Kernels made part of the accelerators for better performance.

PerfectNumbers.Core/Gpu/Accelerators/HeuristicPrimeTesterAccelerator.cs

@@ -104,22 +104,7 @@ internal sealed class HeuristicPrimeTesterAccelerator
 		return cached;
 	}
 
-	[ThreadStatic]
+	public readonly Kernel HeuristicTrialDivisionKernel;
-	public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>>? HeuristicTrialDivisionKernel;
-
-	internal static Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables> GetHeuristicTrialDivisionKernel(Accelerator accelerator)
-	{
-		var pool = HeuristicTrialDivisionKernel ??= [];
-		if (pool.TryGetValue(accelerator, out var cached))
-		{
-			return cached;
-		}
-
-		cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>>();
-		pool[accelerator] = cached;
-		return cached;
-	}
-
 
 	public MemoryBuffer1D<ulong, Stride1D.Dense> Input = null!;
 	public MemoryBuffer1D<byte, Stride1D.Dense> Output = null!;
@@ -152,17 +137,19 @@ internal sealed class HeuristicPrimeTesterAccelerator
 		Context = accelerator.Context;
 
 		Accelerator = accelerator;
-		AcceleratorStream stream = accelerator.CreateStream();
+		EnsureCapacity(minBufferCapacity);
-		// var kernels = _kernels[index]; //GpuKernelPool.Kernels;
+
+		HeuristicTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel));
+
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
+		
 		// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
 
 		_sharedTables = Accelerators.HeuristicGpuTables.EnsureStaticTables(accelerator, stream);
-		var heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
+		_heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
-		_heuristicDivisorTables = heuristicDivisorTables;
 
-		EnsureCapacity(minBufferCapacity);
 		stream.Synchronize();
-		stream.Dispose();
+		AcceleratorStreamPool.Return(stream); 
 	}
 
 	public void EnsureCapacity(int minCapacity)
@@ -172,16 +159,13 @@ internal sealed class HeuristicPrimeTesterAccelerator
 			return;
 		}
 
-		Input?.Dispose();
+		Input.Dispose();
-		Output?.Dispose();
+		Output.Dispose();
 		HeuristicFlag?.Dispose();
 
-		// lock (Accelerator)
+		Input = Accelerator.Allocate1D<ulong>(minCapacity);
-		{
+		Output = Accelerator.Allocate1D<byte>(minCapacity);
-			Input = Accelerator.Allocate1D<ulong>(minCapacity);
+		HeuristicFlag = Accelerator.Allocate1D<int>(minCapacity);
-			Output = Accelerator.Allocate1D<byte>(minCapacity);
-			HeuristicFlag = Accelerator.Allocate1D<int>(minCapacity);
-		}
 
 		BufferCapacity = minCapacity;
 	}

PerfectNumbers.Core/Gpu/Accelerators/PrimeOrderCalculatorAccelerator.cs

@@ -110,6 +110,8 @@ public sealed class PrimeOrderCalculatorAccelerator
 	public readonly Kernel ConvertToStandardKernel;
 	public readonly Kernel KeepMontgomeryKernel;
 	public readonly Kernel PollardRhoKernel;
+	public Kernel SharesFactorKernel;
+
 
 	public void EnsureCapacity(int factorsCount, int primeTesterCapacity)
 	{
@@ -199,6 +201,8 @@ public sealed class PrimeOrderCalculatorAccelerator
 		SmallPrimeSieveKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel));
 
 		PollardRhoKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.TryPollardRhoKernel));
+
+		SharesFactorKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel)); ;
 	}
 
 	public void Dispose()

PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs

@@ -228,7 +228,7 @@ public sealed class HeuristicCombinedPrimeTester
 		var gpu = HeuristicPrimeTesterAccelerator.Rent(1);
 		var accelerator = gpu.Accelerator;
 		var stream = gpu.Stream;
-		var kernel = HeuristicPrimeTesterAccelerator.HeuristicTrialDivisionKernel;
+		var kernel = gpu.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
 		var flagView = flagView1D.AsContiguous();
 
@@ -240,8 +240,9 @@ public sealed class HeuristicCombinedPrimeTester
 		stream!.Synchronize();
 
 
-		kernel(
+		kernel.Launch(
 				stream,
+				1,
 				divisorLength,
 				flagView,
 				n,

PerfectNumbers.Core/HeuristicPrimeTester.cs

@@ -243,9 +243,8 @@ public sealed class HeuristicPrimeTester
 	{
 		// GpuPrimeWorkLimiter.Acquire();
 		var gpu = HeuristicPrimeTesterAccelerator.Rent(1);
-		var accelerator = gpu.Accelerator;
 		var stream = gpu.Stream;
-		var kernel = HeuristicPrimeTesterAccelerator.GetHeuristicTrialDivisionKernel(accelerator);
+		var kernel = gpu.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
 		var flagView = flagView1D.AsContiguous();
 
@@ -255,8 +254,9 @@ public sealed class HeuristicPrimeTester
 		int groupALength = HeuristicPrimeSieves.GroupADivisors.Length;
 
 		flagView1D.CopyFromCPU(stream, ref compositeFlag, 1);
-		kernel(
+		kernel.Launch(
 				stream!,
+				1,
 				groupALength,
 				flagView,
 				n,
@@ -292,8 +292,9 @@ public sealed class HeuristicPrimeTester
 
 				compositeFlag = 0;
 				flagView1D.CopyFromCPU(stream, ref compositeFlag, 1);
-				kernel(
+				kernel.Launch(
-                    stream,
+					stream,
+						1,
 						divisorLength,
 						flagView,
 						n,

PerfectNumbers.Core/PrimeTester.cs

@@ -278,10 +278,10 @@ public sealed class PrimeTester
 		// and implement it.
 
 		int length = values.Length;
-		ArrayPool<ulong> pool = ThreadStaticPools.UlongPool;
 
 		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
 		var accelerator = gpu.Accelerator;
+		var kernel = gpu.SharesFactorKernel;
 		gpu.EnsureCapacity(0, length);
 		MemoryBuffer1D<ulong, Stride1D.Dense>? inputBuffer = gpu.PrimeTestInput;
 		MemoryBuffer1D<byte, Stride1D.Dense>? resultBuffer = gpu.OutputByte;
@@ -289,8 +289,7 @@ public sealed class PrimeTester
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
 		inputBuffer.View.CopyFromCPU(stream, values);
 
-		var kernel = GetSharesFactorKernel(accelerator);
+		kernel.Launch(stream, 1, length, inputBuffer.View, resultBuffer.View);
-		kernel(stream, length, inputBuffer.View, resultBuffer.View);
 		resultBuffer.View.CopyToCPU(stream, in results);
 		stream.Synchronize();
 
@@ -298,19 +297,6 @@ public sealed class PrimeTester
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 	}
 
-	private static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<byte>> GetSharesFactorKernel(Accelerator accelerator)
-	{
-		var pool = _kernel ?? [];
-		if (pool.TryGetValue(accelerator, out var cached))
-		{
-			return cached;
-		}
-
-		cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<byte>>>();
-		pool[accelerator] = cached;
-		return cached;
-	}
-
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	private static ulong BinaryGcd(ulong u, ulong v)
 	{