Add GPU lease warm-up and shared limiter

EvenPerfectBitScanner/Program.cs

@@ -58,14 +58,15 @@ internal static class Program
 			bool startPrimeProvided = _cliArguments.StartPrimeProvided;
 			int threadCount = Math.Max(1, _cliArguments.ThreadCount);
 			int gpuPrimeBatch = Math.Max(1, _cliArguments.GpuPrimeBatch);
+			int gpuPrimeThreads = Math.Max(1, _cliArguments.GpuPrimeThreads);
 			UnboundedTaskScheduler.ConfigureThreadCount(threadCount);
 			PrimeTester.GpuBatchSize = gpuPrimeBatch;
+			GpuPrimeLeaseLimiter.Initialize(gpuPrimeThreads);
 			GpuContextPool.WarmUpPool(threadCount);
-			PrimeTester.WarmUpGpuKernels(threadCount);
+			PrimeTester.WarmUpGpuKernels(gpuPrimeThreads);
 			Console.WriteLine("Starting up threads...");
 			_ = UnboundedTaskScheduler.Instance;
 			int blockSize = Math.Max(1, _cliArguments.BlockSize);
-			int gpuPrimeThreads = Math.Max(1, _cliArguments.GpuPrimeThreads);
 			GpuKernelType kernelType = _cliArguments.KernelType;
 			bool useBitTransform = _cliArguments.UseBitTransform;
 			bool useOrder = _cliArguments.UseOrder;

PerfectNumbers.Core/Gpu/GpuContextPool.cs

@@ -49,6 +49,10 @@ public static class GpuContextPool
 
 	// Allows callers to choose CPU/GPU per use-case, decoupled from ForceCpu.
 	public static GpuContextLease Rent()
+	{
+		GpuPrimeLeaseLimiter.Enter();
+
+		try
 		{
 			if (GpuPool.TryDequeue(out var gpu))
 			{
@@ -58,6 +62,12 @@ public static class GpuContextPool
 			// Create a new accelerator when the pool does not have one available.
 			return new GpuContextLease(new PooledContext());
 		}
+		catch
+		{
+			GpuPrimeLeaseLimiter.Exit();
+			throw;
+		}
+	}
 
 	public static void WarmUpPool(int threadCount)
 	{
@@ -96,23 +106,33 @@ public static class GpuContextPool
 	{
 		ctx.Accelerator.Synchronize();
 		GpuPool.Enqueue(ctx);
+		GpuPrimeLeaseLimiter.Exit();
 	}
 
-	public readonly struct GpuContextLease
+	public struct GpuContextLease
 	{
-		private readonly PooledContext _ctx;
+		private PooledContext? _ctx;
+		private bool _disposed;
 
 		internal GpuContextLease(PooledContext ctx)
 		{
 			_ctx = ctx;
+			_disposed = false;
 		}
 
-		public readonly Context Context => _ctx.Context;
+		public Context Context => _ctx!.Context;
-		public readonly Accelerator Accelerator => _ctx.Accelerator;
+		public Accelerator Accelerator => _ctx!.Accelerator;
 
-		public readonly void Dispose()
+		public void Dispose()
 		{
-			Return(_ctx);
+			if (_disposed)
+			{
+				return;
+			}
+
+			_disposed = true;
+			Return(_ctx!);
+			_ctx = null;
 		}
 	}
 }

PerfectNumbers.Core/Gpu/GpuPrimeLeaseLimiter.cs

@@ -0,0 +1,37 @@
+using System.Threading;
+
+namespace PerfectNumbers.Core.Gpu;
+
+public static class GpuPrimeLeaseLimiter
+{
+        private static SemaphoreSlim _semaphore = new(int.MaxValue, int.MaxValue);
+        private static int _currentLimit = int.MaxValue;
+
+        public static void Initialize(int limit)
+        {
+                if (limit <= 0)
+                {
+                        limit = 1;
+                }
+
+                if (limit == _currentLimit)
+                {
+                        return;
+                }
+
+                var previous = _semaphore;
+                _semaphore = new SemaphoreSlim(limit, limit);
+                _currentLimit = limit;
+                previous.Dispose();
+        }
+
+        public static void Enter()
+        {
+                _semaphore.Wait();
+        }
+
+        public static void Exit()
+        {
+                _semaphore.Release();
+        }
+}

PerfectNumbers.Core/PrimeTester.cs

@@ -159,18 +159,7 @@ public sealed class PrimeTester
 				return;
 			}
 
-			int toWarm = target - WarmedGpuLeaseCount;
+			PrimeTesterGpuContextPool.WarmUp(target, GpuBatchSize);
-			var leases = new PrimeTesterGpuContextPool.PrimeTesterGpuContextLease[toWarm];
-			for (int i = 0; i < toWarm; i++)
-			{
-				leases[i] = PrimeTesterGpuContextPool.Rent(GpuBatchSize);
-			}
-
-			for (int i = 0; i < toWarm; i++)
-			{
-				leases[i].Dispose();
-			}
-
 			WarmedGpuLeaseCount = target;
 		}
 	}
@@ -229,6 +218,8 @@ public sealed class PrimeTester
 	internal static class PrimeTesterGpuContextPool
 	{
 		private static readonly ConcurrentQueue<PrimeTesterGpuContextLease> Pool = new();
+		private static readonly object WarmUpLock = new();
+		private static int WarmedLeaseCount;
 		private static readonly object SharedTablesLock = new();
 		private static readonly Dictionary<string, SharedHeuristicGpuTables> SharedTables = new();
 
@@ -267,7 +258,35 @@ public sealed class PrimeTester
 			}
 		}
 
+		internal static void WarmUp(int target, int minBufferCapacity)
+		{
+			if (target <= 0)
+			{
+				return;
+			}
+
+			lock (WarmUpLock)
+			{
+				if (target <= WarmedLeaseCount)
+				{
+					return;
+				}
+
+				int toCreate = target - WarmedLeaseCount;
+				for (int i = 0; i < toCreate; i++)
+				{
+					Pool.Enqueue(new PrimeTesterGpuContextLease(minBufferCapacity));
+				}
+
+				WarmedLeaseCount = target;
+			}
+		}
+
 		internal static PrimeTesterGpuContextLease Rent(int minBufferCapacity = 1)
+		{
+			GpuPrimeLeaseLimiter.Enter();
+
+			try
 			{
 				if (!Pool.TryDequeue(out var lease))
 				{
@@ -281,11 +300,18 @@ public sealed class PrimeTester
 				lease.ResetReturnFlag();
 				return lease;
 			}
+			catch
+			{
+				GpuPrimeLeaseLimiter.Exit();
+				throw;
+			}
+		}
 
 		internal static void Return(PrimeTesterGpuContextLease lease)
 		{
 			lease.Accelerator.Synchronize();
 			Pool.Enqueue(lease);
+			GpuPrimeLeaseLimiter.Exit();
 		}
 
 		internal static void DisposeAll()
@@ -304,6 +330,8 @@ public sealed class PrimeTester
 
 				SharedTables.Clear();
 			}
+
+			WarmedLeaseCount = 0;
 		}
 
 		internal static void Clear(Accelerator accelerator)
@@ -313,6 +341,8 @@ public sealed class PrimeTester
 				return;
 			}
 
+			lock (WarmUpLock)
+			{
 				var retained = new List<PrimeTesterGpuContextLease>();
 				while (Pool.TryDequeue(out var lease))
 				{
@@ -331,6 +361,9 @@ public sealed class PrimeTester
 				{
 					Pool.Enqueue(retained[i]);
 				}
+
+				WarmedLeaseCount = Pool.Count;
+			}
 		}
 
 		internal sealed class PrimeTesterGpuContextLease : IDisposable