Merge pull request #241 from mlemanczyk/codex/remove-executionscope-structure

Remove GPU context creation lock

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -227,7 +227,6 @@ public sealed class DivisorCycleCache
     {
         int length = divisors.Length;
         var gpuLease = GpuKernelPool.GetKernel(useGpuOrder: true);
-        var execution = gpuLease.EnterExecutionScope();
 
         Accelerator accelerator = gpuLease.Accelerator;
         var kernel = _gpuKernelCache.GetOrAdd(accelerator, LoadKernel);
@@ -309,7 +308,6 @@ public sealed class DivisorCycleCache
             ArrayPool<byte>.Shared.Return(rentedStatus!, clearArray: false);
         }
 
-        execution.Dispose();
         gpuLease.Dispose();
     }
 

PerfectNumbers.Core/Gpu/GpuContextPool.cs

@@ -47,7 +47,6 @@ public static class GpuContextPool
 
 	private static readonly ConcurrentQueue<PooledContext> CpuPool = new();
 	private static readonly ConcurrentQueue<PooledContext> GpuPool = new();
-	private static readonly object CreationLock = new();
 
 	public static GpuContextLease Rent()
 	{
@@ -61,20 +60,17 @@ public static class GpuContextPool
 		{
 			if (preferCpu && CpuPool.TryDequeue(out var cpu))
 			{
-				return new GpuContextLease(cpu, CreationLock);
+				return new GpuContextLease(cpu);
 			}
 
 			if (!preferCpu && GpuPool.TryDequeue(out var gpu))
 			{
-				return new GpuContextLease(gpu, CreationLock);
+				return new GpuContextLease(gpu);
 			}
 		}
 
-		// Serialize accelerator creation to avoid concurrent CL device/stream setup across threads.
-		lock (CreationLock)
-		{
-			return new GpuContextLease(new PooledContext(preferCpu), CreationLock);
-		}
+		// Create a new accelerator when the pool does not have one available.
+		return new GpuContextLease(new PooledContext(preferCpu));
 	}
 
 	public static void DisposeAll()
@@ -113,30 +109,24 @@ public static class GpuContextPool
 		}
 
                 else
-                {
-                        lock (CreationLock)
                 {
                         ctx.Dispose();
                 }
         }
-        }
 
     public struct GpuContextLease
     {
         private readonly PooledContext _ctx;
 
-                internal GpuContextLease(PooledContext ctx, object kernelInitLock)
+                internal GpuContextLease(PooledContext ctx)
         {
             _ctx = ctx;
-                        KernelInitLock = kernelInitLock;
                 }
 
         public Context Context => _ctx.Context;
 
         public Accelerator Accelerator => _ctx.Accelerator;
 
-                public object KernelInitLock { get; }
-
                 public object ExecutionLock => _ctx.ExecutionLock;
 
 

PerfectNumbers.Core/Gpu/GpuKernelLease.cs

@@ -14,7 +14,6 @@ public sealed class GpuKernelLease
     private GpuContextLease? _gpu;
     private KernelContainer? _kernels;
     private AcceleratorStream? _stream;
-    private object? _executionLock;
 
     private GpuKernelLease()
     {
@@ -31,7 +30,6 @@ public sealed class GpuKernelLease
         lease._gpu = gpu;
         lease._kernels = kernels;
         lease._stream = null;
-        lease._executionLock = gpu.ExecutionLock;
         return lease;
     }
 
@@ -53,15 +51,11 @@ public sealed class GpuKernelLease
         }
     }
 
-    public ExecutionScope EnterExecutionScope() => new ExecutionScope(_executionLock);
-
     public Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>> OrderKernel
     {
         get
         {
             var accel = Accelerator; // avoid capturing 'this' in lambda
-            lock (_gpu!.Value.KernelInitLock)
-            {
             return KernelContainer.InitOnce(ref _kernels!.Order, () =>
             {
                 var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
@@ -70,15 +64,12 @@ public sealed class GpuKernelLease
             });
         }
     }
-    }
 
     public 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>> Pow2ModKernel
     {
         get
         {
             var accel = Accelerator;
-            lock (_gpu!.Value.KernelInitLock)
-            {
             return KernelContainer.InitOnce(ref _kernels!.Pow2Mod, () =>
             {
                 var loaded = accel.LoadAutoGroupedStreamKernel<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>>(Pow2ModKernels.Pow2ModKernelScan);
@@ -87,15 +78,12 @@ public sealed class GpuKernelLease
             });
         }
     }
-    }
 
     public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>> IncrementalKernel
     {
         get
         {
             var accel = Accelerator;
-            lock (_gpu!.Value.KernelInitLock)
-            {
             return KernelContainer.InitOnce(ref _kernels!.Incremental, () =>
             {
                 var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
@@ -104,15 +92,12 @@ public sealed class GpuKernelLease
             });
         }
     }
-    }
 
     public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>> IncrementalOrderKernel
     {
         get
         {
             var accel = Accelerator;
-            lock (_gpu!.Value.KernelInitLock)
-            {
             return KernelContainer.InitOnce(ref _kernels!.IncrementalOrder, () =>
             {
                 var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
@@ -121,15 +106,12 @@ public sealed class GpuKernelLease
             });
         }
     }
-    }
 
     public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>> Pow2ModOrderKernel
     {
         get
         {
             var accel = Accelerator;
-            lock (_gpu!.Value.KernelInitLock)
-            {
             return KernelContainer.InitOnce(ref _kernels!.Pow2ModOrder, () =>
             {
                 var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
@@ -138,15 +120,12 @@ public sealed class GpuKernelLease
             });
         }
     }
-    }
 
     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>> SmallPrimeFactorKernel
     {
         get
         {
             var accel = Accelerator;
-            lock (_gpu!.Value.KernelInitLock)
-            {
             return KernelContainer.InitOnce(ref _kernels!.SmallPrimeFactor, () =>
             {
                 var loaded = accel.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);
@@ -155,7 +134,6 @@ public sealed class GpuKernelLease
             });
         }
     }
-    }
 
     public void Dispose()
     {
@@ -165,8 +143,6 @@ public sealed class GpuKernelLease
         _gpu?.Dispose();
         _gpu = null;
 
-        _executionLock = null;
-
         _releaseLimiter?.Invoke();
         _releaseLimiter = null;
 
@@ -174,25 +150,4 @@ public sealed class GpuKernelLease
         Pool.Enqueue(this);
     }
 
-    public readonly struct ExecutionScope
-    {
-        private readonly object? _lock;
-
-        public ExecutionScope(object? sync)
-        {
-            _lock = sync;
-            if (_lock is not null)
-            {
-                Monitor.Enter(_lock);
-            }
-        }
-
-        public void Dispose()
-        {
-            if (_lock is not null)
-            {
-                Monitor.Exit(_lock);
-            }
-        }
-    }
 }

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -293,12 +293,10 @@ public class GpuKernelPool
     public static void Run(Action<Accelerator, AcceleratorStream> action)
     {
         var lease = GetKernel(useGpuOrder: true);
-        var execution = lease.EnterExecutionScope();
         var accelerator = lease.Accelerator;
         var stream = accelerator.CreateStream();
         action(accelerator, stream);
         stream.Dispose();
-        execution.Dispose();
         lease.Dispose();
 
     }

PerfectNumbers.Core/Gpu/MersenneNumberIncrementalGpuTester.cs

@@ -17,7 +17,6 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
         throw new NotImplementedException($"GPU incremental scanning requires the device cycle heuristics implementation (kernel: {_kernelType}, GPU order: {_useGpuOrder}).");
 
         var gpuLease = GpuKernelPool.GetKernel(_useGpuOrder);
-        var execution = gpuLease.EnterExecutionScope();
         var accelerator = gpuLease.Accelerator;
         var stream = gpuLease.Stream;
         int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy
@@ -123,7 +122,6 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
 
         ArrayPool<ulong>.Shared.Return(orderArray);
         orderBuffer.Dispose();
-        execution.Dispose();
         gpuLease.Dispose();
     }
 }

PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs

@@ -10,7 +10,6 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
         public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
         {
                 var gpuLease = GpuKernelPool.GetKernel(_useGpuOrder);
-                var execution = gpuLease.EnterExecutionScope();
                 var accelerator = gpuLease.Accelerator;
                 var stream = gpuLease.Stream;
                 int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy and avoids TDR
@@ -61,7 +60,6 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
                 }
 
                 foundBuffer.Dispose();
-                execution.Dispose();
                 gpuLease.Dispose();
         }
 }

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -194,7 +194,6 @@ internal static partial class PrimeOrderGpuHeuristics
         try
         {
             var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-            var execution = lease.EnterExecutionScope();
             Accelerator accelerator = lease.Accelerator;
             AcceleratorStream stream = lease.Stream;
 
@@ -248,7 +247,6 @@ internal static partial class PrimeOrderGpuHeuristics
             countBuffer.Dispose();
             remainingBuffer.Dispose();
             fullyFactoredBuffer.Dispose();
-            execution.Dispose();
             lease.Dispose();
             return true;
         }
@@ -362,7 +360,6 @@ internal static partial class PrimeOrderGpuHeuristics
         order = 0UL;
 
         var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-        var execution = lease.EnterExecutionScope();
         Accelerator accelerator = lease.Accelerator;
         AcceleratorStream stream = lease.Stream;
 
@@ -467,7 +464,6 @@ internal static partial class PrimeOrderGpuHeuristics
             stackProductBuffer.Dispose();
             resultBuffer.Dispose();
             statusBuffer.Dispose();
-            execution.Dispose();
         }
     }
 
@@ -526,7 +522,6 @@ internal static partial class PrimeOrderGpuHeuristics
     private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
     {
         var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-        var execution = lease.EnterExecutionScope();
         Accelerator accelerator = lease.Accelerator;
         AcceleratorStream stream = lease.Stream;
         var kernel = GetPow2ModKernel(accelerator);
@@ -551,7 +546,6 @@ internal static partial class PrimeOrderGpuHeuristics
 			Console.WriteLine($"Exception for {prime} and exponents: {string.Join(",", exponents.ToArray())}.");
 			exponentBuffer.Dispose();
 			remainderBuffer.Dispose();
-			execution.Dispose();
 			lease.Dispose();
 			throw;
 		}
@@ -559,7 +553,6 @@ internal static partial class PrimeOrderGpuHeuristics
         remainderBuffer.View.CopyToCPU(ref MemoryMarshal.GetReference(results), exponents.Length);
         exponentBuffer.Dispose();
         remainderBuffer.Dispose();
-        execution.Dispose();
         lease.Dispose();
         return true;
     }
@@ -587,7 +580,6 @@ internal static partial class PrimeOrderGpuHeuristics
 
         try
         {
-            var execution = lease.EnterExecutionScope();
             Accelerator accelerator = lease.Accelerator;
             AcceleratorStream stream = lease.Stream;
             var kernel = GetPow2ModWideKernel(accelerator);
@@ -628,7 +620,6 @@ internal static partial class PrimeOrderGpuHeuristics
 
             exponentBuffer.Dispose();
             remainderBuffer.Dispose();
-            execution.Dispose();
             return true;
         }
         catch (Exception)

PerfectNumbers.Core/MersenneNumberTester.cs

@@ -147,7 +147,6 @@ public sealed class MersenneNumberTester(
 		UInt128[] qs = qsBuffer;
 
                 var gpuLease = GpuKernelPool.GetKernel(_useGpuOrder);
-                var execution = gpuLease.EnterExecutionScope();
                 var accelerator = gpuLease.Accelerator;
                 var stream = gpuLease.Stream;
                 var orderKernel = gpuLease.OrderKernel!;
@@ -194,7 +193,6 @@ public sealed class MersenneNumberTester(
                         qBuffer.Dispose();
                 }
 
-                execution.Dispose();
                 gpuLease.Dispose();
                 ArrayPool<UInt128>.Shared.Return(qs);
         }

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -22,7 +22,6 @@ internal static partial class PrimeOrderCalculator
                 remaining = value;
 
                 var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-                var execution = lease.EnterExecutionScope();
                 Accelerator accelerator = lease.Accelerator;
                 AcceleratorStream stream = lease.Stream;
 
@@ -64,7 +63,6 @@ internal static partial class PrimeOrderCalculator
                         counts.Add(primeValue, exponent);
                 }
 
-                execution.Dispose();
                 lease.Dispose();
                 ThreadStaticPools.UlongPool.Return(primeBufferArray);
                 ThreadStaticPools.IntPool.Return(exponentBufferArray);

PerfectNumbers.Core/ULongExtensions.Gpu.cs

@@ -417,7 +417,6 @@ public static partial class ULongExtensions
 			ulong result = 0UL;
 
 			GpuKernelLease lease = GpuKernelPool.GetKernel(useGpuOrder: true);
-			var execution = lease.EnterExecutionScope();
 
 			Accelerator accelerator = lease.Accelerator;
 			// Keep this commented out. It should never happen in production code.
@@ -444,7 +443,6 @@ public static partial class ULongExtensions
 
 			resultBuffer.Dispose();
 			exponentBuffer.Dispose();
-			execution.Dispose();
 			lease.Dispose();
 			// Keep this commented. We don't want to catch any exceptions. All should crash the scanner.
 			// catch (AcceleratorException)