Performance improvements

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -7,6 +7,7 @@ using ILGPU;
 using ILGPU.Algorithms;
 using ILGPU.Runtime;
 using PerfectNumbers.Core.Gpu;
+using PerfectNumbers.Core.Gpu.Accelerators;
 
 namespace PerfectNumbers.Core;
 
@@ -273,11 +274,7 @@ public sealed class DivisorCycleCache
 	private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
 	{
 		int length = divisors.Length;
-		GpuPrimeWorkLimiter.Acquire();
 
-		Accelerator accelerator = _accelerator;
-		var stream = accelerator.CreateStream();
-		var kernel = _gpuKernel;
 		EnsureCapacity(length);
 
 		ulong[]? rentedPow = null;
@@ -300,9 +297,6 @@ public sealed class DivisorCycleCache
 			? stackalloc byte[length]
 			: new Span<byte>(rentedStatus = ThreadStaticPools.BytePool.Rent(length), 0, length);
 
-		ref ulong divisorRef = ref MemoryMarshal.GetReference(divisors);
-		divisorBuffer!.View.CopyFromCPU(stream, ref divisorRef, length);
-
 		for (int i = 0; i < length; i++)
 		{
 			ulong divisor = divisors[i];
@@ -318,17 +312,22 @@ 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);
+		GpuPrimeWorkLimiter.Acquire();
+		var stream = AcceleratorStreamPool.Rent(_accelerator);// accelerator.CreateStream();
+
+		divisorBuffer!.View.CopyFromCPU(stream, divisors);
+		powBuffer!.View.CopyFromCPU(stream, powSpan);
+		orderBuffer!.View.CopyFromCPU(stream, orderSpan);
+		resultBuffer!.View.CopyFromCPU(stream, resultSpan);
+		statusBuffer!.View.CopyFromCPU(stream, statusSpan);
 
 		int pending;
+		var kernel = _gpuKernel;
 		do
 		{
 			kernel(stream, length, _divisorCyclesBatchSize, divisorBuffer.View, powBuffer.View, orderBuffer.View, resultBuffer.View, statusBuffer.View);
 
-			statusBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(statusSpan), length);
+			statusBuffer.View.CopyToCPU(stream, statusSpan);
 			stream.Synchronize();
 
 			pending = 0;
@@ -342,9 +341,9 @@ public sealed class DivisorCycleCache
 		}
 		while (pending > 0);
 
-		resultBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
+		resultBuffer.View.CopyToCPU(stream,resultSpan);
 		stream.Synchronize();
-		stream.Dispose();
+		AcceleratorStreamPool.Return(stream);
 
 		resultSpan.CopyTo(destination);
 		if (rentedPow is not null)

PerfectNumbers.Core/Gpu/Accelerators/AcceleratorStreamPool.cs

@@ -1,24 +1,26 @@
 using System.Collections.Concurrent;
 using ILGPU.Runtime;
+using Open.Collections;
 
 namespace PerfectNumbers.Core.Gpu.Accelerators;
 
 public static class AcceleratorStreamPool
 {
-	private static readonly ConcurrentDictionary<Accelerator, ConcurrentQueue<AcceleratorStream>> _streams = new(20_480, PerfectNumberConstants.RollingAccelerators);
+	private static readonly Dictionary<Accelerator, ConcurrentQueue<AcceleratorStream>> _streams = new(PerfectNumberConstants .RollingAccelerators);
 	// private static readonly ConcurrentDictionary<Accelerator, AcceleratorStream> _streams = new(AcceleratorPool.Shared.Accelerators.ToDictionary(static x => x, CreateStream));
 
 	private static SemaphoreSlim CreateLock() => new(PerfectNumberConstants.ThreadsByStream);
 	private static AcceleratorStream CreateStream(Accelerator accelerator) => accelerator.CreateStream();
 
-	private static readonly ConcurrentDictionary<Accelerator, SemaphoreSlim> _locks = new(AcceleratorPool.Shared.Accelerators.ToDictionary(x => x, static x => CreateLock()));
+	private static readonly Dictionary<Accelerator, SemaphoreSlim> _locks = new(AcceleratorPool.Shared.Accelerators.ToDictionary(x => x, static x => CreateLock()));
 
 	public static AcceleratorStream Rent(Accelerator accelerator)
 	{
-		var streamLock = _locks.GetOrAdd(accelerator, static _ => new SemaphoreSlim(PerfectNumberConstants.ThreadsByStream));
+		var streamLock = _locks[accelerator];
+		// var streamLock = _locks.GetOrAdd(accelerator, static _ => CreateLock());
 
 		streamLock.Wait();
-		var queue = _streams.GetOrAdd(accelerator, static _ => new());
+		var queue = _streams[accelerator];
 
 		return queue.TryDequeue(out var stream)
 			? stream
@@ -28,9 +30,23 @@ public static class AcceleratorStreamPool
 	public static void Return(AcceleratorStream stream)
 	{
 		Accelerator accelerator = stream.Accelerator;
-		var streamLock = _locks.GetOrAdd(accelerator, static _ => CreateLock());
-		var queue = _streams.GetOrAdd(accelerator, static _ => new());
+		var streamLock = _locks[accelerator];
+		var queue = _streams[accelerator];
 		queue.Enqueue(stream);
 		streamLock.Release();
 	}
+
+	public static void WarmUp(Accelerator accelerator)
+	{
+		_streams.Add(accelerator, new());
+		_ = _locks.TryAdd(accelerator, CreateLock());
+
+		// ConcurrentQueue<AcceleratorStream> streamQueue = new();
+		// var stream = accelerator.CreateStream();
+		// // var stream = accelerator.CreateStream();
+		// streamQueue.Enqueue(stream);
+		// _streams[accelerator] = streamQueue;
+		// _locks[accelerator] = CreateLock();
+		// stream.Synchronize();
+	}
 }

PerfectNumbers.Core/Gpu/Accelerators/Pow2MontgomeryAccelerator.cs

@@ -72,6 +72,8 @@ public sealed class Pow2MontgomeryAccelerator
 		{
 			Console.WriteLine($"Preparing accelerator {i}...");
 			var accelerator = accelerators[i];
+			AcceleratorStreamPool.WarmUp(accelerator);
+			// AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
 			AcceleratorStream stream = accelerator.CreateStream();
 			LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
 			// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
@@ -80,6 +82,8 @@ public sealed class Pow2MontgomeryAccelerator
 			// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
 			stream.Synchronize();
 			stream.Dispose();
+
+			// AcceleratorStreamPool.Return(stream);
 		}
 	}
 

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -142,14 +142,9 @@ public class GpuKernelPool
 
 		static void InitSmallPrimeFactorKernelScan(Accelerator accelerator, KernelContainer kernels)
 		{
-			InitOnce(ref kernels.SmallPrimeFactor, () =>
-			{
-				var loaded = accelerator.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 loaded = accelerator.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);
-
-				return kernel.CreateLauncherDelegate<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>>>();
-			});
+			kernels.SmallPrimeFactor = KernelUtil.GetKernel(loaded);
 		}
 	}
 

PerfectNumbers.Core/Gpu/GpuScratchBufferPool.cs

@@ -1,23 +1,20 @@
-using System.Collections.Concurrent;
 using ILGPU;
 using ILGPU.Runtime;
 
 namespace PerfectNumbers.Core.Gpu;
 
-public sealed class ScratchBuffer
+public sealed class GpuScratchBuffer
 {
-	public MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimeFactorPrimeSlots { get; private set; }
-	public MemoryBuffer1D<int, Stride1D.Dense> SmallPrimeFactorExponentSlots { get; private set; }
-	public MemoryBuffer1D<int, Stride1D.Dense> SmallPrimeFactorCountSlot { get; private set; }
-	public MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimeFactorRemainingSlot { get; private set; }
-
-	public MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxFactors { get; private set; }
-	public MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxCandidates { get; private set; }
-	public MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxResult { get; private set; }
-
 	internal readonly Accelerator _accelerator;
+	public readonly MemoryBuffer1D<int, Stride1D.Dense> SmallPrimeFactorCountSlot;
+	public MemoryBuffer1D<int, Stride1D.Dense> SmallPrimeFactorExponentSlots;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimeFactorPrimeSlots;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> SmallPrimeFactorRemainingSlot;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxCandidates;
+	public MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxFactors;
+	public readonly MemoryBuffer1D<ulong, Stride1D.Dense> SpecialMaxResult;
 
-	public ScratchBuffer(Accelerator accelerator, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
+	public GpuScratchBuffer(Accelerator accelerator, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
 	{
 		_accelerator = accelerator;
 
@@ -63,18 +60,29 @@ public sealed class ScratchBuffer
 			}
 		}
 	}
+
+	internal void Dispose()
+	{
+		SmallPrimeFactorPrimeSlots.Dispose();
+		SmallPrimeFactorExponentSlots.Dispose();
+		SmallPrimeFactorCountSlot.Dispose();
+		SmallPrimeFactorRemainingSlot.Dispose();
+		SpecialMaxFactors.Dispose();
+		SpecialMaxCandidates.Dispose();
+		SpecialMaxResult.Dispose();
+	}
 }
 
 public static class GpuScratchBufferPool
 {
 	[ThreadStatic]
-	private static Dictionary<Accelerator, Queue<ScratchBuffer>>? _pools;
+	private static Dictionary<Accelerator, Queue<GpuScratchBuffer>>? _pools;
 
 	// private static readonly ConcurrentDictionary<Accelerator, ConcurrentQueue<ScratchBuffer>> _pools = new();
 	private const int DefaultSmallPrimeFactorSlotCount = 64; // From PrimeOrderCalculator.Gpu.cs
 	private const int DefaultSpecialMaxFactorCapacity = 1024; // A reasonable default, will be resized if needed
 
-	public static ScratchBuffer Rent(Accelerator accelerator, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
+	public static GpuScratchBuffer Rent(Accelerator accelerator, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
 	{
 		const int WarmUpCapacity = 100;
 		// var pool = _pools.GetOrAdd(accelerator, _ => new ConcurrentQueue<ScratchBuffer>());
@@ -107,10 +115,10 @@ public static class GpuScratchBufferPool
 			return buffer;
 		}
 
-		return new ScratchBuffer(accelerator, smallPrimeFactorSlotCount, specialMaxFactorCapacity);
+		return new GpuScratchBuffer(accelerator, smallPrimeFactorSlotCount, specialMaxFactorCapacity);
 	}
 
-	public static void Return(ScratchBuffer buffer)
+	public static void Return(GpuScratchBuffer buffer)
 	{
 		if (_pools!.TryGetValue(buffer._accelerator, out var pool))
 		{
@@ -128,4 +136,22 @@ public static class GpuScratchBufferPool
 		// 	Return(buffer);
 		// }
 	}
+
+	public static void DisposeAll()
+	{
+		var pool = _pools;
+		if (pool is not null)
+		{
+			foreach (var queue in pool.Values)
+			{
+				while (queue.TryDequeue(out var buffer))
+				{
+					buffer.Dispose();
+				}
+			}
+
+			pool.Clear();
+			_pools = null;
+		}
+	}
 }

PerfectNumbers.Core/Gpu/KernelContainer.cs

@@ -22,7 +22,7 @@ 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 Kernel? SmallPrimeFactor;
 	
 	public Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>? SpecialMax;
 

PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs

@@ -27,6 +27,7 @@ public static class Pow2MontgomeryGpuCalculator
 		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
 
 		kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
+		// kernel.Launch(stream, 1, 1, exponentBuffer.View, divisor, resultBuffer.View);
 
 		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
@@ -83,6 +84,7 @@ public static class Pow2MontgomeryGpuCalculator
 		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
 
 		kernelLauncher(stream, 1,exponentBuffer.View, divisor, resultBuffer.View);
+		// kernel.Launch(stream, 1, 1,exponentBuffer.View, divisor, resultBuffer.View);
 
 		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);

PerfectNumbers.Core/MersenneNumberDivisorByDivisorTester.cs

@@ -4,6 +4,7 @@ using System.Threading.Tasks;
 using System.Globalization;
 using System.Runtime.InteropServices;
 using PerfectNumbers.Core.Gpu.Accelerators;
+using PerfectNumbers.Core.Gpu;
 
 namespace PerfectNumbers.Core;
 
@@ -304,6 +305,7 @@ public static class MersenneNumberDivisorByDivisorTester
 
 						Console.WriteLine($"Task finished for range {rangeStart}");
 						Pow2MontgomeryAccelerator.DisposeAll();
+						GpuScratchBufferPool.DisposeAll();
 					},
 					CancellationToken.None,
 					TaskCreationOptions.DenyChildAttach,

PerfectNumbers.Core/PrimeOrderCalculator.Cpu.cs

@@ -86,7 +86,7 @@ internal static partial class PrimeOrderCalculator
 		ulong phi,
 		PartialFactorResult phiFactors)
 	{
-		if (phiFactors.FullyFactored && TrySpecialMaxCpu(phi, prime, phiFactors, divisorData))
+		if (phiFactors.FullyFactored && TrySpecialMaxCpu(gpu, phi, prime, phiFactors, divisorData))
 		{
 			return phi;
 		}
@@ -119,7 +119,7 @@ internal static partial class PrimeOrderCalculator
 	}
 
 	// private static ulong specialMaxHits;
-	private static bool TrySpecialMaxCpu(ulong phi, ulong prime, PartialFactorResult factors, in MontgomeryDivisorData divisorData)
+	private static bool TrySpecialMaxCpu(Pow2MontgomeryAccelerator gpu, ulong phi, ulong prime, PartialFactorResult factors, in MontgomeryDivisorData divisorData)
 	{
 		int length = factors.Count;
 		// The phi factorization on the scanner path always yields at least one entry for phi >= 2,
@@ -140,7 +140,7 @@ internal static partial class PrimeOrderCalculator
 
 		// Atomic.Add(ref specialMaxHits, 1);
 		// Console.WriteLine($"Special max GPU hit {specialMaxHits} ({length})");
-		return EvaluateSpecialMaxCandidatesGpu(stackBuffer, factorSpan, phi, prime, divisorData);
+		return EvaluateSpecialMaxCandidatesGpu(gpu, stackBuffer, factorSpan, phi, prime, divisorData);
 	}
 
 	private static bool EvaluateSpecialMaxCandidates(Span<ulong> buffer, ReadOnlySpan<FactorEntry> factors, ulong phi, ulong prime, in MontgomeryDivisorData divisorData)

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -39,18 +39,18 @@ internal static partial class PrimeOrderCalculator
 		
 		entries.Sort(entries, static (a, b) => a.Key.CompareTo(b.Key));
 
-		// AcceleratorStream stream = gpu.Stream!;
+		ulong order = phi;
+		var kernel = gpu.CheckFactorsKernel;
 		gpu.EnsureCapacity(entryCount, 1);
 
 		GpuPrimeWorkLimiter.Acquire();
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		gpu.Pow2ModEntriesToTestOnDevice.View.CopyFromCPU(stream, entries);
 
-		ulong order = phi;
-		var kernel = gpu.CheckFactorsKernel;
-		stream.Synchronize();
+		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
+		
+		kernelLauncher(stream, entryCount, phi, gpu.Pow2ModEntriesToTestOnDevice.View, divisorData, gpu.OutputUlong.View);
 
-		kernel.Launch(stream, 1, entryCount, phi, gpu.Pow2ModEntriesToTestOnDevice.View, divisorData, gpu.OutputUlong.View);
 		gpu.OutputUlong.View.CopyToCPU(stream, ref order, 1);
 		stream.Synchronize();
 		AcceleratorStreamPool.Return(stream);
@@ -67,19 +67,16 @@ internal static partial class PrimeOrderCalculator
 		Span<int> exponentBuffer = exponentBufferArray.AsSpan(0, GpuSmallPrimeFactorSlots);
 		remaining = value;
 
-		GpuPrimeWorkLimiter.Acquire();
 		Accelerator accelerator = AcceleratorPool.Shared.Rent();
-		var stream = accelerator.CreateStream();
+		GpuScratchBuffer scratch = GpuScratchBufferPool.Rent(accelerator, GpuSmallPrimeFactorSlots, 0);
+
+		GpuPrimeWorkLimiter.Acquire();
+		var stream = AcceleratorStreamPool.Rent(accelerator);
 		KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.SmallPrimeFactorKernelScan);
-
-		ScratchBuffer scratch = GpuScratchBufferPool.Rent(accelerator, GpuSmallPrimeFactorSlots, 0);
-
 		SmallPrimeFactorViews tables = GpuKernelPool.GetSmallPrimeFactorTables(kernels);
 
-		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");
-
 		var kernel = kernels.SmallPrimeFactor!;
+		var kernelLauncher = kernel.CreateLauncherDelegate<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>>>();
 
 		ArrayView1D<int, Stride1D.Dense> smallPrimeFactorCountSlotView = scratch.SmallPrimeFactorCountSlot.View;
 		ArrayView1D<ulong, Stride1D.Dense> smallPrimeFactorRemainingSlotView = scratch.SmallPrimeFactorRemainingSlot.View;
@@ -87,7 +84,7 @@ internal static partial class PrimeOrderCalculator
 		ArrayView1D<int, Stride1D.Dense> smallPrimeFactorExponentSlotsView = scratch.SmallPrimeFactorExponentSlots.View;
 		ArrayView1D<uint, Stride1D.Dense> smallPrimeFactorsPrimesView = kernels.SmallPrimeFactorsPrimes!.View;
 
-		kernel(
+		kernelLauncher(
 				stream,
 				1,
 				value,
@@ -101,26 +98,27 @@ internal static partial class PrimeOrderCalculator
 				smallPrimeFactorRemainingSlotView);
 
 
-		// Span<int> factorCountTemp = stackalloc int[1];
 		factorCount = 0;
 		smallPrimeFactorCountSlotView.CopyToCPU(stream, ref factorCount, 1);
 		stream.Synchronize();
-		// factorCount = factorCountTemp[0];
+
 		factorCount = Math.Min(factorCount, GpuSmallPrimeFactorSlots);
 
 		if (factorCount > 0)
 		{
-			smallPrimeFactorPrimeSlotsView.CopyToCPU(stream, ref MemoryMarshal.GetReference(primeBuffer), factorCount);
+			primeBuffer = primeBuffer[..factorCount];
+			smallPrimeFactorPrimeSlotsView.CopyToCPU(stream, primeBuffer);
 
-			smallPrimeFactorExponentSlotsView.CopyToCPU(stream, ref MemoryMarshal.GetReference(exponentBuffer), factorCount);
+			exponentBuffer = exponentBuffer[..factorCount];
+			smallPrimeFactorExponentSlotsView.CopyToCPU(stream, exponentBuffer);
 		}
 
-		// Span<ulong> remainingTemp = stackalloc ulong[1];
-
 		smallPrimeFactorRemainingSlotView.CopyToCPU(stream, ref remaining, 1);
-		// remaining = remainingTemp[0];
 		stream.Synchronize();
-		stream.Dispose();
+
+		AcceleratorStreamPool.Return(stream);
+		GpuScratchBufferPool.Return(scratch);
+		GpuPrimeWorkLimiter.Release();
 
 		for (int i = 0; i < factorCount; i++)
 		{
@@ -129,12 +127,11 @@ internal static partial class PrimeOrderCalculator
 			counts.Add(primeValue, exponent);
 		}
 
-		GpuScratchBufferPool.Return(scratch);
-		GpuPrimeWorkLimiter.Release();
 		return true;
 	}
 
 	private static bool EvaluateSpecialMaxCandidatesGpu(
+			Pow2MontgomeryAccelerator gpu,
 			Span<ulong> buffer,
 			ReadOnlySpan<FactorEntry> factors,
 			ulong phi,
@@ -154,11 +151,11 @@ internal static partial class PrimeOrderCalculator
 		}
 
 		GpuPrimeWorkLimiter.Acquire();
-		Accelerator accelerator = AcceleratorPool.Shared.Rent();
-		var stream = accelerator.CreateStream();
+		Accelerator accelerator = gpu.Accelerator; //AcceleratorPool.Shared.Rent();
+		var stream = AcceleratorStreamPool.Rent(accelerator);// accelerator.CreateStream();
 		var kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.EvaluateSpecialMaxCandidatesKernel);
 
-		ScratchBuffer scratch = GpuScratchBufferPool.Rent(accelerator, 0, factorCount);
+		GpuScratchBuffer scratch = GpuScratchBufferPool.Rent(accelerator, 0, factorCount);
 
 		ArrayView1D<ulong, Stride1D.Dense> specialMaxFactorsView = scratch.SpecialMaxFactors.View;
 		specialMaxFactorsView.SubView(0, factorCount).CopyFromCPU(stream, ref MemoryMarshal.GetReference(factorSpan), factorCount);
@@ -182,7 +179,9 @@ internal static partial class PrimeOrderCalculator
 		// specialMaxResultView.CopyToCPU(stream, ref result, 1);
 		specialMaxResultView.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
-		stream.Dispose();
+		
+		AcceleratorStreamPool.Return(stream);
+		// stream.Dispose();
 
 		// stream.Synchronize();
 		// return result != 0;

PerfectNumbers.Core/PrimeTester.cs

@@ -94,20 +94,22 @@ public sealed class PrimeTester
 	public static bool IsPrimeGpu(ulong n)
 	{
 		var gpu = Pow2MontgomeryAccelerator.Rent(1);
-		var kernel = gpu.SmallPrimeSieveKernel!;
 
 		// Span<byte> flag = stackalloc byte[1];
 
+
 		var inputView = gpu.PrimeTestInput.View;
+		var outputView = gpu.OutputByte.View;
+		var kernel = gpu.SmallPrimeSieveKernel!;
 
 		GpuPrimeWorkLimiter.Acquire();
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		// var stream = gpu.Stream!;
 		inputView.CopyFromCPU(stream, ref n, 1);
 
-		var outputView = gpu.OutputByte.View;
+		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
 
-		kernel.Launch(
+		kernelLauncher(
 						stream,
 						1,
 						inputView,
@@ -136,18 +138,15 @@ public sealed class PrimeTester
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static bool IsPrimeGpu(Pow2MontgomeryAccelerator gpu, ulong n)
 	{
-		// var stream = gpu.Stream!;
 		var input = gpu.PrimeTestInput;
 		var inputView = input.View;
+		var output = gpu.OutputByte;
+		var outputView = output.View;
+		Kernel kernel = gpu.SmallPrimeSieveKernel!;
 
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		inputView.CopyFromCPU(stream, ref n, 1);
 
-		// Span<byte> flag = stackalloc byte[1];
-		var output = gpu.OutputByte;
-		var outputView = output.View;
-
-		Kernel kernel = gpu.SmallPrimeSieveKernel!;
 		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
 		
 		kernelLauncher(
@@ -238,8 +237,7 @@ public sealed class PrimeTester
 					outputView);
 
 			var resultSlice = results.Slice(pos, count);
-			ref byte resultRef = ref MemoryMarshal.GetReference(resultSlice);
-			outputView.CopyToCPU(stream, ref resultRef, count);
+			outputView.CopyToCPU(stream, resultSlice);
 
 			pos += count;
 		}