Revert "Stream handling simplified to avoid dictionary lookups in favor of array lookups with known indexes."

This reverts commit a84445ef2a.

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -345,7 +345,7 @@ public sealed class DivisorCycleCache
 
 		resultBuffer.View.CopyToCPU(stream, resultSpan);
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 
 		resultSpan.CopyTo(destination);
 		if (rentedPow is not null)

PerfectNumbers.Core/Gpu/Accelerators/AcceleratorStreamPool.cs

@@ -7,7 +7,8 @@ public static class AcceleratorStreamPool
 {
 	private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
 	private static readonly SemaphoreSlim[] _locks = new SemaphoreSlim[PerfectNumberConstants.RollingAccelerators];
-	private static readonly AcceleratorStream[] _streams = new AcceleratorStream[PerfectNumberConstants.RollingAccelerators];
+	private static readonly ConcurrentQueue<AcceleratorStream>[] _streams = new ConcurrentQueue<AcceleratorStream> [PerfectNumberConstants.RollingAccelerators];
+
 	private static SemaphoreSlim CreateLock() => new(PerfectNumberConstants.ThreadsByAccelerator);
 
 
@@ -16,27 +17,26 @@ public static class AcceleratorStreamPool
 	public static AcceleratorStream Rent(int acceleratorIndex)
 	{
 		var streamLock = _locks[acceleratorIndex];
+		var queue = _streams[acceleratorIndex];
 
 		streamLock.Wait();
-		var queue = _streams[acceleratorIndex];
-		if (queue != null)
-		{
-			return queue;
-		}
 
-		queue = _accelerators[acceleratorIndex].CreateStream();
-		_streams[acceleratorIndex] = queue;
-		return queue;
+		return queue.TryDequeue(out var stream)
+			? stream
+			: _accelerators[acceleratorIndex].CreateStream();
 	}
 
-	public static void Return(int acceleratorIndex)
+	public static void Return(int acceleratorIndex, AcceleratorStream stream)
 	{
 		var streamLock = _locks[acceleratorIndex];
+		var queue = _streams[acceleratorIndex];
+		queue.Enqueue(stream);
 		streamLock.Release();
 	}
 
 	public static void WarmUp(int acceleratorIndex)
 	{
+		_streams[acceleratorIndex] = new();
 		_locks[acceleratorIndex] = CreateLock();
 	}
 }

PerfectNumbers.Core/Gpu/Accelerators/HeuristicGpuTables.cs

@@ -45,25 +45,25 @@ internal sealed class HeuristicGpuTables
 		HeuristicGroupBDivisorsEnding9 = CopyUintSpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesLastNineWithoutLastThree);
 		HeuristicGroupBDivisorSquaresEnding9 = CopySpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesPow2LastNineWithoutLastThree);
 
-		// var combinedEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1;
-		// HeuristicCombinedDivisorsEnding1 = CopySpanToDevice(accelerator, stream, combinedEnding1);
-		// var combinedSquaresEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1Squares;
-		// HeuristicCombinedDivisorSquaresEnding1 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding1);
+		var combinedEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1;
+		HeuristicCombinedDivisorsEnding1 = CopySpanToDevice(accelerator, stream, combinedEnding1);
+		var combinedSquaresEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1Squares;
+		HeuristicCombinedDivisorSquaresEnding1 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding1);
 
-		// var combinedEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3;
-		// HeuristicCombinedDivisorsEnding3 = CopySpanToDevice(accelerator, stream, combinedEnding3);
-		// var combinedSquaresEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3Squares;
-		// HeuristicCombinedDivisorSquaresEnding3 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding3);
+		var combinedEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3;
+		HeuristicCombinedDivisorsEnding3 = CopySpanToDevice(accelerator, stream, combinedEnding3);
+		var combinedSquaresEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3Squares;
+		HeuristicCombinedDivisorSquaresEnding3 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding3);
 
-		// var combinedEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7;
-		// HeuristicCombinedDivisorsEnding7 = CopySpanToDevice(accelerator, stream, combinedEnding7);
-		// var combinedSquaresEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7Squares;
-		// HeuristicCombinedDivisorSquaresEnding7 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding7);
+		var combinedEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7;
+		HeuristicCombinedDivisorsEnding7 = CopySpanToDevice(accelerator, stream, combinedEnding7);
+		var combinedSquaresEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7Squares;
+		HeuristicCombinedDivisorSquaresEnding7 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding7);
 
-		// var combinedEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9;
-		// HeuristicCombinedDivisorsEnding9 = CopySpanToDevice(accelerator, stream, combinedEnding9);
-		// var combinedSquaresEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9Squares;
-		// HeuristicCombinedDivisorSquaresEnding9 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding9);
+		var combinedEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9;
+		HeuristicCombinedDivisorsEnding9 = CopySpanToDevice(accelerator, stream, combinedEnding9);
+		var combinedSquaresEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9Squares;
+		HeuristicCombinedDivisorSquaresEnding9 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding9);
 	}
 
 	internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisors;

PerfectNumbers.Core/Gpu/Accelerators/HeuristicPrimeTesterAccelerator.cs

@@ -136,7 +136,7 @@ internal sealed class HeuristicPrimeTesterAccelerator
 		Accelerator = accelerator;
 		EnsureCapacity(minBufferCapacity);
 
-		HeuristicTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeTesterKernels.HeuristicTrialDivisionKernel));
+		HeuristicTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel));
 
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
 		
@@ -146,7 +146,7 @@ internal sealed class HeuristicPrimeTesterAccelerator
 		_heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
 
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(acceleratorIndex); 
+		AcceleratorStreamPool.Return(acceleratorIndex, stream); 
 	}
 
 	public void EnsureCapacity(int minCapacity)

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -25,7 +25,7 @@ public class GpuKernelPool
 
 	internal static KernelContainer GetOrAddKernels(int acceleratorIndex, AcceleratorStream stream, KernelType kernelType)
 	{
-		var pool = _pool ??= new KernelContainer[PerfectNumberConstants.RollingAccelerators];
+		var pool = _pool ??= [];
 		if (pool[acceleratorIndex] is not {} kernels)
 		{
 			kernels = new();

PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs

@@ -83,10 +83,20 @@ internal static class PrimeTesterKernels
         ArrayView<int> resultFlag,
         ulong n,
         ulong maxDivisorSquare,
-		ArrayView1D<ulong, Stride1D.Dense> divisors,
-		ArrayView1D<ulong, Stride1D.Dense> divisorSquares)
+        HeuristicGpuDivisorTableKind tableKind,
+        HeuristicGpuDivisorTables tables)
     {
+        byte nMod10 = (byte)(n % 10UL);
+        ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(tableKind, nMod10);
+        ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(tableKind, nMod10);
+
+        int divisorLength = (int)divisorSquares.Length;
         int threadIndex = index;
+        if (threadIndex >= divisorLength)
+        {
+            return;
+        }
+
         ulong divisorSquare = divisorSquares[threadIndex];
         if (divisorSquare > maxDivisorSquare)
         {

PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs

@@ -32,7 +32,7 @@ public static class Pow2MontgomeryGpuCalculator
 
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		// GpuPrimeWorkLimiter.Release();
 
 		PrimeOrderCalculatorAccelerator.Return(gpu);
@@ -60,7 +60,7 @@ public static class Pow2MontgomeryGpuCalculator
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		// GpuPrimeWorkLimiter.Release();
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 
@@ -86,7 +86,7 @@ public static class Pow2MontgomeryGpuCalculator
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
 		
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 
 		return result;
 	}

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -102,7 +102,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 	private static SmallPrimeDeviceCache GetSmallPrimeDeviceCache(int acceleratorIndex, AcceleratorStream stream)
 	{
-		var pool = _smallPrimeDeviceCache ??= new SmallPrimeDeviceCache[PerfectNumberConstants.RollingAccelerators];
+		var pool = _smallPrimeDeviceCache ??= [];
 		if (pool[acceleratorIndex] is {} cached)
 		{
 			return cached;
@@ -132,7 +132,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 	private static Kernel GetPartialFactorKernel(int acceleratorIndex)
 	{
-		var pool = _partialFactorKernel ??= new Kernel[PerfectNumberConstants.RollingAccelerators];
+		var pool = _partialFactorKernel ??= [];
 		if (pool[acceleratorIndex] is {} cached)
 		{
 			return cached;
@@ -487,7 +487,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 	private static Kernel GetOrderKernel(int acceleratorIndex)
 	{
-		var pool = _orderKernel ??= new Kernel[PerfectNumberConstants.RollingAccelerators];
+		var pool = _orderKernel ??= [];
 		if (pool[acceleratorIndex] is {} cached)
 		{
 			return cached;
@@ -637,7 +637,7 @@ internal static partial class PrimeOrderGpuHeuristics
 		remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
 		stream.Synchronize();
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		for (int i = 0; i < length; i++)
 		{
 			results[i] = (UInt128)resultSpan[i];
@@ -675,7 +675,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 	private static Kernel GetPow2ModKernel(int acceleratorIndex)
 	{
-		var pool = _pow2ModKernel ??= new Kernel[PerfectNumberConstants.RollingAccelerators];
+		var pool = _pow2ModKernel ??= [];
 		if (pool[acceleratorIndex] is {} cached)
 		{
 			return cached;

PerfectNumbers.Core/HeuristicPrimeTester.cs

@@ -3,7 +3,6 @@ using System.Runtime.CompilerServices;
 using ILGPU.Runtime;
 using PerfectNumbers.Core.Gpu;
 using PerfectNumbers.Core.Gpu.Accelerators;
-using ILGPU;
 
 namespace PerfectNumbers.Core;
 
@@ -240,36 +239,33 @@ public sealed class HeuristicPrimeTester
 	}
 
 
-	private static bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
+	private bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
 	{
 		// GpuPrimeWorkLimiter.Acquire();
 		var gpu = HeuristicPrimeTesterAccelerator.Rent(1);
-		var stream = gpu.Stream!;
+		var stream = gpu.Stream;
 		var kernel = gpu.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
+		var flagView = flagView1D.AsContiguous();
 
-		bool compositeDetected;
+		bool compositeDetected = false;
 		int compositeFlag = 0;
 
 		int groupALength = HeuristicPrimeSieves.GroupADivisors.Length;
 
 		flagView1D.CopyFromCPU(stream, ref compositeFlag, 1);
-		var tables = gpu.HeuristicGpuTables;
-        ArrayView1D<ulong, Stride1D.Dense> divisors = tables.GroupADivisors;
-        ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.GroupADivisorSquares;
-
 		kernel.Launch(
 				stream!,
 				1,
 				groupALength,
-				flagView1D,
+				flagView,
 				n,
 				maxDivisorSquare,
-				divisors,
-				divisorSquares);
+				HeuristicGpuDivisorTableKind.GroupA,
+				gpu.HeuristicGpuTables);
 
 		flagView1D.CopyToCPU(stream, ref compositeFlag, 1);
-		stream.Synchronize();
+		stream!.Synchronize();
 
 		compositeDetected = compositeFlag != 0;
 
@@ -300,7 +296,7 @@ public sealed class HeuristicPrimeTester
 					stream,
 						1,
 						divisorLength,
-						flagView1D,
+						flagView,
 						n,
 						maxDivisorSquare,
 						tableKind,

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -898,7 +898,7 @@ public class MersenneDivisorCycles
 			start = end + 1UL;
 		}
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		pool.Return(divisors, clearArray: false);
 		pool.Return(outCycles, clearArray: false);
 
@@ -907,7 +907,7 @@ public class MersenneDivisorCycles
 
 	private static Kernel GetDivisorCycleKernel(int acceleratorIndex)
 	{
-		var pool = _divisorCycleKernel ??= new Kernel[PerfectNumberConstants.RollingAccelerators];
+		var pool = _divisorCycleKernel ??= [];
 		if (pool[acceleratorIndex] is {} cached)
 		{
 			return cached;

PerfectNumbers.Core/MersenneNumberTester.cs

@@ -204,7 +204,7 @@ public sealed class MersenneNumberTester(
 			qBuffer.Dispose();
 		}
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex,stream);
 		// GpuPrimeWorkLimiter.Release();
 		uInt128Pool.Return(qs);
 	}

PerfectNumbers.Core/PerfectNumberConstants.cs

@@ -15,7 +15,7 @@ public static class PerfectNumberConstants
     public const int MaxOddPowersCount = 128;
 	public const int PooledArrayThreshold = 64;
 	public static readonly int RollingAccelerators = 298; //SharedGpuContext.Device.MaxNumThreadsPerGroup;
-	public const int ThreadsByAccelerator = 2;
+	public const int ThreadsByAccelerator = 4096;
     // TODO: Load these limits from the benchmark-driven configuration so CPU and GPU scans stay aligned with the optimal
     // divisor-cycle datasets we generate offline.
     // TODO: Promote these magic numbers into a runtime profile derived from EvenPerfectBitScanner.Benchmarks so we can retune

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -50,7 +50,7 @@ internal static partial class PrimeOrderCalculator
 
 		gpu.OutputUlong.View.CopyToCPU(stream, ref order, 1);
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 
 		// GpuPrimeWorkLimiter.Release();
 		return order;
@@ -114,7 +114,7 @@ internal static partial class PrimeOrderCalculator
 		smallPrimeFactorRemainingSlotView.CopyToCPU(stream, ref remaining, 1);
 		stream.Synchronize();
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		// GpuPrimeWorkLimiter.Release();
 
 		for (int i = 0; i < factorCount; i++)
@@ -164,7 +164,7 @@ internal static partial class PrimeOrderCalculator
 		specialMaxResultView.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
 		
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		// GpuPrimeWorkLimiter.Release();
 		return result != 0;
 	}
@@ -190,7 +190,7 @@ internal static partial class PrimeOrderCalculator
 		gpu.Input.View.CopyToCPU(stream, randomStateSpan);
 		stream.Synchronize();
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		ThreadStaticDeterministicRandomGpu.Exclusive.SetState(randomStateSpan[0]);
 
 		bool factored = factoredSpan[0] != 0;

PerfectNumbers.Core/PrimeTester.cs

@@ -123,7 +123,7 @@ public sealed class PrimeTester
 		outputView.CopyToCPU(stream, ref flag, 1);
 		stream.Synchronize();
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 		// GpuPrimeWorkLimiter.Release();
 
@@ -162,7 +162,7 @@ public sealed class PrimeTester
 
 		outputView.CopyToCPU(stream, ref flag, 1);
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 
 		return flag != 0;
 	}
@@ -239,7 +239,7 @@ public sealed class PrimeTester
 		}
 
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 		// GpuPrimeWorkLimiter.Release();
 	}
@@ -295,7 +295,7 @@ public sealed class PrimeTester
 		resultBuffer.View.CopyToCPU(stream, in results);
 		stream.Synchronize();
 
-		AcceleratorStreamPool.Return(acceleratorIndex);
+		AcceleratorStreamPool.Return(acceleratorIndex, stream);
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 	}