accelerator, acc => new DivisorGeneratorGpuTables(acc));
-        }
-
-        public static void Clear(Accelerator accelerator)
-        {
-                if (accelerator is null)
-                {
-                        return;
-                }
-
-                if (Cache.TryRemove(accelerator, out var tables))
-                {
-                        tables.Dispose();
-                }
-        }
-
-        public static void DisposeAll()
-        {
-                foreach (var entry in Cache)
-                {
-                        entry.Value.Dispose();
-                }
-
-                Cache.Clear();
-        }
-}

PerfectNumbers.Core/Gpu/GpuContextPool.cs

@@ -1,4 +1,3 @@
-using System;
 using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
@@ -12,14 +11,15 @@ public static class GpuContextPool
 	// Default device preference for generic GPU kernels (prime scans, NTT, etc.)
 	internal sealed class PooledContext
 	{
-		public Context Context { get; }
+		public readonly Context Context;
-		public Accelerator Accelerator { get; }
+		public readonly Accelerator Accelerator;
 
-		public PooledContext()
+		public PooledContext(KernelContainer kernels)
 		{
-			Context = Context.CreateDefault();
+			Context = SharedGpuContext.Context;
-			Accelerator = Context.GetPreferredDevice(false).CreateAccelerator(Context);
+			Accelerator = SharedGpuContext.Accelerator;
-			DivisorGeneratorGpuCache.WarmUpByDivisorTables(Accelerator);
+			GpuStaticTableInitializer.EnsureStaticTables(kernels, Accelerator);
+			kernels.Dispose();
 			// 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
@@ -33,11 +33,11 @@ public static class GpuContextPool
 			// Release PrimeTester GPU state for this accelerator (kernel/device primes)
 			// Clear any per-accelerator cached resources to avoid releasing
 			// after the accelerator is destroyed.
-			DivisorGeneratorGpuCache.Clear(Accelerator);
 			PrimeTester.ClearGpuCaches(Accelerator);
 
-			Accelerator.Dispose();
+			// These resources are shared between GPU leases
-			Context.Dispose();
+			//  Accelerator.Dispose();
+			// 	Context.Dispose();
 		}
 	}
 
@@ -53,9 +53,10 @@ public static class GpuContextPool
 		}
 
 		// Create a new accelerator when the pool does not have one available.
-		return new GpuContextLease(new PooledContext());
+		return new GpuContextLease(new PooledContext(GpuKernelPool.GetKernels()));
 	}
 
+
 	public static void WarmUpPool(int threadCount)
 	{
 		if (threadCount <= WarmedGpuContextCount)
@@ -64,15 +65,16 @@ public static class GpuContextPool
 		}
 
 		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++)
 		{
-			contexts[i] = new PooledContext();
+			KernelContainer kernels = GpuKernelPool.GetKernels();
-		}
+			context = new PooledContext(kernels);
-
+			contexts[i] = context;
-		for (int i = 0; i < toCreate; i++)
+			pool.Enqueue(context);
-		{
-			GpuPool.Enqueue(contexts[i]);
 		}
 
 		WarmedGpuContextCount = threadCount;
@@ -85,7 +87,6 @@ public static class GpuContextPool
 			gpu.Dispose();
 		}
 
-		DivisorGeneratorGpuCache.DisposeAll();
 		PrimeTester.DisposeGpuContexts();
 		WarmedGpuContextCount = 0;
 	}
@@ -96,21 +97,30 @@ public static class GpuContextPool
 		GpuPool.Enqueue(ctx);
 	}
 
-	public readonly struct GpuContextLease
+	public struct GpuContextLease
 	{
-		private readonly PooledContext _ctx;
+		private PooledContext? _ctx;
+		private bool _disposed;
 
 		internal GpuContextLease(PooledContext ctx)
 		{
 			_ctx = ctx;
+			_disposed = false;
 		}
 
-		public readonly Context Context => _ctx.Context;
+		public Context Context => _ctx!.Context;
-		public readonly Accelerator Accelerator => _ctx.Accelerator;
+		public Accelerator Accelerator => _ctx!.Accelerator;
 
-		public readonly void Dispose()
+		public void Dispose()
 		{
-			Return(_ctx);
+			if (_disposed)
+			{
+				return;
+			}
+
+			_disposed = true;
+			Return(_ctx!);
+			_ctx = null;
 		}
 	}
 }

PerfectNumbers.Core/Gpu/GpuKernelLease.cs

@@ -2,6 +2,7 @@ using System;
 using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
+using PerfectNumbers.Core;
 using static PerfectNumbers.Core.Gpu.GpuContextPool;
 
 namespace PerfectNumbers.Core.Gpu;
@@ -10,25 +11,24 @@ public sealed class GpuKernelLease
 {
     private static readonly ConcurrentQueue<GpuKernelLease> Pool = new();
 
-    private Action? _releaseLimiter;
     private GpuContextLease? _gpu;
-    private KernelContainer? _kernels;
+    public KernelContainer Kernels;
     private AcceleratorStream? _stream;
 
-    private GpuKernelLease()
+    private GpuKernelLease(KernelContainer kernels)
 	{
+		Kernels = kernels;
 	}
 
-    internal static GpuKernelLease Rent(GpuPrimeWorkLimiter.Lease limiter, GpuContextLease gpu, KernelContainer kernels)
+    internal static GpuKernelLease Rent(GpuContextLease gpu, KernelContainer kernels)
     {
         if (!Pool.TryDequeue(out var lease))
         {
-            lease = new GpuKernelLease();
+            lease = new GpuKernelLease(kernels);
         }
 
-        lease._releaseLimiter = limiter.Dispose;
         lease._gpu = gpu;
-        lease._kernels = kernels;
+        lease.Kernels = kernels;
         lease._stream = null;
         return lease;
     }
@@ -56,7 +56,7 @@ public sealed class GpuKernelLease
         get
         {
             var accel = Accelerator; // avoid capturing 'this' in lambda
-            return KernelContainer.InitOnce(ref _kernels!.Order, () =>
+            return KernelContainer.InitOnce(ref Kernels!.Order, () =>
             {
                 var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
                 var kernel = KernelUtil.GetKernel(loaded);
@@ -70,7 +70,7 @@ public sealed class GpuKernelLease
         get
         {
             var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.Pow2Mod, () =>
+            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);
@@ -84,7 +84,7 @@ public sealed class GpuKernelLease
         get
         {
             var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.Incremental, () =>
+            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);
@@ -98,7 +98,7 @@ public sealed class GpuKernelLease
         get
         {
             var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.IncrementalOrder, () =>
+            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);
@@ -112,7 +112,7 @@ public sealed class GpuKernelLease
         get
         {
             var accel = Accelerator;
-            return KernelContainer.InitOnce(ref _kernels!.Pow2ModOrder, () =>
+            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);
@@ -121,12 +121,26 @@ public sealed class GpuKernelLease
         }
     }
 
+    public Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ushort, Stride1D.Dense>> SpecialMaxKernel
+    {
+        get
+        {
+            var accel = Accelerator;
+            return KernelContainer.InitOnce(ref Kernels!.SpecialMax, () =>
+            {
+                var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ushort, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel);
+                var kernel = KernelUtil.GetKernel(loaded);
+                return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ushort, 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, () =>
+            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);
@@ -143,10 +157,8 @@ public sealed class GpuKernelLease
         _gpu?.Dispose();
         _gpu = null;
 
-        _releaseLimiter?.Invoke();
+		Kernels?.Dispose();
-        _releaseLimiter = null;
+		GpuPrimeWorkLimiter.Release();
-
-        _kernels = null;
         Pool.Enqueue(this);
     }
 

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -1,3 +1,4 @@
+using PerfectNumbers.Core;
 using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
@@ -37,6 +38,7 @@ public sealed class KernelContainer
 	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;
+	public Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ushort, Stride1D.Dense>>? SpecialMax;
 
 	// Optional device buffer with small divisor cycles (<= 4M). Index = divisor, value = cycle length.
 	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallCycles;
@@ -46,6 +48,9 @@ public sealed class KernelContainer
 	public MemoryBuffer1D<int, Stride1D.Dense>? SmallPrimeFactorExponentSlots;
 	public MemoryBuffer1D<int, Stride1D.Dense>? SmallPrimeFactorCountSlot;
 	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorRemainingSlot;
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? SpecialMaxFactors;
+	public MemoryBuffer1D<ulong, Stride1D.Dense>? SpecialMaxCandidates;
+	public MemoryBuffer1D<ushort, Stride1D.Dense>? SpecialMaxResult;
 	public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastOne;
 	public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimesPow2LastOne;
 	public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastSeven;
@@ -67,6 +72,8 @@ public sealed class KernelContainer
 		}
 		return current;
 	}
+
+	public void Dispose() => GpuKernelPool.Return(this);
 }
 
 public readonly struct ResiduePrimeViews(
@@ -81,44 +88,29 @@ public readonly struct ResiduePrimeViews(
 	public readonly ArrayView1D<ulong, Stride1D.Dense> LastSevenPow2 = lastSevenPow2;
 }
 
-public readonly struct SmallPrimeFactorTables
+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;
+	public readonly MemoryBuffer1D<uint, Stride1D.Dense> Primes = primes;
-        Squares = squares;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Squares = squares;
-        Count = count;
+	public readonly int Count = count;
-    }
 
 	public ArrayView1D<uint, Stride1D.Dense> PrimesView => Primes.View;
 	public ArrayView1D<ulong, Stride1D.Dense> SquaresView => Squares.View;
 }
 
-public readonly struct SmallPrimeFactorScratch
+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;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> PrimeSlots = primeSlots;
-        ExponentSlots = exponentSlots;
+	public readonly MemoryBuffer1D<int, Stride1D.Dense> ExponentSlots = exponentSlots;
-        CountSlot = countSlot;
+	public readonly MemoryBuffer1D<int, Stride1D.Dense> CountSlot = countSlot;
-        RemainingSlot = remainingSlot;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> RemainingSlot = remainingSlot;
-    }
 
 	public void Clear()
 	{
@@ -129,44 +121,55 @@ public readonly struct SmallPrimeFactorScratch
 	}
 }
 
+public readonly struct SpecialMaxScratch(
+	MemoryBuffer1D<ulong, Stride1D.Dense> factorValues,
+	MemoryBuffer1D<ulong, Stride1D.Dense> candidateValues,
+	MemoryBuffer1D<ushort, Stride1D.Dense> resultSlot)
+{
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> FactorValues = factorValues;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> CandidateValues = candidateValues;
+	public readonly MemoryBuffer1D<ushort, Stride1D.Dense> ResultSlot = resultSlot;
+
+	public ArrayView1D<ulong, Stride1D.Dense> FactorsView => FactorValues.View;
+	public ArrayView1D<ulong, Stride1D.Dense> CandidatesView => CandidateValues.View;
+	public ArrayView1D<ushort, Stride1D.Dense> ResultView => ResultSlot.View;
+}
+
 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.
+	private static readonly ConcurrentQueue<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.
 
-    private static KernelContainer GetKernels(Accelerator accelerator)
+	public static KernelContainer GetKernels() => KernelCache.TryDequeue(out var container)
+		? container
+		: new KernelContainer();
+
+	public static void Return(KernelContainer container) => KernelCache.Enqueue(container);
+
+	internal static void PreloadStaticTables(KernelContainer kernels, Accelerator accelerator)
 	{
-        return KernelCache.GetOrAdd(accelerator, _ => new KernelContainer());
+		EnsureSmallCyclesOnDevice(kernels, accelerator);
+		EnsureSmallPrimesOnDevice(kernels, accelerator);
+		EnsureSmallPrimeFactorTables(kernels, accelerator);
 	}
 
 	// 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)
+	public static ArrayView1D<ulong, Stride1D.Dense> EnsureSmallCyclesOnDevice(KernelContainer kernels, Accelerator accelerator)
 	{
-        var kernels = GetKernels(accelerator);
 		if (kernels.SmallCycles is { } buffer)
 		{
 			return buffer.View;
 		}
 
-        // Ensure single upload per accelerator even if multiple threads race here.
-        lock (kernels) // TODO: Remove this lock by pre-uploading the immutable small-cycle snapshot during initialization; once no mutation happens at runtime, the pool must expose a simple reference without synchronization.
-        {
-            if (kernels.SmallCycles is { } existing)
-            {
-                return existing.View;
-            }
-
 		var host = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot(); // TODO: Preload this device buffer during startup and keep it immutable so we can delete the lock above in favor of the preloaded snapshot.
 		var device = accelerator.Allocate1D<ulong>(host.Length);
 		device.View.CopyFromCPU(host);
 		kernels.SmallCycles = device;
 		return device.View;
 	}
-    }
 
-    public static ResiduePrimeViews EnsureSmallPrimesOnDevice(Accelerator accelerator)
+	public static ResiduePrimeViews EnsureSmallPrimesOnDevice(KernelContainer kernels, Accelerator accelerator)
 	{
-        var kernels = GetKernels(accelerator);
 		if (kernels.SmallPrimesLastOne is { } lastOne &&
 			kernels.SmallPrimesLastSeven is { } lastSeven &&
 			kernels.SmallPrimesPow2LastOne is { } lastOnePow2 &&
@@ -175,8 +178,6 @@ public class GpuKernelPool
 			return new ResiduePrimeViews(lastOne.View, lastSeven.View, lastOnePow2.View, lastSevenPow2.View);
 		}
 
-        lock (kernels) // TODO: Inline these small-prime uploads into startup initialization alongside the small-cycle snapshot so we can drop runtime locking and keep the GPU pool free of synchronization.
-        {
 		if (kernels.SmallPrimesLastOne is { } existingLastOne &&
 			kernels.SmallPrimesLastSeven is { } existingLastSeven &&
 			kernels.SmallPrimesPow2LastOne is { } existingLastOnePow2 &&
@@ -206,18 +207,14 @@ public class GpuKernelPool
 
 		return new ResiduePrimeViews(deviceLastOne.View, deviceLastSeven.View, deviceLastOnePow2.View, deviceLastSevenPow2.View);
 	}
-    }
 
-    public static SmallPrimeFactorTables EnsureSmallPrimeFactorTables(Accelerator accelerator)
+	public static SmallPrimeFactorTables EnsureSmallPrimeFactorTables(KernelContainer kernels, 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);
@@ -236,12 +233,8 @@ public class GpuKernelPool
 
 		return new SmallPrimeFactorTables(devicePrimes, deviceSquares, hostPrimes.Length);
 	}
-    }
 
-    public static SmallPrimeFactorScratch EnsureSmallPrimeFactorScratch(Accelerator accelerator, int slotCount)
+	public static SmallPrimeFactorScratch EnsureSmallPrimeFactorScratch(KernelContainer kernels, Accelerator accelerator, int slotCount)
-    {
-        var kernels = GetKernels(accelerator);
-        lock (kernels)
 	{
 		MemoryBuffer1D<ulong, Stride1D.Dense>? primeSlots = kernels.SmallPrimeFactorPrimeSlots;
 		if (primeSlots is null || primeSlots.Length < slotCount)
@@ -275,15 +268,41 @@ public class GpuKernelPool
 
 		return new SmallPrimeFactorScratch(primeSlots!, exponentSlots!, countSlot!, remainingSlot!);
 	}
+
+	public static SpecialMaxScratch EnsureSpecialMaxScratch(KernelContainer kernels, Accelerator accelerator, int factorCapacity)
+	{
+		MemoryBuffer1D<ulong, Stride1D.Dense>? factorValues = kernels.SpecialMaxFactors;
+		if (factorValues is null || factorValues.Length < factorCapacity)
+		{
+			factorValues?.Dispose();
+			factorValues = accelerator.Allocate1D<ulong>(factorCapacity);
+			kernels.SpecialMaxFactors = factorValues;
+		}
+
+		MemoryBuffer1D<ulong, Stride1D.Dense>? candidateValues = kernels.SpecialMaxCandidates;
+		if (candidateValues is null || candidateValues.Length < factorCapacity)
+		{
+			candidateValues?.Dispose();
+			candidateValues = accelerator.Allocate1D<ulong>(factorCapacity);
+			kernels.SpecialMaxCandidates = candidateValues;
+		}
+
+		MemoryBuffer1D<ushort, Stride1D.Dense>? resultSlot = kernels.SpecialMaxResult;
+		if (resultSlot is null)
+		{
+			resultSlot = accelerator.Allocate1D<ushort>(1);
+			kernels.SpecialMaxResult = resultSlot;
+		}
+
+		return new SpecialMaxScratch(factorValues!, candidateValues!, resultSlot!);
 	}
 
 	public static GpuKernelLease GetKernel(bool useGpuOrder)
 	{
-        var limiter = GpuPrimeWorkLimiter.Acquire();
+		GpuPrimeWorkLimiter.Acquire();
 		var gpu = Rent();
-        var accelerator = gpu.Accelerator;
+		var kernels = GetKernels();
-        var kernels = GetKernels(accelerator);
+		return GpuKernelLease.Rent(gpu, kernels);
-        return GpuKernelLease.Rent(limiter, gpu, kernels);
 	}
 
 	/// <summary>

PerfectNumbers.Core/Gpu/GpuPrimeWorkLimiter.cs

@@ -1,3 +1,6 @@
+using System;
+using System.Threading;
+
 namespace PerfectNumbers.Core.Gpu;
 
 public static class GpuPrimeWorkLimiter
@@ -5,17 +8,16 @@ public static class GpuPrimeWorkLimiter
     private static SemaphoreSlim _semaphore = new(1, int.MaxValue);
     private static int _currentLimit = 1;
 
-    public static Lease Acquire()
+	public static void Acquire() => _semaphore.Wait();
-    {
+	public static void Release() => _semaphore.Release();
-        // 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 SetLimit(int value)
     {
+        if (value <= 0)
+        {
+            value = 1;
+        }
+
         if (value == _currentLimit)
         {
             return;
@@ -23,23 +25,9 @@ public static class GpuPrimeWorkLimiter
 
         // TODO: Switch to a shared limiter implementation with GpuWorkLimiter so we can coordinate CPU/GPU prime work using
         // the same semaphore pool without reallocating per adjustment.
+        var previous = _semaphore;
         _semaphore = new SemaphoreSlim(value, value);
         _currentLimit = value;
-    }
+        previous.Dispose();
-
-    public sealed class Lease
-    {
-        private readonly SemaphoreSlim _sem;
-
-        internal Lease(SemaphoreSlim sem)
-        {
-            _sem = sem;
-        }
-
-        public void Dispose()
-        {
-            // _sem.Release();
     }
 }
-}
-

PerfectNumbers.Core/Gpu/GpuStaticTableInitializer.cs

@@ -0,0 +1,39 @@
+using System.Collections.Concurrent;
+using System.Collections.Generic;
+using System.Runtime.CompilerServices;
+using ILGPU;
+using ILGPU.Runtime;
+using PerfectNumbers.Core;
+
+namespace PerfectNumbers.Core.Gpu;
+
+internal static class GpuStaticTableInitializer
+{
+    private static readonly ConcurrentDictionary<Accelerator, byte> Initialized = new(AcceleratorReferenceComparer.Instance);
+
+    internal static void EnsureStaticTables(KernelContainer kernels, Accelerator accelerator)
+    {
+        Initialized.GetOrAdd(accelerator, acc =>
+        {
+            PrimeTester.PrimeTesterGpuContextPool.EnsureStaticTables(kernels, acc);
+            GpuKernelPool.PreloadStaticTables(kernels, acc);
+            PrimeOrderGpuHeuristics.PreloadStaticTables( acc);
+            return 0;
+        });
+    }
+
+    private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
+    {
+        internal static AcceleratorReferenceComparer Instance { get; } = new();
+
+        public bool Equals(Accelerator? x, Accelerator? y)
+        {
+            return ReferenceEquals(x, y);
+        }
+
+        public int GetHashCode(Accelerator obj)
+        {
+            return RuntimeHelpers.GetHashCode(obj);
+        }
+    }
+}

PerfectNumbers.Core/Gpu/HeuristicGpuDivisorTables.cs

@@ -8,73 +8,57 @@ internal readonly struct HeuristicGpuDivisorTables
 {
         public static readonly HeuristicGpuDivisorTables Shared = default;
 
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisors;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding1;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding7;
-        public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding9;
-
-        public HeuristicGpuDivisorTables(
-                ArrayView1D<ulong, Stride1D.Dense> groupADivisors,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding1,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding7,
-                ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding9)
-        {
-                GroupADivisors = groupADivisors;
-                GroupBDivisorsEnding1 = groupBDivisorsEnding1;
-                GroupBDivisorsEnding7 = groupBDivisorsEnding7;
-                GroupBDivisorsEnding9 = groupBDivisorsEnding9;
-        }
-
-        [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> SelectDivisors(HeuristicGpuDivisorTableKind kind, byte ending)
-        {
-                return kind switch
-                {
-                        HeuristicGpuDivisorTableKind.GroupA => GroupADivisors,
-                        HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorsEnding1,
-                        HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorsEnding7,
-                        HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorsEnding9,
-                        _ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
-                };
-        }
-
-        internal static void InitializeShared(in HeuristicGpuDivisorTables tables)
-        {
-                Unsafe.AsRef(in Shared) = tables;
-        }
-}
-
-internal readonly struct HeuristicGpuCombinedDivisorTables
-{
-        public static readonly HeuristicGpuCombinedDivisorTables 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 HeuristicGpuCombinedDivisorTables(
+        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> 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)]
@@ -90,7 +74,72 @@ internal readonly struct HeuristicGpuCombinedDivisorTables
                 };
         }
 
-        internal static void InitializeShared(in HeuristicGpuCombinedDivisorTables tables)
+        [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/Kernels/PrimeOrderSpecialMaxKernels.cs

@@ -0,0 +1,80 @@
+using ILGPU;
+using ILGPU.Runtime;
+using PerfectNumbers.Core;
+
+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<ushort, 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]))
+        {
+            resultOut[0] = 0;
+            return;
+        }
+
+        for (int i = 1; i < actual; i++)
+        {
+            if (stepper.ComputeNextIsUnityGpu(candidates[i]))
+            {
+                resultOut[0] = 0;
+                return;
+            }
+        }
+
+        resultOut[0] = 1;
+    }
+}

PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs

@@ -82,25 +82,30 @@ internal static class PrimeTesterKernels
         Index1D index,
         ArrayView<int> resultFlag,
         ulong n,
-        ulong maxDivisor,
+        ulong maxDivisorSquare,
         HeuristicGpuDivisorTableKind tableKind,
-        HeuristicGpuDivisorTables tables,
+        HeuristicGpuDivisorTables tables)
-        HeuristicGpuCombinedDivisorTables combinedTables)
     {
         byte nMod10 = (byte)(n % 10UL);
-        ArrayView1D<ulong, Stride1D.Dense> divisors = tableKind == HeuristicGpuDivisorTableKind.Combined
+        ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(tableKind, nMod10);
-            ? combinedTables.SelectCombined(nMod10)
+        ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(tableKind, nMod10);
-            : tables.SelectDivisors(tableKind, nMod10);
 
-        int divisorLength = (int)divisors.Length;
+        int divisorLength = (int)divisorSquares.Length;
         int threadIndex = index;
-        bool skipThread = threadIndex >= divisorLength;
+        if (threadIndex >= divisorLength)
-        ulong divisor = skipThread ? 1UL : divisors[threadIndex];
+        {
-        skipThread = skipThread | (divisor > maxDivisor);
+            return;
+        }
 
-        ulong remainder = skipThread ? 1UL : n % divisor;
+        ulong divisorSquare = divisorSquares[threadIndex];
-        bool hasFactor = !skipThread & (remainder == 0UL);
+        if (divisorSquare > maxDivisorSquare)
-        if (hasFactor)
+        {
+            return;
+        }
+
+        ulong divisor = divisors[threadIndex];
+        divisor = n % divisor;
+        if (divisor == 0UL)
         {
             resultFlag[0] = 1;
         }

PerfectNumbers.Core/Gpu/MersenneNumberIncrementalGpuTester.cs

@@ -31,11 +31,11 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
         ulong step3 = ((exponent % 3UL) << 1) % 3UL;
         ulong step5 = ((exponent % 5UL) << 1) % 5UL;
         GpuUInt128 twoPGpu = (GpuUInt128)twoP;
-        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(accelerator);
+        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(gpuLease.Kernels, accelerator);
         ResiduePrimeViews primeViews = default;
         if (_kernelType == GpuKernelType.Pow2Mod)
         {
-            primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(accelerator);
+            primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(gpuLease.Kernels, accelerator);
         }
 
         var orderBuffer = accelerator.Allocate1D<ulong>(batchSize);

PerfectNumbers.Core/Gpu/MersenneNumberLucasLehmerGpuTester.cs

@@ -54,13 +54,13 @@ public class MersenneNumberLucasLehmerGpuTester
             return false;
         }
 
-        var limiter = GpuPrimeWorkLimiter.Acquire();
+        GpuPrimeWorkLimiter.Acquire();
         bool result;
         if (exponent >= 128UL)
         {
             if (!TryGetNttParameters(exponent, out var nttMod, out var primitiveRoot))
             {
-                limiter.Dispose();
+				GpuPrimeWorkLimiter.Release();
                 throw new NotSupportedException("NTT parameters for the given exponent are not available.");
             }
 
@@ -80,7 +80,7 @@ public class MersenneNumberLucasLehmerGpuTester
             gpu.Dispose();
         }
 
-        limiter.Dispose();
+		GpuPrimeWorkLimiter.Release();
         return result;
     }
 

PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs

@@ -29,6 +29,8 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 		UInt128 remaining;
 		int currentSize;
 		int found = 0;
+
+		var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(gpuLease.Kernels, accelerator);
 		while (kStart <= maxK && Volatile.Read(ref isPrime))
 		{
 			remaining = maxK - kStart + 1UL;
@@ -44,7 +46,6 @@ 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);
 

PerfectNumbers.Core/Gpu/MersenneNumberResidueGpuTester.cs

@@ -26,7 +26,7 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
     // 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();
+        GpuPrimeWorkLimiter.Acquire();
         var gpuLease = RentAccelerator();
         var accelerator = gpuLease.Accelerator;
 
@@ -35,14 +35,15 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
         var stream = accelerator.CreateStream();
         var kernel = GetPow2ModKernel(accelerator);
 		var resources = RentResources(accelerator, GpuConstants.ScanBatchSize);
+		var gpuKernels = GpuKernelPool.GetKernels();
 
         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);
+        var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(gpuKernels, accelerator);
-        ResiduePrimeViews primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(accelerator);
+        ResiduePrimeViews primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(gpuKernels, accelerator);
 
         UInt128 kStart = 1UL;
         UInt128 limit = maxK + UInt128.One;
@@ -109,8 +110,9 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
         ReturnResources(accelerator, resources);
         stream.Dispose();
 		// Monitor.Exit(gpuLease.ExecutionLock);
+		gpuKernels.Dispose();
 		ReturnAccelerator(gpuLease);
-        limiter.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) =>

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -18,6 +18,11 @@ internal enum GpuPow2ModStatus
 
 internal static partial class PrimeOrderGpuHeuristics
 {
+    internal static void PreloadStaticTables(Accelerator accelerator)
+    {
+        _ = GetSmallPrimeDeviceCache(accelerator);
+    }
+
     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();

PerfectNumbers.Core/Gpu/SharedGpuContext.cs

@@ -0,0 +1,19 @@
+using System;
+using System.Threading;
+using ILGPU;
+using ILGPU.Runtime;
+
+namespace PerfectNumbers.Core.Gpu;
+
+internal static class SharedGpuContext
+{
+    private static readonly Lazy<Context> s_context = new(() => ILGPU.Context.CreateDefault(), LazyThreadSafetyMode.ExecutionAndPublication);
+    private static readonly Lazy<Device> s_device = new(() => s_context.Value.GetPreferredDevice(false), LazyThreadSafetyMode.ExecutionAndPublication);
+    private static readonly Lazy<Accelerator> s_accelerator = new(() => s_device.Value.CreateAccelerator(s_context.Value), LazyThreadSafetyMode.ExecutionAndPublication);
+
+    internal static Context Context => s_context.Value;
+
+    internal static Device Device => s_device.Value;
+
+    internal static Accelerator Accelerator => s_accelerator.Value;
+}

PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs

@@ -1,5 +1,4 @@
 using Open.Numeric.Primes;
-using System;
 using ILGPU.Runtime;
 using System.Runtime.CompilerServices;
 using PerfectNumbers.Core.Gpu;
@@ -225,14 +224,13 @@ public sealed class HeuristicCombinedPrimeTester
 
 	private static bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
 	{
-                var limiter = GpuPrimeWorkLimiter.Acquire();
+		GpuPrimeWorkLimiter.Acquire();
-                var gpu = PrimeTester.PrimeTesterGpuContextPool.Rent(nMod10, 1, PrimeTester.PrimeTesterGpuContextPool.HeuristicGpuTableScope.GroupABWithCombined);
+		var gpu = PrimeTester.PrimeTesterGpuContextPool.Rent(1);
 		var accelerator = gpu.Accelerator;
-                var kernel = gpu.HeuristicTrialDivisionKernel;
+		var kernel = PrimeTester.PrimeTesterGpuContextPool.PrimeTesterGpuContextLease.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
 		var flagView = flagView1D.AsContiguous();
 
-                ulong maxDivisor = ComputeMaxDivisorLimit(maxDivisorSquare);
 		bool compositeDetected = false;
 		int divisorLength = GetCombinedDivisors(nMod10).Length;
 
@@ -242,38 +240,20 @@ public sealed class HeuristicCombinedPrimeTester
 				divisorLength,
 				flagView,
 				n,
-                        maxDivisor,
+				maxDivisorSquare,
 				HeuristicGpuDivisorTableKind.Combined,
-                        gpu.HeuristicGpuTables,
+				gpu.HeuristicGpuTables);
-                        gpu.HeuristicCombinedTables);
 		accelerator.Synchronize();
 		flagView1D.CopyToCPU(ref compositeFlag, 1);
 		compositeDetected = compositeFlag != 0;
 
 		gpu.Dispose();
-                limiter.Dispose();
+		GpuPrimeWorkLimiter.Release();
 		return compositeDetected;
 	}
 
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
-        
-        private static ulong ComputeMaxDivisorLimit(ulong maxDivisorSquare)
-        {
-                ulong maxDivisor = (ulong)Math.Sqrt(maxDivisorSquare);
-
-                while ((maxDivisor + 1) != 0 && (maxDivisor + 1) <= maxDivisorSquare / (maxDivisor + 1))
-                {
-                        maxDivisor++;
-                }
-
-                while (maxDivisor != 0 && maxDivisor > maxDivisorSquare / maxDivisor)
-                {
-                        maxDivisor--;
-                }
-
-                return maxDivisor;
-        }
 	private static bool EvaluateWithOpenNumericFallback(ulong n)
 	{
 		return Prime.Numbers.IsPrime(n);

PerfectNumbers.Core/HeuristicPrimeTester.cs

@@ -240,14 +240,13 @@ public sealed class HeuristicPrimeTester
 
 	private bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
 	{
-                var limiter = GpuPrimeWorkLimiter.Acquire();
+		GpuPrimeWorkLimiter.Acquire();
-                var gpu = PrimeTester.PrimeTesterGpuContextPool.Rent(nMod10, 1, PrimeTester.PrimeTesterGpuContextPool.HeuristicGpuTableScope.GroupAB);
+		var gpu = PrimeTester.PrimeTesterGpuContextPool.Rent(1);
 		var accelerator = gpu.Accelerator;
-                var kernel = gpu.HeuristicTrialDivisionKernel;
+		var kernel = PrimeTester.PrimeTesterGpuContextPool.PrimeTesterGpuContextLease.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
 		var flagView = flagView1D.AsContiguous();
 
-                ulong maxDivisor = ComputeMaxDivisorLimit(maxDivisorSquare);
 		bool compositeDetected = false;
 		int compositeFlag = 0;
 
@@ -258,10 +257,9 @@ public sealed class HeuristicPrimeTester
 				groupALength,
 				flagView,
 				n,
-                        maxDivisor,
+				maxDivisorSquare,
 				HeuristicGpuDivisorTableKind.GroupA,
-                        gpu.HeuristicGpuTables,
+				gpu.HeuristicGpuTables);
-                        gpu.HeuristicCombinedTables);
 		accelerator.Synchronize();
 		flagView1D.CopyToCPU(ref compositeFlag, 1);
 		compositeDetected = compositeFlag != 0;
@@ -293,10 +291,9 @@ public sealed class HeuristicPrimeTester
 						divisorLength,
 						flagView,
 						n,
-                                        maxDivisor,
+						maxDivisorSquare,
 						tableKind,
-                                        gpu.HeuristicGpuTables,
+						gpu.HeuristicGpuTables);
-                        gpu.HeuristicCombinedTables);
 				accelerator.Synchronize();
 				flagView1D.CopyToCPU(ref compositeFlag, 1);
 				compositeDetected = compositeFlag != 0;
@@ -304,7 +301,7 @@ public sealed class HeuristicPrimeTester
 		}
 
 		gpu.Dispose();
-                limiter.Dispose();
+		GpuPrimeWorkLimiter.Release();
 		return compositeDetected;
 	}
 
@@ -315,23 +312,6 @@ public sealed class HeuristicPrimeTester
 	{
 		return Prime.Numbers.IsPrime(n);
 	}
-
-	private static ulong ComputeMaxDivisorLimit(ulong maxDivisorSquare)
-	{
-		ulong maxDivisor = (ulong)Math.Sqrt(maxDivisorSquare);
-
-		while ((maxDivisor + 1) != 0 && (maxDivisor + 1) <= maxDivisorSquare / (maxDivisor + 1))
-		{
-			maxDivisor++;
-		}
-
-		while (maxDivisor != 0 && maxDivisor > maxDivisorSquare / maxDivisor)
-		{
-			maxDivisor--;
-		}
-
-		return maxDivisor;
-	}
 	internal static ulong ComputeHeuristicDivisorSquareLimit(ulong n)
 	{
 		return n;

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -412,7 +412,7 @@ public class MersenneDivisorCycles
 		// Console.WriteLine($"MersenneDivisorCycles.TryFactorIntoCountsInternal hits {Volatile.Read(ref _tryFactorIntoCountsInternalHits)}");
 
 		// HeuristicPrimeTester primeTester = _primeTester ??= new();
-		if (PrimeTester.IsPrime(remaining))
+		if (PrimeTester.IsPrimeGpu(remaining))
 		// if (PrimeTester.IsPrimeCpu(remaining, CancellationToken.None))
 		{
 			AddFactor(counts, remaining);
@@ -943,7 +943,7 @@ public class MersenneDivisorCycles
 		// HeuristicPrimeTester primeTester = _primeTester ??= new();
 		// if (!skipPrimeOrderHeuristic && HeuristicPrimeTester.Exclusive.IsPrimeCpu(divisor, CancellationToken.None))
 		// if (!skipPrimeOrderHeuristic && primeTester.IsPrimeGpu(divisor))
-		if (!skipPrimeOrderHeuristic && PrimeTester.IsPrime(divisor))
+		if (!skipPrimeOrderHeuristic && PrimeTester.IsPrimeGpu(divisor))
 		{
 			ulong computedOrder = PrimeOrderCalculator.Calculate(
 					divisor,

PerfectNumbers.Core/PrimeOrderCalculator.Cpu.cs

@@ -1,4 +1,3 @@
-using System;
 using System.Buffers;
 using System.Diagnostics;
 using System.Numerics;
@@ -108,7 +107,7 @@ internal static partial class PrimeOrderCalculator
 		return candidateOrder;
 	}
 
-
+	// private static ulong specialMaxHits;
 	private static bool TrySpecialMaxCpu(ulong phi, ulong prime, PartialFactorResult factors, in MontgomeryDivisorData divisorData)
 	{
 		int length = factors.Count;
@@ -121,16 +120,16 @@ internal static partial class PrimeOrderCalculator
 
 		ReadOnlySpan<FactorEntry> factorSpan = new(factors.Factors, 0, length);
 
-		if (length <= 32)
-		{
 		Span<ulong> stackBuffer = stackalloc ulong[length];
+
+		if (length <= 7)
+		{
 			return EvaluateSpecialMaxCandidates(stackBuffer, factorSpan, phi, prime, divisorData);
 		}
 
-		ulong[] rented = ThreadStaticPools.UlongPool.Rent(length);
+		// Atomic.Add(ref specialMaxHits, 1);
-		bool result = EvaluateSpecialMaxCandidates(rented.AsSpan(0, length), factorSpan, phi, prime, divisorData);
+		// Console.WriteLine($"Special max GPU hit {specialMaxHits} ({length})");
-		ThreadStaticPools.UlongPool.Return(rented, clearArray: false);
+		return EvaluateSpecialMaxCandidatesGpu(stackBuffer, factorSpan, phi, prime, divisorData);
-		return result;
 	}
 
 	private static bool EvaluateSpecialMaxCandidates(Span<ulong> buffer, ReadOnlySpan<FactorEntry> factors, ulong phi, ulong prime, in MontgomeryDivisorData divisorData)
@@ -1194,7 +1193,7 @@ internal static partial class PrimeOrderCalculator
 				// Console.WriteLine($"Partial factor pending hits {Volatile.Read(ref _partialFactorPendingHits)}");
 
 				// bool isPrime = PrimeTester.IsPrimeCpu(composite, CancellationToken.None);
-				bool isPrime = PrimeTester.IsPrime(composite);
+				bool isPrime = PrimeTester.IsPrimeGpu(composite);
 				// bool isPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(composite);
 
 				entry = entry.WithPrimality(isPrime);
@@ -1231,7 +1230,7 @@ internal static partial class PrimeOrderCalculator
 			// Atomic.Add(ref _partialFactorCofactorHits, 1UL);
 			// Console.WriteLine($"Partial factor cofactor hits {Volatile.Read(ref _partialFactorCofactorHits)}");
 
-			cofactorIsPrime = PrimeTester.IsPrime(cofactor);
+			cofactorIsPrime = PrimeTester.IsPrimeGpu(cofactor);
 			// cofactorIsPrime = PrimeTester.IsPrimeGpu(cofactor);
 			// cofactorIsPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(cofactor);
 		}
@@ -1701,7 +1700,7 @@ internal static partial class PrimeOrderCalculator
 		// HeuristicPrimeTester tester = _tester ??= new();
 
 		// bool isPrime = primeTester.HeuristicIsPrimeGpu(value);
-		bool isPrime = PrimeTester.IsPrime(value);
+		bool isPrime = PrimeTester.IsPrimeGpu(value);
 		// bool isPrime = PrimeTester.IsPrimeGpu(value);
 		// bool isPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(value);
 		if (!knownComposite && isPrime)
@@ -1717,7 +1716,7 @@ internal static partial class PrimeOrderCalculator
 		// Console.WriteLine($"Factor completely after PollardRho hits {Volatile.Read(ref _factorCompletelyPollardRhoHits)}");
 
 		// isPrime = primeTester.HeuristicIsPrimeGpu(factor);
-		isPrime = PrimeTester.IsPrime(factor);
+		isPrime = PrimeTester.IsPrimeGpu(factor);
 		// isPrime = PrimeTester.IsPrimeGpu(factor);
 		// isPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(factor);
 		if (isPrime)
@@ -1754,7 +1753,7 @@ internal static partial class PrimeOrderCalculator
 			// Atomic.Add(ref _partialFactorCofactorHits, 1UL);
 			// Console.WriteLine($"Partial factor cofactor hits {Volatile.Read(ref _partialFactorCofactorHits)}");
 
-			isPrime = PrimeTester.IsPrime(quotient);
+			isPrime = PrimeTester.IsPrimeGpu(quotient);
 			// isPrime = PrimeTester.IsPrimeGpu(quotient);
 			// isPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(quotient);
 		}

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -25,8 +25,8 @@ internal static partial class PrimeOrderCalculator
 		Accelerator accelerator = lease.Accelerator;
 		AcceleratorStream stream = lease.Stream;
 
-                SmallPrimeFactorTables tables = GpuKernelPool.EnsureSmallPrimeFactorTables(accelerator);
+		SmallPrimeFactorTables tables = GpuKernelPool.EnsureSmallPrimeFactorTables(lease.Kernels, accelerator);
-                SmallPrimeFactorScratch scratch = GpuKernelPool.EnsureSmallPrimeFactorScratch(accelerator, GpuSmallPrimeFactorSlots);
+		SmallPrimeFactorScratch scratch = GpuKernelPool.EnsureSmallPrimeFactorScratch(lease.Kernels, accelerator, GpuSmallPrimeFactorSlots);
 
 		var kernel = lease.SmallPrimeFactorKernel;
 		kernel(
@@ -68,4 +68,52 @@ internal static partial class PrimeOrderCalculator
 		ThreadStaticPools.IntPool.Return(exponentBufferArray);
 		return true;
 	}
+
+	private static bool EvaluateSpecialMaxCandidatesGpu(
+			Span<ulong> buffer,
+			ReadOnlySpan<FactorEntry> factors,
+			ulong phi,
+			ulong prime,
+			in MontgomeryDivisorData divisorData)
+	{
+		if (factors.Length == 0)
+		{
+			return true;
 		}
+
+		int factorCount = factors.Length;
+		Span<ulong> factorSpan = buffer[..factorCount];
+		for (int i = 0; i < factorCount; i++)
+		{
+			factorSpan[i] = factors[i].Value;
+		}
+
+		var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
+		Accelerator accelerator = lease.Accelerator;
+		AcceleratorStream stream = lease.Stream;
+
+		SpecialMaxScratch scratch = GpuKernelPool.EnsureSpecialMaxScratch(lease.Kernels, accelerator, factorCapacity: factorCount);
+
+		scratch.FactorsView.SubView(0, factorCount).CopyFromCPU(ref MemoryMarshal.GetReference(factorSpan), factorCount);
+
+		var kernel = lease.SpecialMaxKernel;
+		kernel(
+				stream,
+				1,
+				phi,
+				scratch.FactorsView,
+				factorCount,
+				divisorData,
+				scratch.CandidatesView,
+				scratch.ResultView);
+
+		stream.Synchronize();
+
+		ushort result = 0;
+		scratch.ResultView.CopyToCPU(ref result, 1);
+		lease.Dispose();
+
+		return result != 0;
+	}
+}
+

PerfectNumbers.Core/PrimeTester.cs

@@ -93,50 +93,43 @@ public sealed class PrimeTester
 	}
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
-	public bool IsPrimeGpu(ulong n)
+	public static bool IsPrimeGpu(ulong n)
 	{
-		// bool belowGpuRange = n < 31UL;
+		GpuPrimeWorkLimiter.Acquire();
+		var gpu = PrimeTesterGpuContextPool.Rent(1);
+		var accelerator = gpu.Accelerator;
+		var kernel = PrimeTesterGpuContextPool.PrimeTesterGpuContextLease.Kernel;
 
-		// if (belowGpuRange)
+		ulong value = n;
-		// {
+		byte flag = 0;
-			return IsPrime(n);
-		// }
 
-		// var limiter = GpuPrimeWorkLimiter.Acquire();
+		var input = gpu.Input;
-		// var gpu = PrimeTesterGpuContextPool.Rent(1);
+		var output = gpu.Output;
-		// var accelerator = gpu.Accelerator;
+		var inputView = input.View;
-		// var kernel = gpu.Kernel;
+		var outputView = output.View;
 
-		// ulong value = n;
+		inputView.CopyFromCPU(ref value, 1);
-		// byte flag = 0;
+		kernel(
+						1,
+						inputView,
+						gpu.DevicePrimesDefault.View,
+						gpu.DevicePrimesLastOne.View,
+						gpu.DevicePrimesLastSeven.View,
+						gpu.DevicePrimesLastThree.View,
+						gpu.DevicePrimesLastNine.View,
+						gpu.DevicePrimesPow2Default.View,
+						gpu.DevicePrimesPow2LastOne.View,
+						gpu.DevicePrimesPow2LastSeven.View,
+						gpu.DevicePrimesPow2LastThree.View,
+						gpu.DevicePrimesPow2LastNine.View,
+						outputView);
+		accelerator.Synchronize();
+		outputView.CopyToCPU(ref flag, 1);
 
-		// var input = gpu.Input;
+		gpu.Dispose();
-		// var output = gpu.Output;
+		GpuPrimeWorkLimiter.Release();
-		// var inputView = input.View;
-		// var outputView = output.View;
 
-		// inputView.CopyFromCPU(ref value, 1);
+		return flag != 0;
-		// kernel(
-		// 				1,
-		// 				inputView,
-		// 				gpu.DevicePrimesDefault.View,
-		// 				gpu.DevicePrimesLastOne.View,
-		// 				gpu.DevicePrimesLastSeven.View,
-		// 				gpu.DevicePrimesLastThree.View,
-		// 				gpu.DevicePrimesLastNine.View,
-		// 				gpu.DevicePrimesPow2Default.View,
-		// 				gpu.DevicePrimesPow2LastOne.View,
-		// 				gpu.DevicePrimesPow2LastSeven.View,
-		// 				gpu.DevicePrimesPow2LastThree.View,
-		// 				gpu.DevicePrimesPow2LastNine.View,
-		// 				outputView);
-		// accelerator.Synchronize();
-		// outputView.CopyToCPU(ref flag, 1);
-
-		// gpu.Dispose();
-		// limiter.Dispose();
-
-		// return flag != 0;
 	}
 
 	public static int GpuBatchSize { get; set; } = 262_144;
@@ -159,26 +152,7 @@ public sealed class PrimeTester
 				return;
 			}
 
-			int toWarm = target - WarmedGpuLeaseCount;
+			PrimeTesterGpuContextPool.WarmUp(target, GpuBatchSize);
-			var leases = new PrimeTesterGpuContextPool.PrimeTesterGpuContextLease[toWarm];
-			ReadOnlySpan<byte> supportedDigits = PrimeTesterGpuContextPool.SupportedDigits;
-			int digitIndex = 0;
-			for (int i = 0; i < toWarm; i++)
-			{
-				byte trailingDigit = supportedDigits[digitIndex];
-				leases[i] = PrimeTesterGpuContextPool.Rent(trailingDigit, GpuBatchSize);
-				digitIndex++;
-				if (digitIndex == supportedDigits.Length)
-				{
-					digitIndex = 0;
-				}
-			}
-
-			for (int i = 0; i < toWarm; i++)
-			{
-				leases[i].Dispose();
-			}
-
 			WarmedGpuLeaseCount = target;
 		}
 	}
@@ -186,45 +160,25 @@ public sealed class PrimeTester
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static void IsPrimeBatchGpu(ReadOnlySpan<ulong> values, Span<byte> results)
 	{
-		var limiter = GpuPrimeWorkLimiter.Acquire();
+		GpuPrimeWorkLimiter.Acquire();
+		var gpu = PrimeTesterGpuContextPool.Rent(GpuBatchSize);
+		var accelerator = gpu.Accelerator;
+		var kernel = PrimeTesterGpuContextPool.PrimeTesterGpuContextLease.Kernel;
 		int totalLength = values.Length;
 		int batchSize = GpuBatchSize;
-		int pos = 0;
 
-		while (pos < totalLength)
-		{
-			ulong currentValue = values[pos];
-			byte trailingDigit = (byte)(currentValue % 10UL);
-			var gpu = PrimeTesterGpuContextPool.Rent(trailingDigit, GpuBatchSize);
-			var accelerator = gpu.Accelerator;
-			var kernel = gpu.Kernel;
 		var input = gpu.Input;
 		var output = gpu.Output;
 		var inputView = input.View;
 		var outputView = output.View;
 
-			int blockStart = pos;
+		int pos = 0;
-			while (blockStart < totalLength)
+		while (pos < totalLength)
 		{
-				byte blockDigit = (byte)(values[blockStart] % 10UL);
+			int remaining = totalLength - pos;
-				if (blockDigit != trailingDigit)
+			int count = remaining > batchSize ? batchSize : remaining;
-				{
-					break;
-				}
 
-				int limit = Math.Min(batchSize, totalLength - blockStart);
+			var valueSlice = values.Slice(pos, count);
-				int count = 0;
-				while (count < limit && (byte)(values[blockStart + count] % 10UL) == trailingDigit)
-				{
-					count++;
-				}
-
-				if (count == 0)
-				{
-					break;
-				}
-
-				var valueSlice = values.Slice(blockStart, count);
 			ref ulong valueRef = ref MemoryMarshal.GetReference(valueSlice);
 			inputView.CopyFromCPU(ref valueRef, count);
 
@@ -243,77 +197,94 @@ public sealed class PrimeTester
 					gpu.DevicePrimesPow2LastNine.View,
 					outputView);
 			accelerator.Synchronize();
-				var resultSlice = results.Slice(blockStart, count);
+			var resultSlice = results.Slice(pos, count);
 			ref byte resultRef = ref MemoryMarshal.GetReference(resultSlice);
 			outputView.CopyToCPU(ref resultRef, count);
 
-				blockStart += count;
+			pos += count;
 		}
 
 		gpu.Dispose();
-			pos = blockStart;
+		GpuPrimeWorkLimiter.Release();
-		}
-
-		limiter.Dispose();
 	}
 
 	internal static class PrimeTesterGpuContextPool
 	{
-		internal enum HeuristicGpuTableScope : byte
+		private static readonly ConcurrentQueue<PrimeTesterGpuContextLease> Pool = new();
-		{
+		private static readonly object WarmUpLock = new();
-			GroupAB = 0,
+		private static int WarmedLeaseCount;
-			GroupABWithCombined = 1,
-		}
-
-		private static readonly byte[] SupportedTrailingDigits = [1, 3, 7, 9];
-		private const int ScopeCount = 2;
-		private static readonly ConcurrentQueue<PrimeTesterGpuContextLease>[,] Pools;
 		private static readonly object SharedTablesLock = new();
-		private static readonly Dictionary<string, SharedKernelTables> SharedTables = new();
+		private static readonly Dictionary<Accelerator, SharedHeuristicGpuTables> SharedTables = new(AcceleratorReferenceComparer.Instance);
-		private static readonly Dictionary<(string Key, byte Digit, HeuristicGpuTableScope Scope), TrailingDigitTables> DigitTables = new();
 
-		static PrimeTesterGpuContextPool()
+		private static SharedHeuristicGpuTables RentSharedTables(Accelerator accelerator)
 		{
-			Pools = new ConcurrentQueue<PrimeTesterGpuContextLease>[SupportedTrailingDigits.Length, ScopeCount];
+			lock (SharedTablesLock)
-			for (int digitIndex = 0; digitIndex < SupportedTrailingDigits.Length; digitIndex++)
 			{
-				for (int scopeIndex = 0; scopeIndex < ScopeCount; scopeIndex++)
+				if (!SharedTables.TryGetValue(accelerator, out var tables))
 				{
-					Pools[digitIndex, scopeIndex] = new ConcurrentQueue<PrimeTesterGpuContextLease>();
+					tables = new SharedHeuristicGpuTables(accelerator);
+					SharedTables.Add(accelerator, tables);
 				}
+
+				tables.AddReference();
+				return tables;
 			}
 		}
 
-		internal static ReadOnlySpan<byte> SupportedDigits => SupportedTrailingDigits;
+		private static void ReleaseSharedTables(SharedHeuristicGpuTables tables)
-
-		private static int GetPoolIndex(byte trailingDigit)
 		{
-			return trailingDigit switch
+			if (tables is null)
 			{
-				1 => 0,
+				return;
-				3 => 1,
-				7 => 2,
-				9 => 3,
-				_ => throw new ArgumentOutOfRangeException(nameof(trailingDigit), trailingDigit, "Only trailing digits 1, 3, 7, and 9 are supported."),
-			};
 			}
 
-		private static int GetScopeIndex(HeuristicGpuTableScope scope)
+			lock (SharedTablesLock)
 			{
-			return scope switch
+				tables.Release();
-			{
+			}
-				HeuristicGpuTableScope.GroupAB => 0,
-				HeuristicGpuTableScope.GroupABWithCombined => 1,
-				_ => throw new ArgumentOutOfRangeException(nameof(scope), scope, "Unsupported heuristic table scope."),
-			};
 		}
 
-		internal static PrimeTesterGpuContextLease Rent(byte trailingDigit, int minBufferCapacity = 1, HeuristicGpuTableScope tableScope = HeuristicGpuTableScope.GroupAB)
+		internal static void EnsureStaticTables(KernelContainer kernels, Accelerator accelerator)
 		{
-			var pool = Pools[GetPoolIndex(trailingDigit), GetScopeIndex(tableScope)];
+			var tables = RentSharedTables(accelerator);
-			if (!pool.TryDequeue(out var lease))
+			var heuristicTables = tables.CreateHeuristicDivisorTables();
+			HeuristicGpuDivisorTables.InitializeShared(in heuristicTables);
+			ReleaseSharedTables(tables);
+		}
+
+		internal static void WarmUp(int target, int minBufferCapacity)
 		{
-				lease = new PrimeTesterGpuContextLease(trailingDigit, minBufferCapacity, tableScope);
+			if (target <= 0)
+			{
+				return;
+			}
+
+			lock (WarmUpLock)
+			{
+				if (target <= WarmedLeaseCount)
+				{
+					return;
+				}
+
+				KernelContainer kernels = GpuKernelPool.GetKernels();
+				GpuStaticTableInitializer.EnsureStaticTables(kernels, SharedGpuContext.Accelerator);
+
+				int toCreate = target - WarmedLeaseCount;
+				for (int i = 0; i < toCreate; i++)
+				{
+					Pool.Enqueue(new PrimeTesterGpuContextLease(minBufferCapacity));
+				}
+
+				kernels.Dispose();
+				WarmedLeaseCount = target;
+			}
+		}
+
+                internal static PrimeTesterGpuContextLease Rent(int minBufferCapacity = 1)
+                {
+                        if (!Pool.TryDequeue(out var lease))
+                        {
+                                lease = new PrimeTesterGpuContextLease(minBufferCapacity);
                         }
                         else
                         {
@@ -324,42 +295,33 @@ public sealed class PrimeTester
                         return lease;
                 }
 
+
 		internal static void Return(PrimeTesterGpuContextLease lease)
 		{
 			lease.Accelerator.Synchronize();
-			Pools[GetPoolIndex(lease.TrailingDigit), GetScopeIndex(lease.TableScope)].Enqueue(lease);
+			Pool.Enqueue(lease);
 		}
 
 		internal static void DisposeAll()
 		{
-			for (int digitIndex = 0; digitIndex < SupportedTrailingDigits.Length; digitIndex++)
+			while (Pool.TryDequeue(out var lease))
-			{
-				for (int scopeIndex = 0; scopeIndex < ScopeCount; scopeIndex++)
-				{
-					var pool = Pools[digitIndex, scopeIndex];
-					while (pool.TryDequeue(out var lease))
 			{
 				lease.DisposeResources();
 			}
-				}
-			}
 
 			lock (SharedTablesLock)
 			{
-				foreach (var entry in DigitTables.Values)
-				{
-					entry.Dispose();
-				}
-
-				DigitTables.Clear();
-
 				foreach (var entry in SharedTables.Values)
 				{
 					entry.Dispose();
 				}
 
 				SharedTables.Clear();
+				var emptyTables = default(HeuristicGpuDivisorTables);
+				HeuristicGpuDivisorTables.InitializeShared(in emptyTables);
 			}
+
+			WarmedLeaseCount = 0;
 		}
 
 		internal static void Clear(Accelerator accelerator)
@@ -369,13 +331,10 @@ public sealed class PrimeTester
 				return;
 			}
 
+			lock (WarmUpLock)
+			{
 				var retained = new List<PrimeTesterGpuContextLease>();
-			for (int digitIndex = 0; digitIndex < SupportedTrailingDigits.Length; digitIndex++)
+				while (Pool.TryDequeue(out var lease))
-			{
-				for (int scopeIndex = 0; scopeIndex < ScopeCount; scopeIndex++)
-				{
-					var pool = Pools[digitIndex, scopeIndex];
-					while (pool.TryDequeue(out var lease))
 				{
 					if (lease.Accelerator == accelerator)
 					{
@@ -390,172 +349,46 @@ public sealed class PrimeTester
 				int retainedCount = retained.Count;
 				for (int i = 0; i < retainedCount; i++)
 				{
-						pool.Enqueue(retained[i]);
+					Pool.Enqueue(retained[i]);
 				}
 
-					retained.Clear();
+				WarmedLeaseCount = Pool.Count;
-				}
 			}
 
 			lock (SharedTablesLock)
 			{
-				string key = CreateAcceleratorKey(accelerator);
+				if (SharedTables.Remove(accelerator, out var tables))
-
-				if (SharedTables.Remove(key, out var shared))
-				{
-					shared.Dispose();
-				}
-
-				if (DigitTables.Count > 0)
-				{
-					var toRemove = new List<(string Key, byte Digit, HeuristicGpuTableScope Scope)>();
-					foreach (var entry in DigitTables)
-					{
-						if (entry.Key.Key == key)
-						{
-							entry.Value.Dispose();
-							toRemove.Add(entry.Key);
-						}
-					}
-
-					for (int i = 0; i < toRemove.Count; i++)
-					{
-						DigitTables.Remove(toRemove[i]);
-					}
-				}
-			}
-		}
-
-		private static string CreateAcceleratorKey(Accelerator accelerator)
-		{
-			return accelerator.AcceleratorType + ":" + accelerator.Name;
-		}
-
-		private static SharedKernelTables RentSharedTables(Accelerator accelerator)
-		{
-			string key = CreateAcceleratorKey(accelerator);
-
-			lock (SharedTablesLock)
-			{
-				if (!SharedTables.TryGetValue(key, out var tables))
-				{
-					tables = new SharedKernelTables(accelerator, key);
-					SharedTables.Add(key, tables);
-				}
-
-				tables.AddReference();
-				return tables;
-			}
-		}
-
-		private static void ReleaseSharedTables(SharedKernelTables tables)
-		{
-			if (tables is null)
-			{
-				return;
-			}
-
-			lock (SharedTablesLock)
-			{
-				if (tables.Release() && SharedTables.Remove(tables.Key, out _))
 				{
 					tables.Dispose();
 				}
-			}
-		}
 
-		private static TrailingDigitTables RentTrailingDigitTables(Accelerator accelerator, byte trailingDigit, HeuristicGpuTableScope scope)
+				if (SharedTables.Count == 0)
 				{
-			string key = CreateAcceleratorKey(accelerator);
+					var emptyTables = default(HeuristicGpuDivisorTables);
-			var compositeKey = (key, trailingDigit, scope);
+					HeuristicGpuDivisorTables.InitializeShared(in emptyTables);
-
-			lock (SharedTablesLock)
-			{
-				if (!DigitTables.TryGetValue(compositeKey, out var tables))
-				{
-					tables = new TrailingDigitTables(accelerator, key, trailingDigit, scope);
-					DigitTables.Add(compositeKey, tables);
-				}
-
-				tables.AddReference();
-				return tables;
-			}
-		}
-
-		private static void ReleaseTrailingDigitTables(TrailingDigitTables tables)
-		{
-			if (tables is null)
-			{
-				return;
-			}
-
-			lock (SharedTablesLock)
-			{
-				var compositeKey = (tables.Key, tables.TrailingDigit, tables.Scope);
-				if (tables.Release() && DigitTables.Remove(compositeKey))
-				{
-					tables.Dispose();
 				}
 			}
 		}
 
 		internal sealed class PrimeTesterGpuContextLease : IDisposable
 		{
+
 			private bool _returned;
-		private readonly SharedKernelTables _sharedTables;
+			private readonly SharedHeuristicGpuTables _sharedTables;
-		private readonly TrailingDigitTables _digitTables;
 			private readonly HeuristicGpuDivisorTables _heuristicDivisorTables;
-		private readonly HeuristicGpuCombinedDivisorTables _heuristicCombinedTables;
 
 			internal readonly Context AcceleratorContext;
 			public readonly Accelerator Accelerator;
-		public readonly byte TrailingDigit;
+			
-		public readonly HeuristicGpuTableScope TableScope;
+			public 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>> Kernel = SharedGpuContext.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>>(PrimeTesterKernels.SmallPrimeSieveKernel);
-		public 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>> Kernel;
+
-		public readonly Action<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables, HeuristicGpuCombinedDivisorTables> HeuristicTrialDivisionKernel;
+			public static readonly Action<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables> HeuristicTrialDivisionKernel = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.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;
 
-		internal PrimeTesterGpuContextLease(byte trailingDigit, int minBufferCapacity, HeuristicGpuTableScope tableScope)
-		{
-			TrailingDigit = trailingDigit;
-			TableScope = tableScope;
-
-			AcceleratorContext = Context.CreateDefault();
-			Accelerator = AcceleratorContext.GetPreferredDevice(false).CreateAccelerator(AcceleratorContext);
-			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>>(PrimeTesterKernels.SmallPrimeSieveKernel);
-			HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables, HeuristicGpuCombinedDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
-
-			_sharedTables = RentSharedTables(Accelerator);
-			_digitTables = RentTrailingDigitTables(Accelerator, trailingDigit, tableScope);
-
-			_heuristicDivisorTables = new HeuristicGpuDivisorTables(
-				_sharedTables.HeuristicGroupADivisors.View,
-				_digitTables.GroupBDivisorsEnding1View,
-				_digitTables.GroupBDivisorsEnding7View,
-				_digitTables.GroupBDivisorsEnding9View);
-			HeuristicGpuDivisorTables.InitializeShared(in _heuristicDivisorTables);
-
-			if (_digitTables.HasCombinedTables)
-			{
-				_heuristicCombinedTables = new HeuristicGpuCombinedDivisorTables(
-					_digitTables.CombinedDivisorsEnding1View,
-					_digitTables.CombinedDivisorsEnding3View,
-					_digitTables.CombinedDivisorsEnding7View,
-					_digitTables.CombinedDivisorsEnding9View);
-				HeuristicGpuCombinedDivisorTables.InitializeShared(in _heuristicCombinedTables);
-			}
-			else
-			{
-				_heuristicCombinedTables = HeuristicGpuCombinedDivisorTables.Shared;
-			}
-
-			EnsureCapacity(minBufferCapacity);
-			_returned = false;
-		}
-
 			public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesDefault => _sharedTables.DevicePrimesDefault;
 			public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne => _sharedTables.DevicePrimesLastOne;
 			public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven => _sharedTables.DevicePrimesLastSeven;
@@ -566,11 +399,43 @@ public sealed class PrimeTester
 			public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven => _sharedTables.DevicePrimesPow2LastSeven;
 			public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree => _sharedTables.DevicePrimesPow2LastThree;
 			public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine => _sharedTables.DevicePrimesPow2LastNine;
+			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;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding1 => _sharedTables.HeuristicCombinedDivisorsEnding1;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding1 => _sharedTables.HeuristicCombinedDivisorSquaresEnding1;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding3 => _sharedTables.HeuristicCombinedDivisorsEnding3;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding3 => _sharedTables.HeuristicCombinedDivisorSquaresEnding3;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding7 => _sharedTables.HeuristicCombinedDivisorsEnding7;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding7 => _sharedTables.HeuristicCombinedDivisorSquaresEnding7;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding9 => _sharedTables.HeuristicCombinedDivisorsEnding9;
+			public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding9 => _sharedTables.HeuristicCombinedDivisorSquaresEnding9;
 
 			public HeuristicGpuDivisorTables HeuristicGpuTables => _heuristicDivisorTables;
-		public HeuristicGpuCombinedDivisorTables HeuristicCombinedTables => _heuristicCombinedTables;
 
-		internal void EnsureCapacity(int minCapacity)
+			internal PrimeTesterGpuContextLease(int minBufferCapacity)
+			{
+				AcceleratorContext = SharedGpuContext.Context;
+				Accelerator = SharedGpuContext.Accelerator;
+				// GpuStaticTableInitializer.EnsureStaticTables(Accelerator);
+				// 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>>(PrimeTesterKernels.SmallPrimeSieveKernel);
+				// HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
+
+				_sharedTables = RentSharedTables(Accelerator);
+
+				_heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
+				HeuristicGpuDivisorTables.InitializeShared(in _heuristicDivisorTables);
+
+				EnsureCapacity(minBufferCapacity);
+				_returned = false;
+			}
+
+			public void EnsureCapacity(int minCapacity)
 			{
 				if (minCapacity <= BufferCapacity)
 				{
@@ -599,7 +464,7 @@ public sealed class PrimeTester
 				}
 
 				_returned = true;
-			PrimeTesterGpuContextPool.Return(this);
+				Return(this);
 			}
 
 			internal void DisposeResources()
@@ -607,19 +472,34 @@ public sealed class PrimeTester
 				Input?.Dispose();
 				Output?.Dispose();
 				HeuristicFlag?.Dispose();
-			ReleaseTrailingDigitTables(_digitTables);
 				ReleaseSharedTables(_sharedTables);
-			Accelerator.Dispose();
+				// These resources are shared between GPU leases
-			AcceleratorContext.Dispose();
+				// Accelerator.Dispose();
+				// AcceleratorContext.Dispose();
 			}
 		}
-	private sealed class SharedKernelTables
+
+		private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
+		{
+			internal static AcceleratorReferenceComparer Instance { get; } = new();
+
+			public bool Equals(Accelerator? x, Accelerator? y)
+			{
+				return ReferenceEquals(x, y);
+			}
+
+			public int GetHashCode(Accelerator obj)
+			{
+				return RuntimeHelpers.GetHashCode(obj);
+			}
+		}
+
+		private sealed class SharedHeuristicGpuTables
 		{
 			private int _referenceCount;
 
-			internal SharedKernelTables(Accelerator accelerator, string key)
+			internal SharedHeuristicGpuTables(Accelerator accelerator)
 			{
-				Key = key;
 
 				var primesDefault = DivisorGenerator.SmallPrimes;
 				DevicePrimesDefault = accelerator.Allocate1D<uint>(primesDefault.Length);
@@ -662,10 +542,44 @@ public sealed class PrimeTester
 				DevicePrimesPow2LastNine.View.CopyFromCPU(primesPow2LastNine);
 
 				HeuristicPrimeSieves.EnsureInitialized();
-				HeuristicGroupADivisors = CopySpanToDevice(accelerator, HeuristicPrimeSieves.GroupADivisors);
+				HeuristicCombinedPrimeTester.EnsureInitialized();
+
+				var heuristicGroupA = HeuristicPrimeSieves.GroupADivisors;
+				HeuristicGroupADivisors = CopySpanToDevice(accelerator, heuristicGroupA);
+
+				var heuristicGroupASquares = HeuristicPrimeSieves.GroupADivisorSquares;
+				HeuristicGroupADivisorSquares = CopySpanToDevice(accelerator, heuristicGroupASquares);
+
+				HeuristicGroupBDivisorsEnding1 = CopyUintSpanToDevice(accelerator, DivisorGenerator.SmallPrimesLastOneWithoutLastThree);
+				HeuristicGroupBDivisorSquaresEnding1 = CopySpanToDevice(accelerator, DivisorGenerator.SmallPrimesPow2LastOneWithoutLastThree);
+
+				HeuristicGroupBDivisorsEnding7 = CopyUintSpanToDevice(accelerator, DivisorGenerator.SmallPrimesLastSevenWithoutLastThree);
+				HeuristicGroupBDivisorSquaresEnding7 = CopySpanToDevice(accelerator, DivisorGenerator.SmallPrimesPow2LastSevenWithoutLastThree);
+
+				HeuristicGroupBDivisorsEnding9 = CopyUintSpanToDevice(accelerator, DivisorGenerator.SmallPrimesLastNineWithoutLastThree);
+				HeuristicGroupBDivisorSquaresEnding9 = CopySpanToDevice(accelerator, DivisorGenerator.SmallPrimesPow2LastNineWithoutLastThree);
+
+				var combinedEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1Span;
+				HeuristicCombinedDivisorsEnding1 = CopySpanToDevice(accelerator, combinedEnding1);
+				var combinedSquaresEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1SquaresSpan;
+				HeuristicCombinedDivisorSquaresEnding1 = CopySpanToDevice(accelerator, combinedSquaresEnding1);
+
+				var combinedEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3Span;
+				HeuristicCombinedDivisorsEnding3 = CopySpanToDevice(accelerator, combinedEnding3);
+				var combinedSquaresEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3SquaresSpan;
+				HeuristicCombinedDivisorSquaresEnding3 = CopySpanToDevice(accelerator, combinedSquaresEnding3);
+
+				var combinedEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7Span;
+				HeuristicCombinedDivisorsEnding7 = CopySpanToDevice(accelerator, combinedEnding7);
+				var combinedSquaresEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7SquaresSpan;
+				HeuristicCombinedDivisorSquaresEnding7 = CopySpanToDevice(accelerator, combinedSquaresEnding7);
+
+				var combinedEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9Span;
+				HeuristicCombinedDivisorsEnding9 = CopySpanToDevice(accelerator, combinedEnding9);
+				var combinedSquaresEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9SquaresSpan;
+				HeuristicCombinedDivisorSquaresEnding9 = CopySpanToDevice(accelerator, combinedSquaresEnding9);
 			}
 
-			internal string Key { get; }
 			internal MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesDefault { get; }
 			internal MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne { get; }
 			internal MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven { get; }
@@ -677,122 +591,54 @@ public sealed class PrimeTester
 			internal MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree { get; }
 			internal MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine { get; }
 			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisors { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisorSquares { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding1 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding1 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding7 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding7 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding9 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding9 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding1 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding1 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding3 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding3 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding7 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding7 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding9 { get; }
+			internal MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding9 { get; }
+
+			internal HeuristicGpuDivisorTables CreateHeuristicDivisorTables()
+			{
+				return new HeuristicGpuDivisorTables(
+					HeuristicCombinedDivisorsEnding1.View,
+					HeuristicCombinedDivisorsEnding3.View,
+					HeuristicCombinedDivisorsEnding7.View,
+					HeuristicCombinedDivisorsEnding9.View,
+					HeuristicCombinedDivisorSquaresEnding1.View,
+					HeuristicCombinedDivisorSquaresEnding3.View,
+					HeuristicCombinedDivisorSquaresEnding7.View,
+					HeuristicCombinedDivisorSquaresEnding9.View,
+					HeuristicGroupADivisors.View,
+					HeuristicGroupADivisorSquares.View,
+					HeuristicGroupBDivisorsEnding1.View,
+					HeuristicGroupBDivisorSquaresEnding1.View,
+					HeuristicGroupBDivisorsEnding7.View,
+					HeuristicGroupBDivisorSquaresEnding7.View,
+					HeuristicGroupBDivisorsEnding9.View,
+					HeuristicGroupBDivisorSquaresEnding9.View);
+			}
 
 			internal void AddReference()
 			{
 				_referenceCount++;
 			}
 
-			internal bool Release()
+			internal void Release()
 			{
 				if (_referenceCount > 0)
 				{
 					_referenceCount--;
 				}
-
-				return _referenceCount == 0;
-			}
-
-			internal void Dispose()
-			{
-				DevicePrimesDefault.Dispose();
-				DevicePrimesLastOne.Dispose();
-				DevicePrimesLastSeven.Dispose();
-				DevicePrimesLastThree.Dispose();
-				DevicePrimesLastNine.Dispose();
-				DevicePrimesPow2Default.Dispose();
-				DevicePrimesPow2LastOne.Dispose();
-				DevicePrimesPow2LastSeven.Dispose();
-				DevicePrimesPow2LastThree.Dispose();
-				DevicePrimesPow2LastNine.Dispose();
-				HeuristicGroupADivisors.Dispose();
-			}
-		}
-
-		private sealed class TrailingDigitTables
-		{
-			private int _referenceCount;
-
-			internal TrailingDigitTables(Accelerator accelerator, string key, byte trailingDigit, HeuristicGpuTableScope scope)
-			{
-				Key = key;
-				TrailingDigit = trailingDigit;
-				Scope = scope;
-
-				HeuristicPrimeSieves.EnsureInitialized();
-				HeuristicCombinedPrimeTester.EnsureInitialized();
-
-				if (trailingDigit == 1 || trailingDigit == 7 || trailingDigit == 9)
-				{
-					GroupBDivisorsEnding1 = CopyUintSpanToDevice(accelerator, DivisorGenerator.SmallPrimesLastOneWithoutLastThree);
-				}
-
-				if (trailingDigit == 3 || trailingDigit == 7 || trailingDigit == 9)
-				{
-					GroupBDivisorsEnding7 = CopyUintSpanToDevice(accelerator, DivisorGenerator.SmallPrimesLastSevenWithoutLastThree);
-				}
-
-				if (trailingDigit == 1 || trailingDigit == 3 || trailingDigit == 9)
-				{
-					GroupBDivisorsEnding9 = CopyUintSpanToDevice(accelerator, DivisorGenerator.SmallPrimesLastNineWithoutLastThree);
-				}
-
-				if (scope == HeuristicGpuTableScope.GroupABWithCombined)
-				{
-					CombinedDivisorsEnding1 = trailingDigit == 1 ? CopySpanToDevice(accelerator, HeuristicCombinedPrimeTester.CombinedDivisorsEnding1Span) : null;
-					CombinedDivisorsEnding3 = trailingDigit == 3 ? CopySpanToDevice(accelerator, HeuristicCombinedPrimeTester.CombinedDivisorsEnding3Span) : null;
-					CombinedDivisorsEnding7 = trailingDigit == 7 ? CopySpanToDevice(accelerator, HeuristicCombinedPrimeTester.CombinedDivisorsEnding7Span) : null;
-					CombinedDivisorsEnding9 = trailingDigit == 9 ? CopySpanToDevice(accelerator, HeuristicCombinedPrimeTester.CombinedDivisorsEnding9Span) : null;
-				}
-			}
-
-			internal string Key { get; }
-			internal byte TrailingDigit { get; }
-			internal HeuristicGpuTableScope Scope { get; }
-
-			internal bool HasCombinedTables => Scope == HeuristicGpuTableScope.GroupABWithCombined;
-
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? CombinedDivisorsEnding1 { get; }
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? CombinedDivisorsEnding3 { get; }
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? CombinedDivisorsEnding7 { get; }
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? CombinedDivisorsEnding9 { get; }
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? GroupBDivisorsEnding1 { get; }
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? GroupBDivisorsEnding7 { get; }
-			internal MemoryBuffer1D<ulong, Stride1D.Dense>? GroupBDivisorsEnding9 { get; }
-
-			internal ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding1View => CombinedDivisorsEnding1?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-			internal ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding3View => CombinedDivisorsEnding3?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-			internal ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding7View => CombinedDivisorsEnding7?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-			internal ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding9View => CombinedDivisorsEnding9?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-			internal ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding1View => GroupBDivisorsEnding1?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-			internal ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding7View => GroupBDivisorsEnding7?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-			internal ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding9View => GroupBDivisorsEnding9?.View ?? ArrayView1D<ulong, Stride1D.Dense>.Empty;
-
-			internal void AddReference()
-			{
-				_referenceCount++;
-			}
-
-			internal bool Release()
-			{
-				if (_referenceCount > 0)
-				{
-					_referenceCount--;
-				}
-
-				return _referenceCount == 0;
-			}
-
-			internal void Dispose()
-			{
-				CombinedDivisorsEnding1?.Dispose();
-				CombinedDivisorsEnding3?.Dispose();
-				CombinedDivisorsEnding7?.Dispose();
-				CombinedDivisorsEnding9?.Dispose();
-				GroupBDivisorsEnding1?.Dispose();
-				GroupBDivisorsEnding7?.Dispose();
-				GroupBDivisorsEnding9?.Dispose();
-			}
 			}
 
 			private static MemoryBuffer1D<ulong, Stride1D.Dense> CopySpanToDevice(Accelerator accelerator, ReadOnlySpan<ulong> span)
@@ -800,8 +646,8 @@ public sealed class PrimeTester
 				var buffer = accelerator.Allocate1D<ulong>(span.Length);
 				if (!span.IsEmpty)
 				{
-				ref ulong source = ref MemoryMarshal.GetReference(span);
+					ref ulong sourceRef = ref MemoryMarshal.GetReference(span);
-				buffer.View.CopyFromCPU(ref source, span.Length);
+					buffer.View.CopyFromCPU(ref sourceRef, span.Length);
 				}
 
 				return buffer;
@@ -823,6 +669,37 @@ public sealed class PrimeTester
 
 				return buffer;
 			}
+
+			internal void Dispose()
+			{
+				DevicePrimesDefault.Dispose();
+				DevicePrimesLastOne.Dispose();
+				DevicePrimesLastSeven.Dispose();
+				DevicePrimesLastThree.Dispose();
+				DevicePrimesLastNine.Dispose();
+				DevicePrimesPow2Default.Dispose();
+				DevicePrimesPow2LastOne.Dispose();
+				DevicePrimesPow2LastSeven.Dispose();
+				DevicePrimesPow2LastThree.Dispose();
+				DevicePrimesPow2LastNine.Dispose();
+				HeuristicGroupADivisors.Dispose();
+				HeuristicGroupADivisorSquares.Dispose();
+				HeuristicGroupBDivisorsEnding1.Dispose();
+				HeuristicGroupBDivisorSquaresEnding1.Dispose();
+				HeuristicGroupBDivisorsEnding7.Dispose();
+				HeuristicGroupBDivisorSquaresEnding7.Dispose();
+				HeuristicGroupBDivisorsEnding9.Dispose();
+				HeuristicGroupBDivisorSquaresEnding9.Dispose();
+				HeuristicCombinedDivisorsEnding1.Dispose();
+				HeuristicCombinedDivisorSquaresEnding1.Dispose();
+				HeuristicCombinedDivisorsEnding3.Dispose();
+				HeuristicCombinedDivisorSquaresEnding3.Dispose();
+				HeuristicCombinedDivisorsEnding7.Dispose();
+				HeuristicCombinedDivisorSquaresEnding7.Dispose();
+				HeuristicCombinedDivisorsEnding9.Dispose();
+				HeuristicCombinedDivisorSquaresEnding9.Dispose();
+			}
+		}
 	}
 
 	// Expose cache clearing for accelerator disposal coordination
@@ -855,7 +732,7 @@ public sealed class PrimeTester
 		// TODO: Route this batch helper through the shared GPU kernel pool from
 		// GpuUInt128BinaryGcdBenchmarks so we reuse cached kernels, pinned host buffers,
 		// and divisor-cycle staging instead of allocating new device buffers per call.
-		var gpu = PrimeTesterGpuContextPool.Rent(1);
+		var gpu = PrimeTesterGpuContextPool.Rent();
 		var accelerator = gpu.Accelerator;
 		var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel);
 

docs/gpu-context-rework-plan.md

@@ -1,8 +0,0 @@
-# GPU context pooling and heuristic divisor updates
-
-- [x] Review existing GPU context pooling in `PerfectNumbers.Core/PrimeTester.cs` and determine shared table allocations that can be scoped per trailing digit.
-- [x] Split `PrimeTesterGpuContextPool` into trailing-digit specific pools so each lease only uploads the divisor tables required for that digit.
-- [x] Update GPU warm-up and rental code paths to pick pools based on `n % 10` and ensure thread-safe reuse.
-- [x] Replace square-table usage in GPU heuristic trial division with square-root limits so divisor tables no longer need squared values on the accelerator.
-- [x] Remove GPU square buffers and adjust kernels, table types, and callers accordingly.
-- [x] Build solution to confirm the new pooling and divisor logic compiles; tests documented in summary.