PrimeTesterByLastDigitAccelerator integrated into Pow2MontgomeryGpuCalculator for better performance.

PerfectNumbers.Core/Gpu/AcceleratorPool.cs

@@ -28,7 +28,7 @@ namespace PerfectNumbers.Core.Gpu
 			// Intentionally left empty - there's nothing to do here. We're not really renting anything.
 		}
 
-		public static readonly AcceleratorPool Shared = new(PerfectNumberConstants.RollingAccelerators);
+		public static readonly AcceleratorPool Shared = new(1);
 		private readonly int Capacity;
 
 		public AcceleratorPool(int capacity)
@@ -38,7 +38,20 @@ namespace PerfectNumbers.Core.Gpu
 			var device = context.GetPreferredDevice(false);
 			Context = context;
 			Device = device;
-			Accelerators = [.. Enumerable.Range(0, capacity).Select(_ => device.CreateAccelerator(context))];
+			int acceleratorsInContext = 0;
+			Accelerators = [.. Enumerable.Range(0, capacity).Select(_ =>
+			{
+				Accelerator accelerator = device.CreateAccelerator(context);
+				acceleratorsInContext++;
+				if (acceleratorsInContext >= 128)
+				{
+					context = Context.CreateDefault();
+					device = context.GetPreferredDevice(false);
+					acceleratorsInContext = 0;
+				}
+
+				return accelerator;
+			})];
 		}
 	}
 }

PerfectNumbers.Core/Gpu/Accelerators/HeuristicGpuTables.cs

@@ -6,7 +6,7 @@ namespace PerfectNumbers.Core.Gpu.Accelerators;
 
 internal sealed class HeuristicGpuTables
 {
-	private static readonly ConcurrentDictionary<Accelerator, HeuristicGpuTables> _sharedTables = new(20_480, PerfectNumberConstants.RollingAccelerators);
+	private static readonly ConcurrentDictionary<Accelerator, HeuristicGpuTables> _sharedTables = new(20_480, 1);
 
 	internal static HeuristicGpuTables EnsureStaticTables(Accelerator accelerator, AcceleratorStream stream)
 		=> _sharedTables.GetOrAdd(accelerator, _ => new HeuristicGpuTables(accelerator, stream));

PerfectNumbers.Core/Gpu/Accelerators/LastDigitGpuTables.cs

@@ -6,7 +6,7 @@ namespace PerfectNumbers.Core.Gpu.Accelerators;
 
 internal sealed class LastDigitGpuTables
 {
-	private static readonly ConcurrentDictionary<Accelerator, LastDigitGpuTables> _sharedTables = new(20_480, PerfectNumberConstants.RollingAccelerators);
+	private static readonly ConcurrentDictionary<Accelerator, LastDigitGpuTables> _sharedTables = new(20_480, PerfectNumberConstants.RollingAccelerators * 2);
 
 	internal static LastDigitGpuTables EnsureStaticTables(Accelerator accelerator, AcceleratorStream stream)
 		=> _sharedTables.GetOrAdd(accelerator, _ => new LastDigitGpuTables(accelerator, stream));

PerfectNumbers.Core/Gpu/Accelerators/Pow2MontgomeryAccelerator.cs

@@ -12,20 +12,21 @@ public sealed class Pow2MontgomeryAccelerator
 	[ThreadStatic]
 	private static Queue<Pow2MontgomeryAccelerator>? _pool;
 
-	private static AcceleratorPool _accelerators = new(PerfectNumberConstants.RollingAccelerators);
+	private static AcceleratorPool _accelerators = new(PerfectNumberConstants.RollingAccelerators << 1);
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
-	public static Pow2MontgomeryAccelerator Rent()
+	public static Pow2MontgomeryAccelerator Rent(int primeTesterCapacity)
 	{
 		var pool = _pool ??= [];
 		if (!pool.TryDequeue(out var gpu))
 		{
 			Accelerator accelerator = _accelerators.Rent();
-			gpu = new(accelerator, PerfectNumberConstants.DefaultFactorsBuffer);
+			gpu = new(accelerator, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity);
 		}
 		else
 		{
 			gpu.Stream = gpu.Accelerator.CreateStream();
+			gpu.EnsureCapacity(PerfectNumberConstants.DefaultFactorsBuffer,primeTesterCapacity);
 		}
 
 		return gpu;
@@ -39,13 +40,69 @@ public sealed class Pow2MontgomeryAccelerator
 		_pool!.Enqueue(gpu);
 	}
 
+	internal static void Clear()
+	{
+		Queue<Pow2MontgomeryAccelerator>? pool = _pool;
+		if (pool != null)
+		{
+			while (pool.TryDequeue(out var lease))
+			{
+				lease.Dispose();
+			}
+		}
+	}
+
+	internal static void DisposeAll()
+	{
+		Queue<Pow2MontgomeryAccelerator>? pool = _pool;
+		if (pool != null)
+		{
+			while (pool.TryDequeue(out var lease))
+			{
+				lease.Dispose();
+			}
+		}
+	}
+
+	internal static void WarmUp()
+	{
+		Accelerator[]? accelerators = _accelerators.Accelerators;
+		int acceleratorCount = accelerators.Length;
+		for (var i = 0; i < acceleratorCount; i++)
+		{
+			Console.WriteLine($"Preparing accelerator {i}...");
+			var accelerator = accelerators[i];
+			AcceleratorStream stream = accelerator.CreateStream();
+			LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
+			// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
+			// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
+			// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
+			// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
+			stream.Synchronize();
+			stream.Dispose();
+		}
+	}
+
 	#endregion
 
 	public readonly Accelerator Accelerator;
 	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Input;
 	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Output;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> PrimeTestInput;
+	public MemoryBuffer1D<byte, Stride1D.Dense> PrimeTestOutput;
+	public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag;
 	public MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> Pow2ModEntriesToTestOnDevice;
 	public Dictionary<ulong, int> Pow2ModEntriesToTestOnHost = [];
+	public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
+	public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
+	public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree;
+	public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
+	public readonly Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> SmallPrimeSieveKernel;
+
 	public AcceleratorStream? Stream;
 
 	public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernel;
@@ -53,7 +110,23 @@ public sealed class Pow2MontgomeryAccelerator
 
 	public readonly Kernel CheckFactorsKernel;
 
-	public void EnsureCapacity(int factorsCount)
+	[ThreadStatic]
+	private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>? SmallPrimeSieveKernelPool;
+
+	internal static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetSmallPrimeSieveKernel(Accelerator accelerator)
+	{
+		var pool = SmallPrimeSieveKernelPool ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
+		pool[accelerator] = cached;
+		return cached;
+	}
+
+	public void EnsureCapacity(int factorsCount, int primeTesterCapacity)
 	{
 		if (factorsCount > Pow2ModEntriesToTestOnDevice.Length)
 		{
@@ -61,16 +134,45 @@ public sealed class Pow2MontgomeryAccelerator
 			Pow2ModEntriesToTestOnDevice.Dispose();
 			Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
 		}
+
+		if (primeTesterCapacity > PrimeTestInput.Length)
+		{
+
+			PrimeTestInput?.Dispose();
+			PrimeTestOutput?.Dispose();
+			HeuristicFlag?.Dispose();
+
+			// lock (Accelerator)
+			{
+				PrimeTestInput = Accelerator.Allocate1D<ulong>(primeTesterCapacity);
+				PrimeTestOutput = Accelerator.Allocate1D<byte>(primeTesterCapacity);
+				HeuristicFlag = Accelerator.Allocate1D<int>(primeTesterCapacity);
+			}
+		}
 	}
 
-	public Pow2MontgomeryAccelerator(Accelerator accelerator, int factorsCount)
+	public Pow2MontgomeryAccelerator(Accelerator accelerator, int factorsCount, int primeTesterCapacity)
 	{
 		Accelerator = accelerator;
-		Stream = accelerator.CreateStream();
+		AcceleratorStream stream = accelerator.CreateStream();
+		Stream = stream;
 
 		Input = accelerator.Allocate1D<ulong>(1);
 		Output = accelerator.Allocate1D<ulong>(1);
 		Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
+		PrimeTestInput = Accelerator.Allocate1D<ulong>(primeTesterCapacity);
+		PrimeTestOutput = Accelerator.Allocate1D<byte>(primeTesterCapacity);
+		HeuristicFlag = Accelerator.Allocate1D<int>(primeTesterCapacity);
+
+		var sharedTables = LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
+		DevicePrimesLastOne = sharedTables.DevicePrimesLastOne;
+		DevicePrimesLastSeven = sharedTables.DevicePrimesLastSeven;
+		DevicePrimesLastThree = sharedTables.DevicePrimesLastThree;
+		DevicePrimesLastNine = sharedTables.DevicePrimesLastNine;
+		DevicePrimesPow2LastOne = sharedTables.DevicePrimesPow2LastOne;
+		DevicePrimesPow2LastSeven = sharedTables.DevicePrimesPow2LastSeven;
+		DevicePrimesPow2LastThree = sharedTables.DevicePrimesPow2LastThree;
+		DevicePrimesPow2LastNine = sharedTables.DevicePrimesPow2LastNine;
 
 		var keepKernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery);
 		var keepLauncher = KernelUtil.GetKernel(keepKernel);
@@ -84,6 +186,8 @@ public sealed class Pow2MontgomeryAccelerator
 
 		var checkFactorsKernel = accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel);
 		CheckFactorsKernel = KernelUtil.GetKernel(checkFactorsKernel);
+
+		SmallPrimeSieveKernel = (Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>)KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>().Clone();
 	}
 
 	public void Dispose()
@@ -92,7 +196,17 @@ public sealed class Pow2MontgomeryAccelerator
 		Output.Dispose();
 		Stream?.Dispose();
 		Pow2ModEntriesToTestOnDevice.Dispose();
-		// Accelerator is shared with other threads
+		PrimeTestInput?.Dispose();
+		PrimeTestOutput?.Dispose();
+		HeuristicFlag?.Dispose();
+
+		// Stream is disposed upon return to the queue, so we don't need to dispose it here.
+		// Stream.Dispose();
+
+		// These resources are shared between GPU leases
+		// ReleaseSharedTables(_sharedTables);
+		// Accelerator.Dispose();
+		// Context.Dispose();
 		// Accelerator.Dispose();
 	}
 }

PerfectNumbers.Core/Gpu/Accelerators/PrimeTesterByLastDigitAccelerator.cs

@@ -1,192 +0,0 @@
-using System.Runtime.CompilerServices;
-using ILGPU;
-using ILGPU.Runtime;
-
-namespace PerfectNumbers.Core.Gpu.Accelerators;
-
-internal sealed class PrimeTesterByLastDigitAccelerator
-{
-	#region Pool
-
-	[ThreadStatic]
-	private static Queue<PrimeTesterByLastDigitAccelerator>? Pool;
-
-	// internal static void EnsureStaticTables(Accelerator accelerator)
-	// {
-	// 	var tables = CreateSharedTables(accelerator);
-	// }
-
-	internal static void WarmUp()
-	{
-		Accelerator[]? accelerators = AcceleratorPool.Shared.Accelerators;
-		int acceleratorCount = accelerators.Length;
-		Span<AcceleratorStream> streams = new(new AcceleratorStream[acceleratorCount]);
-		for (var i = 0; i < acceleratorCount; i++)
-		{
-			Console.WriteLine($"Preparing accelerator {i}...");
-			var accelerator = accelerators[i];
-			AcceleratorStream stream = accelerator.CreateStream();
-			LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
-			// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
-			// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
-			// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
-			// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
-			streams[i] = stream;
-		}
-
-		for (var i = 0; i < acceleratorCount; i++)
-		{
-			AcceleratorStream stream = streams[i];
-			stream.Synchronize();
-			stream.Dispose();
-			Console.WriteLine($"Accelerator {i} is ready");
-		}
-	}
-
-	internal static PrimeTesterByLastDigitAccelerator Rent(int minBufferCapacity = 1)
-	{
-		var pool = Pool ??= new();
-		if (!pool.TryDequeue(out var lease))
-		{
-			lease = new PrimeTesterByLastDigitAccelerator(minBufferCapacity);
-		}
-		else
-		{
-			lease.EnsureCapacity(minBufferCapacity);
-		}
-
-		return lease;
-	}
-
-	internal static void Return(PrimeTesterByLastDigitAccelerator lease)
-	{
-		var pool = Pool ??= new();
-		pool.Enqueue(lease);
-	}
-
-	internal static void DisposeAll()
-	{
-		if (Pool != null)
-		{
-			while (Pool.TryDequeue(out var lease))
-			{
-				lease.DisposeResources();
-			}
-		}
-	}
-
-	internal static void Clear()
-	{
-		if (Pool != null)
-		{
-			var retained = new List<PrimeTesterByLastDigitAccelerator>();
-			while (Pool.TryDequeue(out var lease))
-			{
-				lease.DisposeResources();
-			}
-		}
-	}
-
-	#endregion
-
-	private readonly LastDigitGpuTables _sharedTables;
-	internal readonly Context Context;
-	public readonly Accelerator Accelerator;
-	private AcceleratorStream? _stream;
-	public AcceleratorStream Stream
-	{
-		[MethodImpl(MethodImplOptions.AggressiveInlining)]
-		get => _stream ?? Accelerator.CreateStream();
-	}
-
-	[ThreadStatic]
-	public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>? SmallPrimeSieveKernel;
-
-	internal static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetSmallPrimeSieveKernel(Accelerator accelerator)
-	{
-		var pool = SmallPrimeSieveKernel ??= [];
-		if (pool.TryGetValue(accelerator, out var cached))
-		{
-			return cached;
-		}
-
-		cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
-		pool[accelerator] = cached;
-		return cached;
-	}
-
-	[ThreadStatic]
-	public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>>? 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;
-
-	public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne => _sharedTables.DevicePrimesLastOne;
-	public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven => _sharedTables.DevicePrimesLastSeven;
-	public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree => _sharedTables.DevicePrimesLastThree;
-	public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine => _sharedTables.DevicePrimesLastNine;
-	public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne => _sharedTables.DevicePrimesPow2LastOne;
-	public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven => _sharedTables.DevicePrimesPow2LastSeven;
-	public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree => _sharedTables.DevicePrimesPow2LastThree;
-	public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine => _sharedTables.DevicePrimesPow2LastNine;
-
-	internal PrimeTesterByLastDigitAccelerator(int minBufferCapacity)
-	{
-		Accelerator accelerator = AcceleratorPool.Shared.Rent();
-		Context = accelerator.Context;
-
-		Accelerator = accelerator;
-		AcceleratorStream stream = accelerator.CreateStream();
-
-		var sharedTables = LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
-		_sharedTables = sharedTables;
-
-		EnsureCapacity(minBufferCapacity);
-		stream.Synchronize();
-		stream.Dispose();
-	}
-
-	public void EnsureCapacity(int minCapacity)
-	{
-		if (minCapacity <= BufferCapacity)
-		{
-			return;
-		}
-
-		Input?.Dispose();
-		Output?.Dispose();
-		HeuristicFlag?.Dispose();
-
-		// lock (Accelerator)
-		{
-			Input = Accelerator.Allocate1D<ulong>(minCapacity);
-			Output = Accelerator.Allocate1D<byte>(minCapacity);
-			HeuristicFlag = Accelerator.Allocate1D<int>(minCapacity);
-		}
-
-		BufferCapacity = minCapacity;
-	}
-
-	public void Dispose()
-	{
-		_stream?.Dispose();
-		_stream = null;
-		Return(this);
-	}
-
-	internal void DisposeResources()
-	{
-		Input?.Dispose();
-		Output?.Dispose();
-		HeuristicFlag?.Dispose();
-		_stream?.Dispose();
-		_stream = null;
-		// ReleaseSharedTables(_sharedTables);
-		// These resources are shared between GPU leases
-		// Accelerator.Dispose();
-		// AcceleratorContext.Dispose();
-	}
-}

PerfectNumbers.Core/Gpu/HeuristicGpuDivisorTables.cs

@@ -4,7 +4,7 @@ using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 
-internal readonly struct HeuristicGpuDivisorTables
+public readonly struct HeuristicGpuDivisorTables
 {
 	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding1;
 	public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding3;

PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs

@@ -5,7 +5,7 @@ using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 
-internal enum HeuristicGpuDivisorTableKind : byte
+public enum HeuristicGpuDivisorTableKind : byte
 {
     GroupA = 0,
     GroupBEnding1 = 1,

PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs

@@ -10,9 +10,8 @@ public static class Pow2MontgomeryGpuCalculator
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static ulong CalculateKeep(ulong exponent, in MontgomeryDivisorData divisor)
 	{
-		GpuPrimeWorkLimiter.Acquire();
 
-		var gpu = Pow2MontgomeryAccelerator.Rent();
+		var gpu = Pow2MontgomeryAccelerator.Rent(1);
 		Accelerator? accelerator = gpu.Accelerator;
 		AcceleratorStream stream = gpu.Stream!;
 		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
@@ -23,23 +22,24 @@ public static class Pow2MontgomeryGpuCalculator
 		// resultBuffer.MemSetToZero(stream);
 
 		var kernel = gpu.KeepMontgomeryKernel;
+
+		GpuPrimeWorkLimiter.Acquire();
 		kernel(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
 
 		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
+		GpuPrimeWorkLimiter.Release();
 
 		Pow2MontgomeryAccelerator.Return(gpu);
-		GpuPrimeWorkLimiter.Release();
 		return result;
 	}
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static ulong CalculateConvert(ulong exponent, in MontgomeryDivisorData divisor)
 	{
-		GpuPrimeWorkLimiter.Acquire();
 
-		var gpu = Pow2MontgomeryAccelerator.Rent();
+		var gpu = Pow2MontgomeryAccelerator.Rent(1);
 		AcceleratorStream stream = gpu.Stream!;
 		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
 		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.Output;
@@ -49,14 +49,16 @@ public static class Pow2MontgomeryGpuCalculator
 		// resultBuffer.MemSetToZero(stream);
 
 		var kernel = gpu.ConvertToStandardKernel;
+
+		GpuPrimeWorkLimiter.Acquire();
 		kernel(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
 
 		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
+		GpuPrimeWorkLimiter.Release();
 
 		Pow2MontgomeryAccelerator.Return(gpu);
-		GpuPrimeWorkLimiter.Release();
 		return result;
 	}
 

PerfectNumbers.Core/PrimeOrderCalculator.Cpu.cs

@@ -52,14 +52,14 @@ internal static partial class PrimeOrderCalculator
 
 		ulong phi = prime - 1UL;
 
-		Pow2MontgomeryAccelerator gpu = Pow2MontgomeryAccelerator.Rent();
+		Pow2MontgomeryAccelerator gpu = Pow2MontgomeryAccelerator.Rent(1);
 		if (IsGpuHeuristicDevice && PrimeOrderGpuHeuristics.TryCalculateOrder(prime, previousOrder, config, divisorData, out ulong gpuOrder))
 		{
 			Pow2MontgomeryAccelerator.Return(gpu);
 			return gpuOrder;
 		}
 
-		PartialFactorResult phiFactors = PartialFactor(phi, config);
+		PartialFactorResult phiFactors = PartialFactor(gpu, phi, config);
 
 		ulong result;
 		if (phiFactors.Factors is null)
@@ -106,7 +106,7 @@ internal static partial class PrimeOrderCalculator
 			return CalculateByFactorizationGpu(gpu, prime, divisorData);
 		}
 
-		if (TryHeuristicFinishCpu(prime, candidateOrder, previousOrder, divisorData, config, phiFactors, orderFactors, out ulong order))
+		if (TryHeuristicFinishCpu(gpu, prime, candidateOrder, previousOrder, divisorData, config, phiFactors, orderFactors, out ulong order))
 		{
 			GpuPrimeWorkLimiter.Release();
 			return order;
@@ -347,7 +347,7 @@ internal static partial class PrimeOrderCalculator
 			return false;
 		}
 
-		PartialFactorResult factorization = PartialFactor(order, config);
+		PartialFactorResult factorization = PartialFactor(gpu, order, config);
 		if (factorization.Factors is null)
 		{
 			factorization.Dispose();
@@ -488,6 +488,7 @@ internal static partial class PrimeOrderCalculator
 	}
 
 	private static bool TryHeuristicFinishCpu(
+		Pow2MontgomeryAccelerator gpu,
 		ulong prime,
 		ulong order,
 		ulong? previousOrder,
@@ -505,7 +506,7 @@ internal static partial class PrimeOrderCalculator
 		}
 
 		// Reuse the partial factorization from TryConfirmOrderCpu when available.
-		PartialFactorResult orderFactors = cachedOrderFactors ?? PartialFactor(order, config);
+		PartialFactorResult orderFactors = cachedOrderFactors ?? PartialFactor(gpu, order, config);
 		try
 		{
 			if (orderFactors.Factors is null)
@@ -621,7 +622,7 @@ internal static partial class PrimeOrderCalculator
 
 					bool equalsOne = powStepper.InitializeCpuIsUnity(candidate);
 					powStepperInitialized = true;
-					if (equalsOne && TryConfirmCandidateCpu(prime, candidate, divisorData, config, ref powUsed, powBudget))
+					if (equalsOne && TryConfirmCandidateCpu(gpu, prime, candidate, divisorData, config, ref powUsed, powBudget))
 					{
 						if (gpuPool is not null)
 						{
@@ -658,7 +659,7 @@ internal static partial class PrimeOrderCalculator
 						continue;
 					}
 
-					if (!TryConfirmCandidateCpu(prime, candidate, divisorData, config, ref powUsed, powBudget))
+					if (!TryConfirmCandidateCpu(gpu, prime, candidate, divisorData, config, ref powUsed, powBudget))
 					{
 						continue;
 					}
@@ -863,9 +864,9 @@ internal static partial class PrimeOrderCalculator
 		}
 	}
 
-	private static bool TryConfirmCandidateCpu(ulong prime, ulong candidate, in MontgomeryDivisorData divisorData, in PrimeOrderSearchConfig config, ref int powUsed, int powBudget)
+	private static bool TryConfirmCandidateCpu(Pow2MontgomeryAccelerator gpu, ulong prime, ulong candidate, in MontgomeryDivisorData divisorData, in PrimeOrderSearchConfig config, ref int powUsed, int powBudget)
 	{
-		PartialFactorResult factorization = PartialFactor(candidate, config);
+		PartialFactorResult factorization = PartialFactor(gpu, candidate, config);
 		try
 		{
 			if (factorization.Factors is null)
@@ -1056,7 +1057,7 @@ internal static partial class PrimeOrderCalculator
 	// private static ulong _partialFactorPendingHits;
 	// private static ulong _partialFactorCofactorHits;
 
-	private static PartialFactorResult PartialFactor(ulong value, in PrimeOrderSearchConfig config)
+	private static PartialFactorResult PartialFactor(Pow2MontgomeryAccelerator gpu, ulong value, in PrimeOrderSearchConfig config)
 	{
 		if (value <= 1UL)
 		{
@@ -1125,6 +1126,8 @@ internal static partial class PrimeOrderCalculator
 			bool pollardRhoDeadlineReached = Stopwatch.GetTimestamp() > deadlineTimestamp;
 			if (!pollardRhoDeadlineReached)
 			{
+				GpuPrimeWorkLimiter.Acquire();
+
 				while (compositeStack.Count > 0)
 				{
 					ulong composite = compositeStack.Pop();
@@ -1141,7 +1144,7 @@ internal static partial class PrimeOrderCalculator
 					// Console.WriteLine($"Partial factor hits {Volatile.Read(ref _partialFactorHits)}");
 
 					// bool isPrime = PrimeTester.IsPrimeCpu(composite, CancellationToken.None);
-					bool isPrime = PrimeTester.IsPrime(composite);
+					bool isPrime = PrimeTester.IsPrimeGpu(gpu, composite);
 					// bool isPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(composite);
 
 					if (isPrime)
@@ -1173,6 +1176,8 @@ internal static partial class PrimeOrderCalculator
 				{
 					pollardRhoDeadlineReached = Stopwatch.GetTimestamp() > deadlineTimestamp;
 				}
+
+				GpuPrimeWorkLimiter.Release();
 			}
 
 			if (pollardRhoDeadlineReached)
@@ -1188,6 +1193,7 @@ internal static partial class PrimeOrderCalculator
 		bool cofactorContainsComposite = false;
 		int pendingCount = pending.Count;
 		int index = 0;
+		GpuPrimeWorkLimiter.Acquire();
 		for (; index < pendingCount; index++)
 		{
 			PendingEntry entry = pending[index];
@@ -1209,7 +1215,7 @@ 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.IsPrimeGpu(composite);
+				bool isPrime = PrimeTester.IsPrimeGpu(gpu, composite);
 				// bool isPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(composite);
 
 				entry = entry.WithPrimality(isPrime);
@@ -1228,7 +1234,7 @@ internal static partial class PrimeOrderCalculator
 				cofactorContainsComposite = true;
 			}
 		}
-
+		
 		bool cofactorIsPrime;
 		if (cofactor <= 1UL)
 		{
@@ -1246,11 +1252,13 @@ internal static partial class PrimeOrderCalculator
 			// Atomic.Add(ref _partialFactorCofactorHits, 1UL);
 			// Console.WriteLine($"Partial factor cofactor hits {Volatile.Read(ref _partialFactorCofactorHits)}");
 
-			cofactorIsPrime = PrimeTester.IsPrimeGpu(cofactor);
+			cofactorIsPrime = PrimeTester.IsPrimeGpu(gpu, cofactor);
 			// cofactorIsPrime = PrimeTester.IsPrimeGpu(cofactor);
 			// cofactorIsPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(cofactor);
 		}
 
+		GpuPrimeWorkLimiter.Release();
+
 		ArrayPool<FactorEntry> pool = ThreadStaticPools.FactorEntryPool;
 		bool temp;
 		if (factorCount == 0)

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -38,7 +38,7 @@ internal static partial class PrimeOrderCalculator
 		entries.Sort(entries, static (a, b) => a.Key.CompareTo(b.Key));
 
 		AcceleratorStream stream = gpu.Stream!;
-		gpu.EnsureCapacity(entryCount);
+		gpu.EnsureCapacity(entryCount, 1);
 		gpu.Pow2ModEntriesToTestOnDevice.View.CopyFromCPU(stream, entries);
 
 		ulong order = phi;

PerfectNumbers.Core/PrimeTester.cs

@@ -94,17 +94,17 @@ public sealed class PrimeTester
 	public static bool IsPrimeGpu(ulong n)
 	{
 		GpuPrimeWorkLimiter.Acquire();
-		var gpu = PrimeTesterByLastDigitAccelerator.Rent(1);
+		var gpu = Pow2MontgomeryAccelerator.Rent(1);
 		var accelerator = gpu.Accelerator;
-		var stream = gpu.Stream;
+		var stream = gpu.Stream!;
-		var kernel = PrimeTesterByLastDigitAccelerator.GetSmallPrimeSieveKernel(accelerator);
+		var kernel = gpu.SmallPrimeSieveKernel;
 
 		ulong value = n;
 		// Span<byte> flag = stackalloc byte[1];
 		byte flag = 0;
 
-		var input = gpu.Input;
+		var input = gpu.PrimeTestInput;
-		var output = gpu.Output;
+		var output = gpu.PrimeTestOutput;
 		var inputView = input.View;
 		var outputView = output.View;
 
@@ -131,6 +131,42 @@ public sealed class PrimeTester
 		return flag != 0;
 	}
 
+	[MethodImpl(MethodImplOptions.AggressiveInlining)]
+	public static bool IsPrimeGpu(Pow2MontgomeryAccelerator gpu, ulong n)
+	{
+		var accelerator = gpu.Accelerator;
+		var stream = gpu.Stream!;
+		Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>? kernel = gpu.SmallPrimeSieveKernel;
+
+		ulong value = n;
+		// Span<byte> flag = stackalloc byte[1];
+		byte flag = 0;
+
+		var input = gpu.PrimeTestInput;
+		var output = gpu.PrimeTestOutput;
+		var inputView = input.View;
+		var outputView = output.View;
+
+		inputView.CopyFromCPU(stream, ref value, 1);
+		kernel(
+						stream,
+						1,
+						inputView,
+						gpu.DevicePrimesLastOne.View,
+						gpu.DevicePrimesLastSeven.View,
+						gpu.DevicePrimesLastThree.View,
+						gpu.DevicePrimesLastNine.View,
+						gpu.DevicePrimesPow2LastOne.View,
+						gpu.DevicePrimesPow2LastSeven.View,
+						gpu.DevicePrimesPow2LastThree.View,
+						gpu.DevicePrimesPow2LastNine.View,
+						outputView);
+		outputView.CopyToCPU(stream, ref flag, 1);
+		stream.Synchronize();
+
+		return flag != 0;
+	}
+
 	public static int GpuBatchSize { get; set; } = 262_144;
 
 	private static readonly object GpuWarmUpLock = new();
@@ -151,7 +187,7 @@ public sealed class PrimeTester
 				return;
 			}
 
-			PrimeTesterByLastDigitAccelerator.WarmUp();
+			Pow2MontgomeryAccelerator.WarmUp();
 			WarmedGpuLeaseCount = target;
 		}
 	}
@@ -160,15 +196,15 @@ public sealed class PrimeTester
 	public static void IsPrimeBatchGpu(ReadOnlySpan<ulong> values, Span<byte> results)
 	{
 		GpuPrimeWorkLimiter.Acquire();
-		var gpu = PrimeTesterByLastDigitAccelerator.Rent(GpuBatchSize);
+		var gpu = Pow2MontgomeryAccelerator.Rent(GpuBatchSize);
 		var accelerator = gpu.Accelerator;
-		var stream = gpu.Stream;
+		var stream = gpu.Stream!;
-		var kernel = PrimeTesterByLastDigitAccelerator.GetSmallPrimeSieveKernel(accelerator);
+		var kernel = gpu.SmallPrimeSieveKernel;
 		int totalLength = values.Length;
 		int batchSize = GpuBatchSize;
 
-		var input = gpu.Input;
+		var input = gpu.PrimeTestInput;
-		var output = gpu.Output;
+		var output = gpu.PrimeTestOutput;
 		var inputView = input.View;
 		var outputView = output.View;
 
@@ -211,12 +247,12 @@ public sealed class PrimeTester
 	// Expose cache clearing for accelerator disposal coordination
 	public static void ClearGpuCaches()
 	{
-		PrimeTesterByLastDigitAccelerator.Clear();
+		Pow2MontgomeryAccelerator.Clear();
 	}
 
 	internal static void DisposeGpuContexts()
 	{
-		PrimeTesterByLastDigitAccelerator.DisposeAll();
+		Pow2MontgomeryAccelerator.DisposeAll();
 		lock (GpuWarmUpLock)
 		{
 			WarmedGpuLeaseCount = 0;
@@ -241,9 +277,9 @@ 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 = PrimeTesterByLastDigitAccelerator.Rent();
+		var gpu = Pow2MontgomeryAccelerator.Rent(1);
 		var accelerator = gpu.Accelerator;
-		var stream = gpu.Stream;
+		var stream = gpu.Stream!;
 
 		int length = values.Length;
 		ArrayPool<ulong> pool = ThreadStaticPools.UlongPool;