Cache GPU divisor partial data on GPU path

PerfectNumbers.Core/Gpu/ExponentRemainderStepperGpu.cs

@@ -15,17 +15,17 @@ internal struct ExponentRemainderStepperGpu
     private GpuUInt128 _currentResidue;
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public ExponentRemainderStepperGpu(in MontgomeryDivisorData divisor)
+    public ExponentRemainderStepperGpu(in GpuDivisorPartialData divisor)
     {
         ulong modulus = divisor.Modulus;
         _modulus = modulus;
-        _modulusWide = new ReadOnlyGpuUInt128(modulus);
+        _modulusWide = divisor.ModulusWide;
         PreviousExponent = 0UL;
         _currentResidue = new GpuUInt128(1UL);
     }
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public bool MatchesDivisor(in MontgomeryDivisorData divisor)
+    public bool MatchesDivisor(in GpuDivisorPartialData divisor)
     {
         return _modulus == divisor.Modulus;
     }

PerfectNumbers.Core/Gpu/GpuDivisorPartialData.cs

@@ -0,0 +1,38 @@
+using System.Runtime.CompilerServices;
+using PerfectNumbers.Core;
+
+namespace PerfectNumbers.Core.Gpu;
+
+internal readonly struct GpuDivisorPartialData
+{
+    public readonly ulong Modulus;
+    public readonly ReadOnlyGpuUInt128 ModulusWide;
+    private readonly MontgomeryDivisorData _montgomeryData;
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private GpuDivisorPartialData(ulong modulus, ReadOnlyGpuUInt128 modulusWide, in MontgomeryDivisorData montgomeryData)
+    {
+        Modulus = modulus;
+        ModulusWide = modulusWide;
+        _montgomeryData = montgomeryData;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static GpuDivisorPartialData Create(ulong modulus)
+    {
+        MontgomeryDivisorData montgomery = MontgomeryDivisorData.FromModulus(modulus);
+        return new GpuDivisorPartialData(modulus, new ReadOnlyGpuUInt128(modulus), in montgomery);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    public static GpuDivisorPartialData FromMontgomery(in MontgomeryDivisorData divisor)
+    {
+        return new GpuDivisorPartialData(divisor.Modulus, new ReadOnlyGpuUInt128(divisor.Modulus), in divisor);
+    }
+
+    public MontgomeryDivisorData MontgomeryData
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get => _montgomeryData;
+    }
+}

PerfectNumbers.Core/Gpu/Kernels/DivisorByDivisorKernels.cs

@@ -30,7 +30,7 @@ internal static class DivisorByDivisorKernels
 
     public static void EvaluateDivisorWithStepperKernel(
         Index1D _,
-        MontgomeryDivisorData divisor,
+        GpuDivisorPartialData divisor,
         ulong divisorCycle,
         ulong firstCycleRemainder,
         ArrayView<ulong> exponents,
@@ -101,12 +101,12 @@ internal static class DivisorByDivisorKernels
         hits[globalIndex] = montgomeryResult == 1UL ? (byte)1 : (byte)0;
     }
 
-    public static void ComputeMontgomeryExponentKernel(Index1D index, MontgomeryDivisorData divisor, ArrayView<ulong> exponents, ArrayView<ulong> results)
+    public static void ComputeMontgomeryExponentKernel(Index1D index, GpuDivisorPartialData divisor, ArrayView<ulong> exponents, ArrayView<ulong> results)
     {
         ulong modulus = divisor.Modulus;
         // Each divisor flowing through this kernel follows q = 2kp + 1 with k >= 1, so modulus is always odd and exceeds one.
         ulong exponent = exponents[index];
-        ulong montgomeryResult = ULongExtensions.Pow2MontgomeryModWindowedGpuKeepMontgomery(divisor, exponent);
+        ulong montgomeryResult = ULongExtensions.Pow2MontgomeryModWindowedGpuKeepMontgomery(divisor.MontgomeryData, exponent);
         results[index] = montgomeryResult;
     }
 }

PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs

@@ -11,7 +11,7 @@ using ILGPU.Runtime;
 namespace PerfectNumbers.Core.Gpu;
 
 /// <summary>
-/// Scans Mersenne divisors on the GPU for prime exponents p >= 31 using per-divisor Montgomery data.
+/// Scans Mersenne divisors on the GPU for prime exponents p >= 31 using cached GPU divisor partial data.
 /// Consumers must call <see cref="ConfigureFromMaxPrime"/> before invoking other members so the divisor limits are populated.
 /// </summary>
 public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDivisorByDivisorTester
@@ -24,7 +24,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
     // private bool _isConfigured;
 
     private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>> _kernelCache = new();
-    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, MontgomeryDivisorData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>>> _stepperKernelCache = new(AcceleratorReferenceComparer.Instance);
+    private readonly ConcurrentDictionary<Accelerator, Action<Index1D, GpuDivisorPartialData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>>> _stepperKernelCache = new(AcceleratorReferenceComparer.Instance);
     private readonly ConcurrentDictionary<Accelerator, ConcurrentBag<BatchResources>> _resourcePools = new(AcceleratorReferenceComparer.Instance);
     private readonly ConcurrentBag<GpuContextPool.GpuContextLease> _acceleratorPool = new();
     private readonly ConcurrentBag<DivisorScanSession> _sessionPool = new();
@@ -32,8 +32,8 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
     private Action<Index1D, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> GetKernel(Accelerator accelerator) =>
         _kernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>(DivisorByDivisorKernels.CheckKernel));
 
-    private Action<Index1D, MontgomeryDivisorData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>> GetStepperKernel(Accelerator accelerator) =>
+    private Action<Index1D, GpuDivisorPartialData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>> GetStepperKernel(Accelerator accelerator) =>
-        _stepperKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, MontgomeryDivisorData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>>(DivisorByDivisorKernels.EvaluateDivisorWithStepperKernel));
+        _stepperKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, GpuDivisorPartialData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>>(DivisorByDivisorKernels.EvaluateDivisorWithStepperKernel));
 
 
     public int GpuBatchSize
@@ -258,7 +258,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
             for (int i = 0; i < filteredCount; i++)
             {
                 ulong divisorValue = filteredDivisorsSpan[i];
-                MontgomeryDivisorData divisorDataValue = MontgomeryDivisorData.FromModulus(divisorValue);
+                GpuDivisorPartialData divisorDataValue = GpuDivisorPartialData.Create(divisorValue);
                 ulong divisorCycle = ResolveDivisorCycle(divisorValue, prime, divisorDataValue);
                 if (divisorCycle != prime)
                 {
@@ -305,11 +305,11 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
     }
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static ulong ResolveDivisorCycle(ulong divisor, ulong prime, in MontgomeryDivisorData divisorData)
+    private static ulong ResolveDivisorCycle(ulong divisor, ulong prime, in GpuDivisorPartialData divisorData)
     {
         if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentGpu(divisor, prime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
         {
-            return MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
+            return MersenneDivisorCycles.CalculateCycleLengthGpu(divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
         }
 
         return computedCycle;
@@ -384,7 +384,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
         private readonly MersenneNumberDivisorByDivisorGpuTester _owner;
         private readonly GpuContextPool.GpuContextLease _lease;
         private readonly Accelerator _accelerator;
-        private Action<Index1D, MontgomeryDivisorData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>> _stepperKernel = null!;
+        private Action<Index1D, GpuDivisorPartialData, ulong, ulong, ArrayView<ulong>, ArrayView<byte>> _stepperKernel = null!;
         private MemoryBuffer1D<ulong, Stride1D.Dense> _exponentsBuffer = null!;
         private MemoryBuffer1D<byte, Stride1D.Dense> _hitBuffer = null!;
         private ulong[] _hostBuffer = null!;
@@ -459,10 +459,10 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
             //     return;
             // }
 
-            MontgomeryDivisorData cachedData = divisorData;
+            GpuDivisorPartialData cachedData = GpuDivisorPartialData.FromMontgomery(divisorData);
             if (cachedData.Modulus != divisor)
             {
-                cachedData = MontgomeryDivisorData.FromModulus(divisor);
+                cachedData = GpuDivisorPartialData.Create(divisor);
             }
 
             ulong cycle = divisorCycle;
@@ -472,7 +472,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
             {
                 if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentGpu(divisor, firstPrime, cachedData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
                 {
-                    cycle = MersenneDivisorCycles.CalculateCycleLength(divisor, cachedData, skipPrimeOrderHeuristic: primeOrderFailed);
+                    cycle = MersenneDivisorCycles.CalculateCycleLengthGpu(divisor, cachedData, skipPrimeOrderHeuristic: primeOrderFailed);
                 }
                 else
                 {

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -408,6 +408,22 @@ public class MersenneDivisorCycles
         return success;
     }
 
+    internal static bool TryCalculateCycleLengthForExponentGpu(
+        ulong divisor,
+        ulong exponent,
+        in GpuDivisorPartialData divisorData,
+        out ulong cycleLength,
+        out bool primeOrderFailed) =>
+        TryCalculateCycleLengthForExponentGpu(divisor, exponent, divisorData.MontgomeryData, out cycleLength, out primeOrderFailed);
+
+    internal static ulong CalculateCycleLengthGpu(
+        ulong divisor,
+        in GpuDivisorPartialData divisorData,
+        bool skipPrimeOrderHeuristic = false)
+    {
+        return CalculateCycleLength(divisor, divisorData.MontgomeryData, skipPrimeOrderHeuristic);
+    }
+
     private static bool AccumulateFactors(
         ulong value,
         Dictionary<ulong, int> counts)