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

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;
@@ -176,7 +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)
@@ -626,7 +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/GpuContextPool.cs

@@ -1,4 +1,3 @@
-using System;
 using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
@@ -15,14 +14,12 @@ public static class GpuContextPool
 		public readonly Context Context;
 		public readonly Accelerator Accelerator;
 
-		private static readonly Context _sharedContext = Context.CreateDefault();
+		public PooledContext(KernelContainer kernels)
-		private static readonly Device _sharedDevice = _sharedContext.GetPreferredDevice(false);
-		private static readonly Accelerator _sharedAccelerator = _sharedDevice.CreateAccelerator(_sharedContext);
-
-		public PooledContext()
 		{
-			Context = _sharedContext;
+			Context = SharedGpuContext.Context;
-			Accelerator = _sharedAccelerator;
+			Accelerator = SharedGpuContext.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
@@ -56,7 +53,7 @@ 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()));
 	}
 
 
@@ -74,7 +71,8 @@ public static class GpuContextPool
 		PooledContext context;
 		for (int i = 0; i < toCreate; i++)
 		{
-			context = new PooledContext();
+			KernelContainer kernels = GpuKernelPool.GetKernels();
+			context = new PooledContext(kernels);
 			contexts[i] = context;
 			pool.Enqueue(context);
 		}

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

@@ -8,14 +8,8 @@ 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)
     {
@@ -36,19 +30,4 @@ public static class GpuPrimeWorkLimiter
         _currentLimit = value;
         previous.Dispose();
     }
-
-    public sealed class Lease : IDisposable
-    {
-        private readonly SemaphoreSlim _sem;
-
-        internal Lease(SemaphoreSlim sem)
-        {
-            _sem = sem;
-        }
-
-        public void Dispose()
-        {
-            _sem.Release();
-        }
-    }
 }

PerfectNumbers.Core/Gpu/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/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/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;
@@ -225,10 +224,10 @@ 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(1);
 		var accelerator = gpu.Accelerator;
-                var kernel = gpu.HeuristicTrialDivisionKernel;
+		var kernel = PrimeTester.PrimeTesterGpuContextPool.PrimeTesterGpuContextLease.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
 		var flagView = flagView1D.AsContiguous();
 
@@ -249,7 +248,7 @@ public sealed class HeuristicCombinedPrimeTester
 		compositeDetected = compositeFlag != 0;
 
 		gpu.Dispose();
-                limiter.Dispose();
+		GpuPrimeWorkLimiter.Release();
 		return compositeDetected;
 	}
 

PerfectNumbers.Core/HeuristicPrimeTester.cs

@@ -240,10 +240,10 @@ public sealed class HeuristicPrimeTester
 
 	private bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
 	{
-                var limiter = GpuPrimeWorkLimiter.Acquire();
+		GpuPrimeWorkLimiter.Acquire();
 		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();
 
@@ -301,7 +301,7 @@ public sealed class HeuristicPrimeTester
 		}
 
 		gpu.Dispose();
-                limiter.Dispose();
+		GpuPrimeWorkLimiter.Release();
 		return compositeDetected;
 	}
 

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;
@@ -167,10 +160,10 @@ 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 = gpu.Kernel;
+		var kernel = PrimeTesterGpuContextPool.PrimeTesterGpuContextLease.Kernel;
 		int totalLength = values.Length;
 		int batchSize = GpuBatchSize;
 
@@ -212,7 +205,7 @@ public sealed class PrimeTester
 		}
 
 		gpu.Dispose();
-		limiter.Dispose();
+		GpuPrimeWorkLimiter.Release();
 	}
 
 	internal static class PrimeTesterGpuContextPool
@@ -221,23 +214,16 @@ public sealed class PrimeTester
 		private static readonly object WarmUpLock = new();
 		private static int WarmedLeaseCount;
 		private static readonly object SharedTablesLock = new();
-		private static readonly Dictionary<string, SharedHeuristicGpuTables> SharedTables = new();
+		private static readonly Dictionary<Accelerator, SharedHeuristicGpuTables> SharedTables = new(AcceleratorReferenceComparer.Instance);
-
-		private static string CreateAcceleratorKey(Accelerator accelerator)
-		{
-			return accelerator.AcceleratorType + ":" + accelerator.Name;
-		}
 
 		private static SharedHeuristicGpuTables RentSharedTables(Accelerator accelerator)
 		{
-			string key = CreateAcceleratorKey(accelerator);
-
 			lock (SharedTablesLock)
 			{
-				if (!SharedTables.TryGetValue(key, out var tables))
+				if (!SharedTables.TryGetValue(accelerator, out var tables))
 				{
-					tables = new SharedHeuristicGpuTables(accelerator, key);
+					tables = new SharedHeuristicGpuTables(accelerator);
-					SharedTables.Add(key, tables);
+					SharedTables.Add(accelerator, tables);
 				}
 
 				tables.AddReference();
@@ -258,6 +244,14 @@ public sealed class PrimeTester
 			}
 		}
 
+		internal static void EnsureStaticTables(KernelContainer kernels, Accelerator accelerator)
+		{
+			var tables = RentSharedTables(accelerator);
+			var heuristicTables = tables.CreateHeuristicDivisorTables();
+			HeuristicGpuDivisorTables.InitializeShared(in heuristicTables);
+			ReleaseSharedTables(tables);
+		}
+
 		internal static void WarmUp(int target, int minBufferCapacity)
 		{
 			if (target <= 0)
@@ -272,12 +266,16 @@ public sealed class PrimeTester
 					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;
 			}
 		}
@@ -319,6 +317,8 @@ public sealed class PrimeTester
 				}
 
 				SharedTables.Clear();
+				var emptyTables = default(HeuristicGpuDivisorTables);
+				HeuristicGpuDivisorTables.InitializeShared(in emptyTables);
 			}
 
 			WarmedLeaseCount = 0;
@@ -354,6 +354,20 @@ public sealed class PrimeTester
 
 				WarmedLeaseCount = Pool.Count;
 			}
+
+			lock (SharedTablesLock)
+			{
+				if (SharedTables.Remove(accelerator, out var tables))
+				{
+					tables.Dispose();
+				}
+
+				if (SharedTables.Count == 0)
+				{
+					var emptyTables = default(HeuristicGpuDivisorTables);
+					HeuristicGpuDivisorTables.InitializeShared(in emptyTables);
+				}
+			}
 		}
 
 		internal sealed class PrimeTesterGpuContextLease : IDisposable
@@ -365,8 +379,11 @@ public sealed class PrimeTester
 
 			internal readonly Context AcceleratorContext;
 			public readonly Accelerator Accelerator;
-			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> HeuristicTrialDivisionKernel;
+			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 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!;
@@ -401,36 +418,17 @@ public sealed class PrimeTester
 
 			public HeuristicGpuDivisorTables HeuristicGpuTables => _heuristicDivisorTables;
 
-			private static readonly Context _sharedContext = Context.CreateDefault();
-			private static readonly Device _sharedDevice = _sharedContext.GetPreferredDevice(false);
-			private static readonly Accelerator _sharedAccelerator = _sharedDevice.CreateAccelerator(_sharedContext);
-
 			internal PrimeTesterGpuContextLease(int minBufferCapacity)
 			{
-				AcceleratorContext = _sharedContext;// Context.CreateDefault();
+				AcceleratorContext = SharedGpuContext.Context;
-				Accelerator = _sharedAccelerator; //AcceleratorContext.GetPreferredDevice(false).CreateAccelerator(AcceleratorContext);
+				Accelerator = SharedGpuContext.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);
+				// GpuStaticTableInitializer.EnsureStaticTables(Accelerator);
-				HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
+				// 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 = new HeuristicGpuDivisorTables(
+				_heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
-					_sharedTables.HeuristicCombinedDivisorsEnding1.View,
-					_sharedTables.HeuristicCombinedDivisorsEnding3.View,
-					_sharedTables.HeuristicCombinedDivisorsEnding7.View,
-					_sharedTables.HeuristicCombinedDivisorsEnding9.View,
-					_sharedTables.HeuristicCombinedDivisorSquaresEnding1.View,
-					_sharedTables.HeuristicCombinedDivisorSquaresEnding3.View,
-					_sharedTables.HeuristicCombinedDivisorSquaresEnding7.View,
-					_sharedTables.HeuristicCombinedDivisorSquaresEnding9.View,
-					_sharedTables.HeuristicGroupADivisors.View,
-					_sharedTables.HeuristicGroupADivisorSquares.View,
-					_sharedTables.HeuristicGroupBDivisorsEnding1.View,
-					_sharedTables.HeuristicGroupBDivisorSquaresEnding1.View,
-					_sharedTables.HeuristicGroupBDivisorsEnding7.View,
-					_sharedTables.HeuristicGroupBDivisorSquaresEnding7.View,
-					_sharedTables.HeuristicGroupBDivisorsEnding9.View,
-					_sharedTables.HeuristicGroupBDivisorSquaresEnding9.View);
 				HeuristicGpuDivisorTables.InitializeShared(in _heuristicDivisorTables);
 
 				EnsureCapacity(minBufferCapacity);
@@ -481,13 +479,27 @@ public sealed class PrimeTester
 			}
 		}
 
+		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 SharedHeuristicGpuTables(Accelerator accelerator, string key)
+			internal SharedHeuristicGpuTables(Accelerator accelerator)
 			{
-				Key = key;
 
 				var primesDefault = DivisorGenerator.SmallPrimes;
 				DevicePrimesDefault = accelerator.Allocate1D<uint>(primesDefault.Length);
@@ -568,7 +580,6 @@ public sealed class PrimeTester
 				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; }
@@ -596,6 +607,27 @@ public sealed class PrimeTester
 			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++;