GpuPrimeWorkLimiter.Acquire() / .Release() removed to save time on the hot-path with billions of calls. We'll rely on AcceleratorStreamPool limits and synchronization.

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -41,8 +41,9 @@ public sealed class DivisorCycleCache
 	private DivisorCycleCache(int divisorCyclesBatchSize)
 	{
 		_divisorCyclesBatchSize = Math.Max(1, divisorCyclesBatchSize);
-		_accelerator = SharedGpuContext.Device.CreateAccelerator(SharedGpuContext.Context);
-		_gpuKernel = LoadKernel(_accelerator);
+		var accelerator = AcceleratorPool.Shared.Rent();
+		_accelerator = accelerator;
+		_gpuKernel = LoadKernel(accelerator);
 		_snapshot = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
 	}
 
@@ -83,7 +84,6 @@ public sealed class DivisorCycleCache
 				throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
 			}
 
-			// ObserveCycleCacheDivisor(divisor);
 			return cached;
 		}
 
@@ -118,7 +118,6 @@ public sealed class DivisorCycleCache
 					throw new InvalidDataException($"Divisor cycle is missing for {divisor}");
 				}
 
-				// ObserveCycleCacheDivisor(divisor);
 				cycles[i] = cached;
 			}
 			else
@@ -257,18 +256,15 @@ public sealed class DivisorCycleCache
 			powBuffer!.Dispose();
 			orderBuffer!.Dispose();
 			resultBuffer!.Dispose();
-			statusBuffer!.Dispose();			
+			statusBuffer!.Dispose();
 		}
 
 		Accelerator accelerator = _accelerator;
-		// lock(accelerator)
-		{
-			divisorBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
-			powBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
-			orderBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
-			resultBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
-			statusBuffer = accelerator.Allocate1D<byte>(requiredCapacity);			
-		}
+		divisorBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+		powBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+		orderBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+		resultBuffer = accelerator.Allocate1D<ulong>(requiredCapacity);
+		statusBuffer = accelerator.Allocate1D<byte>(requiredCapacity);
 	}
 
 	private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
@@ -312,8 +308,10 @@ public sealed class DivisorCycleCache
 			statusSpan[i] = ByteZero;
 		}
 
-		GpuPrimeWorkLimiter.Acquire();
-		var stream = AcceleratorStreamPool.Rent(_accelerator);// accelerator.CreateStream();
+		// GpuPrimeWorkLimiter.Acquire();
+		int pending;
+		var kernel = _gpuKernel;
+		var stream = AcceleratorStreamPool.Rent(_accelerator);
 
 		divisorBuffer!.View.CopyFromCPU(stream, divisors);
 		powBuffer!.View.CopyFromCPU(stream, powSpan);
@@ -321,8 +319,6 @@ public sealed class DivisorCycleCache
 		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);
@@ -341,7 +337,7 @@ public sealed class DivisorCycleCache
 		}
 		while (pending > 0);
 
-		resultBuffer.View.CopyToCPU(stream,resultSpan);
+		resultBuffer.View.CopyToCPU(stream, resultSpan);
 		stream.Synchronize();
 		AcceleratorStreamPool.Return(stream);
 
@@ -354,7 +350,7 @@ public sealed class DivisorCycleCache
 			ThreadStaticPools.BytePool.Return(rentedStatus!, clearArray: false);
 		}
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 	}
 
 	private static Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> LoadKernel(Accelerator accelerator)

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -280,13 +280,13 @@ public class GpuKernelPool
 	/// <param name="action">Action to run with (Accelerator, Stream).</param>
 	public static void Run(Action<Accelerator, AcceleratorStream> action)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var accelerator = SharedGpuContext.CreateAccelerator();
 		var stream = accelerator.CreateStream();
 		action(accelerator, stream);
 		stream.Synchronize();
 		stream.Dispose();
 		accelerator.Dispose();
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 	}
 }

PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs

@@ -103,7 +103,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
         ulong processedCount;
         ulong lastProcessed;
 
-        GpuPrimeWorkLimiter.Acquire();
+        // GpuPrimeWorkLimiter();
         var accelerator = AcceleratorPool.Shared.Rent();
         var stream = accelerator.CreateStream();
 
@@ -137,7 +137,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 		stream.Dispose();
         ReturnBatchResources(resources);
 		// Monitor.Exit(gpuLease.ExecutionLock);
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		
         if (composite)
         {
@@ -509,7 +509,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
         internal DivisorScanSession(MersenneNumberDivisorByDivisorGpuTester owner)
         {
             _owner = owner;
-            GpuPrimeWorkLimiter.Acquire();
+            // GpuPrimeWorkLimiter.Acquire();
 			Accelerator accelerator = AcceleratorPool.Shared.Rent();
             _accelerator = accelerator;
             _stream = accelerator.CreateStream();

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}).");
 
-        GpuPrimeWorkLimiter.Acquire();
+        // GpuPrimeWorkLimiter.Acquire();
         var accelerator = AcceleratorPool.Shared.Rent();
         var stream = accelerator.CreateStream();
         int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy
@@ -125,6 +125,6 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
 		stream.Dispose();
         ulongPool.Return(orderArray);
 		orderBuffer.Dispose();
-        GpuPrimeWorkLimiter.Release();
+        // GpuPrimeWorkLimiter.Release();
     }
 }

PerfectNumbers.Core/Gpu/MersenneNumberLucasLehmerGpuTester.cs

@@ -159,7 +159,7 @@ public class MersenneNumberLucasLehmerGpuTester
         }
         else
         {
-        	GpuPrimeWorkLimiter.Acquire();
+        	// GpuPrimeWorkLimiter.Acquire();
             var accelerator = AcceleratorPool.Shared.Rent();
             var stream = accelerator.CreateStream();
             var modulus = new GpuUInt128(((UInt128)1 << (int)exponent) - 1UL); // TODO: Cache these Mersenne moduli per exponent so LL GPU runs skip rebuilding them every launch.
@@ -173,7 +173,7 @@ public class MersenneNumberLucasLehmerGpuTester
             stream.Dispose();
         }
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
         return result;
     }
 

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)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		// var accelerator = SharedGpuContext.Accelerator;
 		var accelerator = AcceleratorPool.Shared.Rent();
 		var stream = accelerator.CreateStream();
@@ -71,6 +71,6 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
 
 		stream.Dispose();
 		foundBuffer.Dispose();
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 	}
 }

PerfectNumbers.Core/Gpu/MersenneNumberResidueGpuTester.cs

@@ -25,7 +25,7 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
 	// GPU residue variant: check 2^p % q == 1 for q = 2*p*k + 1.
 	public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var accelerator = AcceleratorPool.Shared.Rent();
 
 		// Monitor.Enter(gpuLease.ExecutionLock);
@@ -110,7 +110,7 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
 		stream.Dispose();
 		// Monitor.Exit(gpuLease.ExecutionLock);
 		gpuKernels.Dispose();
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 	}
 
 	private static Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> GetPow2ModKernel(Accelerator accelerator)

PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs

@@ -10,28 +10,29 @@ public static class Pow2MontgomeryGpuCalculator
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static ulong CalculateKeep(ulong exponent, in MontgomeryDivisorData divisor)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 
+		ulong result = 0UL;
 		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
 		Accelerator? accelerator = gpu.Accelerator;
-		AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
 		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
+		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
+		var kernel = gpu.KeepMontgomeryKernel;
+
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
 
 		exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
 		// We don't need to worry about any left-overs here.
 		// resultBuffer.MemSetToZero(stream);
 
-		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
-		var kernel = gpu.KeepMontgomeryKernel;
 		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
 
 		kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
 
-		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
 		AcceleratorStreamPool.Return(stream);
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 		return result;
@@ -40,50 +41,48 @@ public static class Pow2MontgomeryGpuCalculator
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static ulong CalculateConvert(ulong exponent, in MontgomeryDivisorData divisor)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 
+		ulong result = 0UL;
 		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
-		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
+		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
 
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
 		// We don't need to worry about any left-overs here.
 		// resultBuffer.MemSetToZero(stream);
 
-		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
 		var kernel = gpu.ConvertToStandardKernel;
-
 		kernel.Launch(stream, 1, 1, exponentBuffer.View, divisor, resultBuffer.View);
-
-		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
-		AcceleratorStreamPool.Return(stream);
-		GpuPrimeWorkLimiter.Release();
 
+		AcceleratorStreamPool.Return(stream);
+		// GpuPrimeWorkLimiter.Release();
 		PrimeOrderCalculatorAccelerator.Return(gpu);
+
 		return result;
 	}
 
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static ulong CalculateConvert(PrimeOrderCalculatorAccelerator gpu, ulong exponent, in MontgomeryDivisorData divisor)
 	{
-		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
+		ulong result = 0UL;
 		MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
+		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
+		var kernel = gpu.ConvertToStandardKernel;
 
+		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
 		// We don't need to worry about any left-overs here.
 		// resultBuffer.MemSetToZero(stream);
 
-		MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
-		var kernel = gpu.ConvertToStandardKernel;
 		var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
-
 		kernelLauncher(stream, 1,exponentBuffer.View, divisor, resultBuffer.View);
-	
-		ulong result = 0UL;
 		resultBuffer.View.CopyToCPU(stream, ref result, 1);
 		stream.Synchronize();
+		
 		AcceleratorStreamPool.Return(stream);
 
 		return result;

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -225,7 +225,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 		try
 		{
-			GpuPrimeWorkLimiter.Acquire();
+			// GpuPrimeWorkLimiter.Acquire();
 			Accelerator accelerator = AcceleratorPool.Shared.Rent();
 			AcceleratorStream stream = accelerator.CreateStream();
 
@@ -290,7 +290,7 @@ internal static partial class PrimeOrderGpuHeuristics
 			countBuffer.Dispose();
 			remainingBuffer.Dispose();
 			fullyFactoredBuffer.Dispose();
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 			return true;
 		}
 		catch (CLException ex)
@@ -402,7 +402,7 @@ internal static partial class PrimeOrderGpuHeuristics
 	{
 		order = 0UL;
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		Accelerator accelerator = AcceleratorPool.Shared.Rent();
 		AcceleratorStream stream = accelerator.CreateStream();
 
@@ -494,7 +494,7 @@ internal static partial class PrimeOrderGpuHeuristics
 		stackProductBuffer.Dispose();
 		resultBuffer.Dispose();
 		statusBuffer.Dispose();
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 
 		PrimeOrderKernelStatus kernelStatus = (PrimeOrderKernelStatus)status;
 		if (kernelStatus == PrimeOrderKernelStatus.Fallback)
@@ -574,7 +574,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 	private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		Accelerator accelerator = AcceleratorPool.Shared.Rent();
 		AcceleratorStream? stream = accelerator.CreateStream();
 		// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
@@ -604,7 +604,7 @@ internal static partial class PrimeOrderGpuHeuristics
 			stream = null;
 			exponentBuffer.Dispose();
 			remainderBuffer.Dispose();
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 		}
 		catch (Exception)
 		{
@@ -612,7 +612,7 @@ internal static partial class PrimeOrderGpuHeuristics
 			stream?.Dispose();
 			exponentBuffer.Dispose();
 			remainderBuffer.Dispose();
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 			throw;
 		}
 
@@ -638,7 +638,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
 		GpuUInt128[]? rentedExponents = null;
 		GpuUInt128[]? rentedResults = null;
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
 
 		try
@@ -704,7 +704,7 @@ internal static partial class PrimeOrderGpuHeuristics
 				gpuUInt128Pool.Return(rentedResults!, clearArray: false);
 			}
 
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 		}
 	}
 

PerfectNumbers.Core/HeuristicCombinedPrimeTester.cs

@@ -224,11 +224,11 @@ public sealed class HeuristicCombinedPrimeTester
 
 	private static bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var gpu = HeuristicPrimeTesterAccelerator.Rent(1);
 		var accelerator = gpu.Accelerator;
 		var stream = gpu.Stream;
-		var kernel = HeuristicPrimeTesterAccelerator.GetHeuristicTrialDivisionKernel(accelerator);
+		var kernel = HeuristicPrimeTesterAccelerator.HeuristicTrialDivisionKernel;
 		var flagView1D = gpu.HeuristicFlag.View;
 		var flagView = flagView1D.AsContiguous();
 
@@ -251,7 +251,7 @@ public sealed class HeuristicCombinedPrimeTester
 		flagView1D.CopyToCPU(stream, ref compositeFlag, 1);
 		stream.Synchronize();
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 
 		compositeDetected = compositeFlag != 0;
 		HeuristicPrimeTesterAccelerator.Return(gpu);

PerfectNumbers.Core/HeuristicPrimeTester.cs

@@ -241,7 +241,7 @@ public sealed class HeuristicPrimeTester
 
 	private bool HeuristicTrialDivisionGpuDetectsDivisor(ulong n, ulong maxDivisorSquare, byte nMod10)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var gpu = HeuristicPrimeTesterAccelerator.Rent(1);
 		var accelerator = gpu.Accelerator;
 		var stream = gpu.Stream;
@@ -307,7 +307,7 @@ public sealed class HeuristicPrimeTester
 		}
 
 		gpu.Dispose();
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		return compositeDetected;
 	}
 

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -804,7 +804,7 @@ public class MersenneDivisorCycles
 
 		divisors = pool.Rent(count);
 		outCycles = pool.Rent(count);
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var accelerator = AcceleratorPool.Shared.Rent();
 		var stream = accelerator.CreateStream();
 
@@ -897,7 +897,7 @@ public class MersenneDivisorCycles
 		pool.Return(divisors, clearArray: false);
 		pool.Return(outCycles, clearArray: false);
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 	}
 
 	private static Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> GetDivisorCycleKernel(Accelerator accelerator)

PerfectNumbers.Core/MersenneNumberTester.cs

@@ -147,7 +147,7 @@ public sealed class MersenneNumberTester(
 
 		UInt128[] qs = qsBuffer;
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var accelerator = AcceleratorPool.Shared.Rent();
 		var stream = accelerator.CreateStream();
 		var orderKernel = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.OrderKernelScan).Order!;
@@ -164,11 +164,8 @@ public sealed class MersenneNumberTester(
 		while (offset < idx)
 		{
 			int count = Math.Min(batchSize, idx - offset);
-			// lock (accelerator)
-			{
-				qBuffer = accelerator.Allocate1D<GpuUInt128>(count);
-				orderBuffer = accelerator.Allocate1D<ulong>(count);
-			}
+			qBuffer = accelerator.Allocate1D<GpuUInt128>(count);
+			orderBuffer = accelerator.Allocate1D<ulong>(count);
 			
 			// Convert to device-friendly GpuUInt128 on the fly
 			var tmp = gpuUInt128Pool.Rent(count);
@@ -204,7 +201,7 @@ public sealed class MersenneNumberTester(
 		}
 
 		stream.Dispose();
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		uInt128Pool.Return(qs);
 	}
 

PerfectNumbers.Core/PrimeOrderCalculator.Cpu.cs

@@ -91,30 +91,30 @@ internal static partial class PrimeOrderCalculator
 			return phi;
 		}
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		ulong candidateOrder = InitializeStartingOrderCpu(gpu, prime, phi, divisorData);
 		candidateOrder = ExponentLoweringCpu(candidateOrder, prime, phiFactors, divisorData);
 
 		if (TryConfirmOrderCpu(gpu, prime, candidateOrder, divisorData, config, out PartialFactorResult? orderFactors))
 		{
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 			return candidateOrder;
 		}
 
 		if (config.Mode == PrimeOrderMode.Strict)
 		{
 			orderFactors?.Dispose();
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 			return CalculateByFactorizationGpu(gpu, prime, divisorData);
 		}
 
 		if (TryHeuristicFinishCpu(gpu, prime, candidateOrder, previousOrder, divisorData, config, phiFactors, orderFactors, out ulong order))
 		{
-			GpuPrimeWorkLimiter.Release();
+			// GpuPrimeWorkLimiter.Release();
 			return order;
 		}
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		return candidateOrder;
 	}
 
@@ -1148,7 +1148,7 @@ internal static partial class PrimeOrderCalculator
 			bool pollardRhoDeadlineReached = Stopwatch.GetTimestamp() > deadlineTimestamp;
 			if (!pollardRhoDeadlineReached)
 			{
-				GpuPrimeWorkLimiter.Acquire();
+				// GpuPrimeWorkLimiter.Acquire();
 
 				while (compositeStack.Count > 0)
 				{
@@ -1200,7 +1200,7 @@ internal static partial class PrimeOrderCalculator
 					pollardRhoDeadlineReached = Stopwatch.GetTimestamp() > deadlineTimestamp;
 				}
 
-				GpuPrimeWorkLimiter.Release();
+				// GpuPrimeWorkLimiter.Release();
 			}
 
 			if (pollardRhoDeadlineReached)
@@ -1216,7 +1216,7 @@ internal static partial class PrimeOrderCalculator
 		bool cofactorContainsComposite = false;
 		int pendingCount = pending.Count;
 		int index = 0;
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		for (; index < pendingCount; index++)
 		{
 			PendingEntry entry = pending[index];
@@ -1280,7 +1280,7 @@ internal static partial class PrimeOrderCalculator
 			// cofactorIsPrime = Open.Numeric.Primes.Prime.Numbers.IsPrime(cofactor);
 		}
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 
 		// ArrayPool<FactorEntry> pool = ThreadStaticPools.FactorEntryPool;
 		bool temp;
@@ -1331,19 +1331,14 @@ internal static partial class PrimeOrderCalculator
 			// 	continue;
 			// }
 
-			// factorsArray[arrayIndex] = new FactorEntry(primeValue, exponentValue, true);
 			factorsArray[arrayIndex] = primeValue;
 			exponentsArray[arrayIndex] = exponentValue;
-			// cofactorIsPrimesArray[arrayIndex] = true;
 			arrayIndex++;
 		}
 
-		// Span<FactorEntry> arraySpan = factorsArray.AsSpan(0, arrayIndex);
-		// Span<ulong> arraySpan = factorsArray.AsSpan(0, arrayIndex);
-		// arraySpan.Sort(static (a, b) => a.CompareTo(b));
 		Array.Sort(factorsArray, exponentsArray, 0, arrayIndex);
 		temp = cofactor == 1UL;
-		// result = PartialFactorResult.Rent(factorsArray, exponentsArray, cofactorIsPrimesArray, cofactor, temp, arrayIndex, cofactorIsPrime);
+		
 		result = PartialFactorResult.Rent(factorsArray, exponentsArray, cofactor, temp, arrayIndex, cofactorIsPrime);
 
 	ReturnResult:
@@ -1716,7 +1711,7 @@ internal static partial class PrimeOrderCalculator
 		ulong order = phi;
 		int entryCount = entries.Length;
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		for (int i = 0; i < entryCount; i++)
 		{
 			ulong primeFactor = entries[i].Key;
@@ -1744,7 +1739,7 @@ internal static partial class PrimeOrderCalculator
 			}
 		}
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		return order;
 	}
 

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -39,7 +39,7 @@ internal static partial class PrimeOrderCalculator
 		var kernel = gpu.CheckFactorsKernel;
 		gpu.EnsureCapacity(entryCount, 1);
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		gpu.Pow2ModEntriesToTestOnDevice.View.CopyFromCPU(stream, entries);
 
@@ -51,7 +51,7 @@ internal static partial class PrimeOrderCalculator
 		stream.Synchronize();
 		AcceleratorStreamPool.Return(stream);
 
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		return order;
 	}
 
@@ -66,7 +66,7 @@ internal static partial class PrimeOrderCalculator
 		gpu.EnsureSmallPrimeFactorSlotsCapacity(GpuSmallPrimeFactorSlots);
 		var accelerator = gpu.Accelerator;
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var stream = AcceleratorStreamPool.Rent(accelerator);
 		KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.SmallPrimeFactorKernelScan);
 		SmallPrimeFactorViews tables = GpuKernelPool.GetSmallPrimeFactorTables(kernels);
@@ -113,7 +113,7 @@ internal static partial class PrimeOrderCalculator
 		stream.Synchronize();
 
 		AcceleratorStreamPool.Return(stream);
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 
 		for (int i = 0; i < factorCount; i++)
 		{
@@ -136,7 +136,7 @@ internal static partial class PrimeOrderCalculator
 		gpu.EnsureSpecialMaxFactorsCapacity(factorCount);
 
 		Accelerator accelerator = gpu.Accelerator;
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var stream = AcceleratorStreamPool.Rent(accelerator);
 		var kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.EvaluateSpecialMaxCandidatesKernel);
 
@@ -162,7 +162,7 @@ internal static partial class PrimeOrderCalculator
 		stream.Synchronize();
 		
 		AcceleratorStreamPool.Return(stream);
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 		return result != 0;
 	}
 
@@ -171,11 +171,12 @@ internal static partial class PrimeOrderCalculator
 		Span<ulong> randomStateSpan = stackalloc ulong[1]; 
 		randomStateSpan[0] = ThreadStaticDeterministicRandomGpu.Exclusive.State;
 
+		Span<byte> factoredSpan = stackalloc byte[1];
+		Span<ulong> factorSpan = stackalloc ulong[1];
+		
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
 		gpu.Input.View.CopyFromCPU(stream, randomStateSpan);
 
-		Span<byte> factoredSpan = stackalloc byte[1];
-		Span<ulong> factorSpan = stackalloc ulong[1];
 		var kernelLauncher = gpu.PollardRhoKernel.CreateLauncherDelegate<Action<AcceleratorStream, ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
 
 		kernelLauncher(stream, n, 1, gpu.Input.View, gpu.OutputByte.View, gpu.OutputUlong.View);

PerfectNumbers.Core/PrimeTester.cs

@@ -1,7 +1,6 @@
 using System.Buffers;
 using System.Numerics;
 using System.Runtime.CompilerServices;
-using System.Runtime.InteropServices;
 using ILGPU;
 using ILGPU.Runtime;
 using PerfectNumbers.Core.Gpu;
@@ -93,15 +92,15 @@ public sealed class PrimeTester
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static bool IsPrimeGpu(ulong n)
 	{
-		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+		byte flag = 0;
 
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
 		var inputView = gpu.PrimeTestInput.View;
 		var outputView = gpu.OutputByte.View;
 		var kernel = gpu.SmallPrimeSieveKernel!;
 
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(gpu.Accelerator);
-		// var stream = gpu.Stream!;
 		inputView.CopyFromCPU(stream, ref n, 1);
 
 		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>>>();
@@ -120,13 +119,12 @@ public sealed class PrimeTester
 						gpu.DevicePrimesPow2LastNine.View,
 						outputView);
 
-		byte flag = 0;
 		outputView.CopyToCPU(stream, ref flag, 1);
 		stream.Synchronize();
 
 		AcceleratorStreamPool.Return(stream);
 		PrimeOrderCalculatorAccelerator.Return(gpu);
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 
 		return flag != 0;
 	}
@@ -134,6 +132,7 @@ public sealed class PrimeTester
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static bool IsPrimeGpu(PrimeOrderCalculatorAccelerator gpu, ulong n)
 	{
+		byte flag = 0;
 		var input = gpu.PrimeTestInput;
 		var inputView = input.View;
 		var output = gpu.OutputByte;
@@ -159,7 +158,6 @@ public sealed class PrimeTester
 						gpu.DevicePrimesPow2LastNine.View,
 						outputView);
 
-		byte flag = 0;
 		outputView.CopyToCPU(stream, ref flag, 1);
 		stream.Synchronize();
 		AcceleratorStreamPool.Return(stream);
@@ -195,7 +193,7 @@ public sealed class PrimeTester
 	[MethodImpl(MethodImplOptions.AggressiveInlining)]
 	public static void IsPrimeBatchGpu(ReadOnlySpan<ulong> values, Span<byte> results)
 	{
-		GpuPrimeWorkLimiter.Acquire();
+		// GpuPrimeWorkLimiter.Acquire();
 		var gpu = PrimeOrderCalculatorAccelerator.Rent(GpuBatchSize);
 		var accelerator = gpu.Accelerator;
 		var kernel = gpu.SmallPrimeSieveKernel!;
@@ -240,7 +238,7 @@ public sealed class PrimeTester
 		stream.Synchronize();
 		AcceleratorStreamPool.Return(stream);
 		PrimeOrderCalculatorAccelerator.Return(gpu);
-		GpuPrimeWorkLimiter.Release();
+		// GpuPrimeWorkLimiter.Release();
 	}
 
 	// Expose cache clearing for accelerator disposal coordination
@@ -276,31 +274,27 @@ 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 = PrimeOrderCalculatorAccelerator.Rent(1);
-		var accelerator = gpu.Accelerator;
+		// Check in benchmarks, which implementation was the fastest, is compatible with GPU,
+		// and implement it.
 
 		int length = values.Length;
 		ArrayPool<ulong> pool = ThreadStaticPools.UlongPool;
-		ulong[] temp = pool.Rent(length);
-		values.CopyTo(temp);
 
-		MemoryBuffer1D<ulong, Stride1D.Dense>? inputBuffer;
-		MemoryBuffer1D<byte, Stride1D.Dense>? resultBuffer;
-
-		inputBuffer = accelerator.Allocate1D<ulong>(length);
-		resultBuffer = accelerator.Allocate1D<byte>(length);
+		var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
+		var accelerator = gpu.Accelerator;
+		gpu.EnsureCapacity(0, length);
+		MemoryBuffer1D<ulong, Stride1D.Dense>? inputBuffer = gpu.PrimeTestInput;
+		MemoryBuffer1D<byte, Stride1D.Dense>? resultBuffer = gpu.OutputByte;
 
 		AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
-		inputBuffer.View.CopyFromCPU(stream, ref temp[0], length);
+		inputBuffer.View.CopyFromCPU(stream, values);
+
 		var kernel = GetSharesFactorKernel(accelerator);
 		kernel(stream, length, inputBuffer.View, resultBuffer.View);
 		resultBuffer.View.CopyToCPU(stream, in results);
 		stream.Synchronize();
 
 		AcceleratorStreamPool.Return(stream);
-		pool.Return(temp, clearArray: false);
-		resultBuffer.Dispose();
-		inputBuffer.Dispose();
 		PrimeOrderCalculatorAccelerator.Return(gpu);
 	}
 
@@ -323,6 +317,8 @@ public sealed class PrimeTester
 		// TODO: Swap this handwritten binary GCD for the optimized helper measured in
 		// GpuUInt128BinaryGcdBenchmarks so CPU callers share the faster subtract-less
 		// ladder once the common implementation is promoted into PerfectNumbers.Core.
+		// You have an example in the benchmarks and you already implemented it in
+		// GpuUInt128, iirc.
 		if (u == 0UL)
 		{
 			return v;