GPU stabilization improvements to prevent calculation crashes

PerfectNumbers.Core/Gpu/Kernels/NttMontgomeryKernels.cs

@@ -30,7 +30,7 @@ internal static class NttMontgomeryKernels
         data[i2] = new GpuUInt128(0UL, diff);
     }
 
-    public static void ToMont64Kernel(Index1D index, AcceleratorStream stream, ArrayView<GpuUInt128> data, ulong modulus, ulong nPrime, ulong r2)
+    public static void ToMont64Kernel(Index1D index, ArrayView<GpuUInt128> data, ulong modulus, ulong nPrime, ulong r2)
     {
         ulong a = data[index].Low;
         ulong v = MontMul64(a, r2, modulus, nPrime);

PerfectNumbers.Core/Gpu/Kernels/Pow2MontgomeryKernels.cs

@@ -1,8 +1,7 @@
 using ILGPU;
 using ILGPU.Runtime;
-using PerfectNumbers.Core;
 
-namespace PerfectNumbers.Core.Gpu;
+namespace PerfectNumbers.Core.Gpu.Kernels;
 
 internal static class Pow2MontgomeryKernels
 {

PerfectNumbers.Core/Gpu/Kernels/PrimeOrderKernels.cs

@@ -1,10 +1,7 @@
-using System;
 using System.Numerics;
 using System.Runtime.CompilerServices;
 using ILGPU;
-using ILGPU.Algorithms.Random;
 using ILGPU.Runtime;
-using PerfectNumbers.Core;
 
 namespace PerfectNumbers.Core.Gpu;
 
@@ -1171,13 +1168,20 @@ internal static partial class PrimeOrderGpuHeuristics
 	}
 
     private static Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>> GetPow2ModWideKernel(Accelerator accelerator)
-    {
-        return Pow2ModKernelWideCache.GetOrAdd(accelerator, static accel =>
-        {
-            var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(Pow2ModKernelWide);
-            var kernel = KernelUtil.GetKernel(loaded);
-            return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
-        });
+	{
+		var pool = _pow2ModWideKernel ??= [];
+		if (pool.TryGetValue(accelerator, out var cached))
+		{
+			return cached;
+		}
+
+		var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(Pow2ModKernelWide);
+
+		var kernel = KernelUtil.GetKernel(loaded);
+
+		cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
+		pool[accelerator] = cached;
+		return cached;
     }
 
 	/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.