ConcurrentDictionary removed. Kernel made dependent on the shared GPU context.

PerfectNumbers.Core.Tests/Gpu/GpuKernelCompilationTests.cs

@@ -349,7 +349,7 @@ public class GpuKernelCompilationTests
     private static void CompileGpuKernelPoolKernels(Accelerator accelerator)
     {
         _ = accelerator;
-        var lease = GpuKernelPool.GetKernel();
+        var lease = GpuKernelPool.Rent();
         try
         {
             _ = lease.OrderKernel;
@@ -363,7 +363,7 @@ public class GpuKernelCompilationTests
             lease.Dispose();
         }
 
-        var orderLease = GpuKernelPool.GetKernel();
+        var orderLease = GpuKernelPool.Rent();
         try
         {
             _ = orderLease.OrderKernel;

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -16,7 +16,7 @@ public sealed class DivisorCycleCache
 	private const byte ByteOne = 1;
 	private const int StackBufferThreshold = 128;
 
-    private readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>> _gpuKernelCache = new(AcceleratorReferenceComparer.Instance);
+	private readonly Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> _gpuKernel = LoadKernel(SharedGpuContext.Accelerator);
 	private volatile ulong[] _snapshot;
 	private volatile bool _useGpuGeneration = true;
 	private readonly int _divisorCyclesBatchSize;
@@ -224,20 +224,55 @@ public sealed class DivisorCycleCache
 		}
 	}
 
+	private const int DefaultCapacity = 128;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? divisorBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? powBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? orderBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<ulong, Stride1D.Dense>? resultBuffer;
+
+	[ThreadStatic]
+	private static MemoryBuffer1D<byte, Stride1D.Dense>? statusBuffer;
+
+	private static void EnsureCapacity(int requiredCapacity)
+	{
+		if (requiredCapacity <= (divisorBuffer?.Length ?? 0))
+		{
+			return;
+		}
+
+		if (divisorBuffer != null)
+		{
+			divisorBuffer.Dispose();
+			powBuffer!.Dispose();
+			orderBuffer!.Dispose();
+			resultBuffer!.Dispose();
+			statusBuffer!.Dispose();			
+		}
+
+		divisorBuffer = SharedGpuContext.Accelerator.Allocate1D<ulong>(requiredCapacity);
+		powBuffer = SharedGpuContext.Accelerator.Allocate1D<ulong>(requiredCapacity);
+		orderBuffer = SharedGpuContext.Accelerator.Allocate1D<ulong>(requiredCapacity);
+		resultBuffer = SharedGpuContext.Accelerator.Allocate1D<ulong>(requiredCapacity);
+		statusBuffer = SharedGpuContext.Accelerator.Allocate1D<byte>(requiredCapacity);
+	}
+
 	private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
 	{
 		int length = divisors.Length;
-        var gpuLease = GpuKernelPool.GetKernel();
+		var gpuLease = GpuKernelPool.Rent();
 
 		Accelerator accelerator = gpuLease.Accelerator;
 		var stream = gpuLease.Stream;
-        var kernel = _gpuKernelCache.GetOrAdd(accelerator, LoadKernel);
-
-        using MemoryBuffer1D<ulong, Stride1D.Dense> divisorBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<ulong, Stride1D.Dense> powBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<ulong, Stride1D.Dense> orderBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = accelerator.Allocate1D<ulong>(length);
-        using MemoryBuffer1D<byte, Stride1D.Dense> statusBuffer = accelerator.Allocate1D<byte>(length);
+		var kernel = _gpuKernel;
+		EnsureCapacity(length);
 
 		ulong[]? rentedPow = null;
 		Span<ulong> powSpan = length <= StackBufferThreshold
@@ -260,7 +295,7 @@ public sealed class DivisorCycleCache
 			: new Span<byte>(rentedStatus = ThreadStaticPools.BytePool.Rent(length), 0, length);
 
 		ref ulong divisorRef = ref MemoryMarshal.GetReference(divisors);
-        divisorBuffer.View.CopyFromCPU(stream, ref divisorRef, length);
+		divisorBuffer!.View.CopyFromCPU(stream, ref divisorRef, length);
 
 		for (int i = 0; i < length; i++)
 		{
@@ -277,10 +312,10 @@ public sealed class DivisorCycleCache
 			statusSpan[i] = ByteZero;
 		}
 
-        powBuffer.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(powSpan), length);
-        orderBuffer.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(orderSpan), length);
-        resultBuffer.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
-        statusBuffer.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(statusSpan), length);
+		powBuffer!.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(powSpan), length);
+		orderBuffer!.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(orderSpan), length);
+		resultBuffer!.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
+		statusBuffer!.View.CopyFromCPU(stream, ref MemoryMarshal.GetReference(statusSpan), length);
 
 		int pending;
 		do
@@ -303,6 +338,7 @@ public sealed class DivisorCycleCache
 
 		resultBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
 		stream.Synchronize();
+
 		resultSpan.CopyTo(destination);
 		if (rentedPow is not null)
 		{

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -83,7 +83,7 @@ public readonly struct ResiduePrimeViews(
 	public readonly ArrayView1D<ulong, Stride1D.Dense> LastSevenPow2 = lastSevenPow2;
 }
 
-public readonly struct SmallPrimeFactorTables(
+public readonly struct SmallPrimeFactorViews(
 	MemoryBuffer1D<uint, Stride1D.Dense> primes,
 	MemoryBuffer1D<ulong, Stride1D.Dense> squares,
 	int count)
@@ -184,18 +184,13 @@ public class GpuKernelPool
 		return new ResiduePrimeViews(deviceLastOne.View, deviceLastSeven.View, deviceLastOnePow2.View, deviceLastSevenPow2.View);
 	}
 
-	public static SmallPrimeFactorTables EnsureSmallPrimeFactorTables(KernelContainer kernels, Accelerator accelerator, AcceleratorStream stream)
+	public static SmallPrimeFactorViews EnsureSmallPrimeFactorTables(KernelContainer kernels, Accelerator accelerator, AcceleratorStream stream)
 	{
 		if (kernels.SmallPrimeFactorsPrimes is { } primeBuffer && kernels.SmallPrimeFactorsSquares is { } squareBuffer)
 		{
-			return new SmallPrimeFactorTables(primeBuffer, squareBuffer, (int)primeBuffer.Length);
+			return new SmallPrimeFactorViews(primeBuffer, squareBuffer, (int)primeBuffer.Length);
 		}
 
-		// if (kernels.SmallPrimeFactorsPrimes is { } existingPrimes && kernels.SmallPrimeFactorsSquares is { } existingSquares)
-		// {
-		// 	return new SmallPrimeFactorTables(existingPrimes, existingSquares, (int)existingPrimes.Length);
-		// }
-
 		var hostPrimes = PrimesGenerator.SmallPrimes;
 		var hostSquares = PrimesGenerator.SmallPrimesPow2;
 
@@ -207,17 +202,13 @@ public class GpuKernelPool
 		kernels.SmallPrimeFactorsPrimes = devicePrimes;
 		kernels.SmallPrimeFactorsSquares = deviceSquares;
 
-		return new SmallPrimeFactorTables(devicePrimes, deviceSquares, hostPrimes.Length);
+		return new SmallPrimeFactorViews(devicePrimes, deviceSquares, hostPrimes.Length);
 	}
 
-
-
-
-
-	public static GpuKernelLease GetKernel()
+	public static GpuKernelLease Rent()
 	{
 		GpuPrimeWorkLimiter.Acquire();
-		var gpu = Rent();
+		var gpu = GpuContextPool.Rent();
 		var stream = gpu.Accelerator.CreateStream();
 		var kernels = GetKernels(gpu.Accelerator, stream);
 		return GpuKernelLease.Rent(gpu, stream, kernels);
@@ -229,7 +220,7 @@ public class GpuKernelPool
 	/// <param name="action">Action to run with (Accelerator, Stream).</param>
 	public static void Run(Action<Accelerator, AcceleratorStream> action)
 	{
-		var lease = GetKernel();
+		var lease = Rent();
 		var accelerator = lease.Accelerator;
 		var stream = lease.Stream;
 		action(accelerator, stream);

PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs

@@ -95,7 +95,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
         ulong processedCount;
         ulong lastProcessed;
 
-        var gpuLease = GpuKernelPool.GetKernel();
+        var gpuLease = GpuKernelPool.Rent();
         var accelerator = gpuLease.Accelerator;
         var stream = gpuLease.Stream;
 
@@ -501,7 +501,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
         internal DivisorScanSession(MersenneNumberDivisorByDivisorGpuTester owner)
         {
             _owner = owner;
-            _lease = GpuKernelPool.GetKernel();
+            _lease = GpuKernelPool.Rent();
             _accelerator = _lease.Accelerator;
             _stream = _lease.Stream;
             _capacity = Math.Max(1, owner._gpuBatchSize);

PerfectNumbers.Core/Gpu/MersenneNumberIncrementalGpuTester.cs

@@ -16,7 +16,7 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
     {
         throw new NotImplementedException($"GPU incremental scanning requires the device cycle heuristics implementation (kernel: {_kernelType}, GPU order: {_useGpuOrder}).");
 
-        var gpuLease = GpuKernelPool.GetKernel();
+        var gpuLease = GpuKernelPool.Rent();
         var accelerator = gpuLease.Accelerator;
         var stream = gpuLease.Stream;
         int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy

PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs

@@ -9,7 +9,7 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 
 	public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
 	{
-		var gpuLease = GpuKernelPool.GetKernel();
+		var gpuLease = GpuKernelPool.Rent();
 		var accelerator = gpuLease.Accelerator;
 		var stream = gpuLease.Stream;
 		int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy and avoids TDR

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -198,7 +198,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
         try
         {
-            var lease = GpuKernelPool.GetKernel();
+            var lease = GpuKernelPool.Rent();
             Accelerator accelerator = lease.Accelerator;
             AcceleratorStream stream = lease.Stream;
 
@@ -365,7 +365,7 @@ internal static partial class PrimeOrderGpuHeuristics
     {
         order = 0UL;
 
-        var lease = GpuKernelPool.GetKernel();
+        var lease = GpuKernelPool.Rent();
         Accelerator accelerator = lease.Accelerator;
         AcceleratorStream stream = lease.Stream;
 
@@ -526,7 +526,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
     private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
     {
-        var lease = GpuKernelPool.GetKernel();
+        var lease = GpuKernelPool.Rent();
         Accelerator accelerator = lease.Accelerator;
         AcceleratorStream stream = lease.Stream;
         var kernel = GetPow2ModKernel(accelerator);
@@ -581,7 +581,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
         GpuUInt128[]? rentedExponents = null;
         GpuUInt128[]? rentedResults = null;
-        var lease = GpuKernelPool.GetKernel();
+        var lease = GpuKernelPool.Rent();
 		ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
 
         try

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -804,7 +804,7 @@ public class MersenneDivisorCycles
 
 		divisors = pool.Rent(count);
 		outCycles = pool.Rent(count);
-		var lease = GpuKernelPool.GetKernel();
+		var lease = GpuKernelPool.Rent();
 		var accelerator = lease.Accelerator;
 		var stream = lease.Stream;
 		var loaded = accelerator.LoadAutoGroupedStreamKernel<

PerfectNumbers.Core/MersenneNumberTester.cs

@@ -147,7 +147,7 @@ public sealed class MersenneNumberTester(
 
 		UInt128[] qs = qsBuffer;
 
-		var gpuLease = GpuKernelPool.GetKernel();
+		var gpuLease = GpuKernelPool.Rent();
 		var accelerator = gpuLease.Accelerator;
 		var stream = gpuLease.Stream;
 		var orderKernel = gpuLease.OrderKernel!;

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -19,14 +19,14 @@ internal static partial class PrimeOrderCalculator
 		Span<int> exponentBuffer = exponentBufferArray.AsSpan(0, GpuSmallPrimeFactorSlots);
 		remaining = value;
 
-		var lease = GpuKernelPool.GetKernel();
+		var lease = GpuKernelPool.Rent();
 		Accelerator accelerator = lease.Accelerator;
 		var stream = lease.Stream;
 		KernelContainer kernels = lease.Kernels;
 
 		ScratchBuffer scratch = GpuScratchBufferPool.Rent(accelerator, GpuSmallPrimeFactorSlots, 0);
 
-		SmallPrimeFactorTables tables = GpuKernelPool.EnsureSmallPrimeFactorTables(kernels, accelerator, stream);
+		SmallPrimeFactorViews tables = GpuKernelPool.EnsureSmallPrimeFactorTables(kernels, accelerator, stream);
 
 		if (kernels.SmallPrimeFactorsPrimes is null) throw new NullReferenceException($"{nameof(kernels.SmallPrimeFactorsPrimes)} is null");
 		if (kernels.SmallPrimeFactorsSquares is null) throw new NullReferenceException($"{nameof(kernels.SmallPrimeFactorsSquares)} is null");
@@ -100,7 +100,7 @@ internal static partial class PrimeOrderCalculator
 			factorSpan[i] = factors[i].Value;
 		}
 
-		GpuKernelLease lease = GpuKernelPool.GetKernel();
+		GpuKernelLease lease = GpuKernelPool.Rent();
 		Accelerator accelerator = lease.Accelerator;
 
 		var stream = lease.Stream;

PerfectNumbers.Core/ULongExtensions.Gpu.cs

@@ -440,7 +440,7 @@ public static partial class ULongExtensions
 
 	private static class Pow2MontgomeryGpuExecutor
 	{
-		private static readonly ConcurrentDictionary<Accelerator, Pow2MontgomeryKernelGroup> KernelCache = new();
+		private static readonly Pow2MontgomeryKernelGroup Kernel = Pow2MontgomeryKernelGroup.Create(SharedGpuContext.Accelerator);
 
 		public static ulong ExecuteKeep(ulong exponent, in MontgomeryDivisorData divisor)
 		{
@@ -452,24 +452,25 @@ public static partial class ULongExtensions
 			return Execute(exponent, divisor, useKeepKernel: false);
 		}
 
+		[ThreadStatic]
+		private static MemoryBuffer1D<ulong, Stride1D.Dense>? _exponentBuffer;
+		[ThreadStatic]
+		private static MemoryBuffer1D<ulong, Stride1D.Dense>? _resultBuffer;
+
 		private static ulong Execute(ulong exponent, in MontgomeryDivisorData divisor, bool useKeepKernel)
 		{
 			ulong result = 0UL;
 
-			GpuKernelLease lease = GpuKernelPool.GetKernel();
+			GpuKernelLease lease = GpuKernelPool.Rent();
 
 			Accelerator accelerator = lease.Accelerator;
 			AcceleratorStream stream = lease.Stream;
-			// Keep this commented out. It should never happen in production code.
-			// if (accelerator.AcceleratorType == AcceleratorType.CPU)
-			// {
-			//         return 0UL;
-			// }
+			var kernelGroup = Kernel;
 
-			var kernelGroup = KernelCache.GetOrAdd(accelerator, static accel => Pow2MontgomeryKernelGroup.Create(accel));
+			EnsureCapacity();
 
-			var exponentBuffer = accelerator.Allocate1D<ulong>(1);
-			var resultBuffer = accelerator.Allocate1D<ulong>(1);
+			MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = _exponentBuffer!;
+			MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = _resultBuffer!;
 
 			exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
 			// We don't need to worry about any left-overs here.
@@ -483,19 +484,19 @@ public static partial class ULongExtensions
 			resultBuffer.Dispose();
 			exponentBuffer.Dispose();
 			lease.Dispose();
-			// Keep this commented. We don't want to catch any exceptions. All should crash the scanner.
-			// catch (AcceleratorException)
-			// {
-			//         return 0UL;
-			// }
-			// catch (NotSupportedException)
-			// {
-			//         return 0UL;
-			// }
-			// Intentionally avoid exception handling here; any accelerator failure should crash the scanner.
 			return result;
 		}
 
+		private static void EnsureCapacity()
+		{
+			if (_exponentBuffer == null)
+			{
+				Accelerator accelerator = SharedGpuContext.Accelerator;
+				_exponentBuffer = accelerator.Allocate1D<ulong>(1);
+				_resultBuffer = accelerator.Allocate1D<ulong>(1);
+			}
+		}
+
 		private sealed class Pow2MontgomeryKernelGroup
 		{
 			internal Pow2MontgomeryKernelGroup(