Tie GPU scratch buffers to context leases

PerfectNumbers.Core/Gpu/GpuContextPool.cs

@@ -13,13 +13,14 @@ public static class GpuContextPool
 	{
 		public readonly Context Context;
 		public readonly Accelerator Accelerator;
+		public readonly KernelContainer Kernels;
 
 		public PooledContext(KernelContainer kernels)
 		{
 			Context = SharedGpuContext.Context;
 			Accelerator = SharedGpuContext.Accelerator;
+			Kernels = kernels;
 			GpuStaticTableInitializer.EnsureStaticTables(kernels, Accelerator);
-			kernels.Dispose();
 			// NOTE: Avoid loading/compiling any kernel here to prevent implicit
 			// CL stream/queue creation during accelerator construction.
 			// Some OpenCL drivers are fragile when a queue is created immediately
@@ -34,6 +35,7 @@ public static class GpuContextPool
 			// Clear any per-accelerator cached resources to avoid releasing
 			// after the accelerator is destroyed.
 			PrimeTester.ClearGpuCaches(Accelerator);
+			Kernels.Dispose();
 
 			// These resources are shared between GPU leases
 			//  Accelerator.Dispose();
@@ -53,7 +55,9 @@ public static class GpuContextPool
 		}
 
 		// Create a new accelerator when the pool does not have one available.
-		return new GpuContextLease(new PooledContext(GpuKernelPool.GetKernels()));
+		KernelContainer kernels = GpuKernelPool.GetKernels();
+		var context = new PooledContext(kernels);
+		return new GpuContextLease(context);
 	}
 
 
@@ -111,6 +115,8 @@ public static class GpuContextPool
 		public Context Context => _ctx!.Context;
 		public Accelerator Accelerator => _ctx!.Accelerator;
 
+		public KernelContainer Kernels => _ctx!.Kernels;
+
 		public void Dispose()
 		{
 			if (_disposed)

PerfectNumbers.Core/Gpu/GpuKernelLease.cs

@@ -12,23 +12,22 @@ public sealed class GpuKernelLease
     private static readonly ConcurrentQueue<GpuKernelLease> Pool = new();
 
     private GpuContextLease? _gpu;
-    public KernelContainer Kernels;
+    public KernelContainer Kernels = null!;
     private AcceleratorStream? _stream;
 
-    private GpuKernelLease(KernelContainer kernels)
+    private GpuKernelLease()
 	{
-		Kernels = kernels;
 	}
 
-    internal static GpuKernelLease Rent(GpuContextLease gpu, KernelContainer kernels)
+    internal static GpuKernelLease Rent(GpuContextLease gpu)
     {
         if (!Pool.TryDequeue(out var lease))
         {
-            lease = new GpuKernelLease(kernels);
+            lease = new GpuKernelLease();
         }
 
         lease._gpu = gpu;
-        lease.Kernels = kernels;
+        lease.Kernels = gpu.Kernels;
         lease._stream = null;
         return lease;
     }
@@ -156,9 +155,9 @@ public sealed class GpuKernelLease
 
         _gpu?.Dispose();
         _gpu = null;
+        Kernels = null!;
 
-		Kernels?.Dispose();
-		GpuPrimeWorkLimiter.Release();
+        GpuPrimeWorkLimiter.Release();
         Pool.Enqueue(this);
     }
 

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -299,10 +299,11 @@ public class GpuKernelPool
 
 	public static GpuKernelLease GetKernel(bool useGpuOrder)
 	{
+		_ = useGpuOrder;
+
 		GpuPrimeWorkLimiter.Acquire();
 		var gpu = Rent();
-		var kernels = GetKernels();
-		return GpuKernelLease.Rent(gpu, kernels);
+		return GpuKernelLease.Rent(gpu);
 	}
 
 	/// <summary>