Merge pull request #184 from mlemanczyk/codex/modify-pow2montgomerymodwindowedgpu-for-gpu-wlp4qt

Add GPU small prime factor path

PerfectNumbers.Core/Gpu/GpuKernelPool.cs

@@ -37,9 +37,16 @@ public sealed class KernelContainer
             ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>? IncrementalOrder;
     public Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong,
         ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>? Pow2ModOrder;
+    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>>? SmallPrimeFactor;
 
     // Optional device buffer with small divisor cycles (<= 4M). Index = divisor, value = cycle length.
     public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallCycles;
+    public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimeFactorsPrimes;
+    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorsSquares;
+    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorPrimeSlots;
+    public MemoryBuffer1D<int, Stride1D.Dense>? SmallPrimeFactorExponentSlots;
+    public MemoryBuffer1D<int, Stride1D.Dense>? SmallPrimeFactorCountSlot;
+    public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimeFactorRemainingSlot;
     public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastOne;
     public MemoryBuffer1D<ulong, Stride1D.Dense>? SmallPrimesPow2LastOne;
     public MemoryBuffer1D<uint, Stride1D.Dense>? SmallPrimesLastSeven;
@@ -75,6 +82,54 @@ public readonly struct ResiduePrimeViews(
     public readonly ArrayView1D<ulong, Stride1D.Dense> LastSevenPow2 = lastSevenPow2;
 }
 
+public readonly struct SmallPrimeFactorTables
+{
+    public readonly MemoryBuffer1D<uint, Stride1D.Dense> Primes;
+    public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Squares;
+    public readonly int Count;
+
+    public SmallPrimeFactorTables(
+        MemoryBuffer1D<uint, Stride1D.Dense> primes,
+        MemoryBuffer1D<ulong, Stride1D.Dense> squares,
+        int count)
+    {
+        Primes = primes;
+        Squares = squares;
+        Count = count;
+    }
+
+    public ArrayView1D<uint, Stride1D.Dense> PrimesView => Primes.View;
+    public ArrayView1D<ulong, Stride1D.Dense> SquaresView => Squares.View;
+}
+
+public readonly struct SmallPrimeFactorScratch
+{
+    public readonly MemoryBuffer1D<ulong, Stride1D.Dense> PrimeSlots;
+    public readonly MemoryBuffer1D<int, Stride1D.Dense> ExponentSlots;
+    public readonly MemoryBuffer1D<int, Stride1D.Dense> CountSlot;
+    public readonly MemoryBuffer1D<ulong, Stride1D.Dense> RemainingSlot;
+
+    public SmallPrimeFactorScratch(
+        MemoryBuffer1D<ulong, Stride1D.Dense> primeSlots,
+        MemoryBuffer1D<int, Stride1D.Dense> exponentSlots,
+        MemoryBuffer1D<int, Stride1D.Dense> countSlot,
+        MemoryBuffer1D<ulong, Stride1D.Dense> remainingSlot)
+    {
+        PrimeSlots = primeSlots;
+        ExponentSlots = exponentSlots;
+        CountSlot = countSlot;
+        RemainingSlot = remainingSlot;
+    }
+
+    public void Clear()
+    {
+        PrimeSlots.MemSetToZero();
+        ExponentSlots.MemSetToZero();
+        CountSlot.MemSetToZero();
+        RemainingSlot.MemSetToZero();
+    }
+}
+
 public ref struct GpuKernelLease(IDisposable limiter, GpuContextLease gpu, KernelContainer kernels)
 {
     private bool disposedValue;
@@ -189,6 +244,82 @@ public ref struct GpuKernelLease(IDisposable limiter, GpuContextLease gpu, Kerne
         }
     }
 
+    public readonly 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>>(SmallPrimeFactorKernelScan);
+                    var kernel = KernelUtil.GetKernel(loaded);
+                    return kernel.CreateLauncherDelegate<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>>>();
+                });
+            }
+        }
+    }
+
+    private static void SmallPrimeFactorKernelScan(
+        Index1D index,
+        ulong value,
+        uint limit,
+        ArrayView1D<uint, Stride1D.Dense> primes,
+        ArrayView1D<ulong, Stride1D.Dense> squares,
+        int primeCount,
+        ArrayView1D<ulong, Stride1D.Dense> primeSlots,
+        ArrayView1D<int, Stride1D.Dense> exponentSlots,
+        ArrayView1D<int, Stride1D.Dense> countSlot,
+        ArrayView1D<ulong, Stride1D.Dense> remainingSlot)
+    {
+        if (index != 0)
+        {
+            return;
+        }
+
+        uint effectiveLimit = limit == 0 ? uint.MaxValue : limit;
+        ulong remaining = value;
+        int count = 0;
+        int capacity = (int)primeSlots.Length;
+
+        for (int i = 0; i < primeCount && remaining > 1UL && count < capacity; i++)
+        {
+            uint primeCandidate = primes[i];
+            if (primeCandidate == 0U || primeCandidate > effectiveLimit)
+            {
+                break;
+            }
+
+            ulong primeSquare = squares[i];
+            if (primeSquare != 0UL && primeSquare > remaining)
+            {
+                break;
+            }
+
+            ulong primeValue = primeCandidate;
+            if ((remaining % primeValue) != 0UL)
+            {
+                continue;
+            }
+
+            int exponent = 0;
+            do
+            {
+                remaining /= primeValue;
+                exponent++;
+            }
+            while ((remaining % primeValue) == 0UL);
+
+            primeSlots[count] = primeValue;
+            exponentSlots[count] = exponent;
+            count++;
+        }
+
+        countSlot[0] = count;
+        remainingSlot[0] = remaining;
+    }
+
     // TODO: Plumb the small-cycles device buffer into all kernels that can benefit
     // (some already accept it). Consider a compact type (byte/ushort) for memory footprint.
 
@@ -779,6 +910,75 @@ public class GpuKernelPool
         }
     }
 
+    public static SmallPrimeFactorTables EnsureSmallPrimeFactorTables(Accelerator accelerator)
+    {
+        var kernels = GetKernels(accelerator);
+        if (kernels.SmallPrimeFactorsPrimes is { } primeBuffer && kernels.SmallPrimeFactorsSquares is { } squareBuffer)
+        {
+            return new SmallPrimeFactorTables(primeBuffer, squareBuffer, (int)primeBuffer.Length);
+        }
+
+        lock (kernels)
+        {
+            if (kernels.SmallPrimeFactorsPrimes is { } existingPrimes && kernels.SmallPrimeFactorsSquares is { } existingSquares)
+            {
+                return new SmallPrimeFactorTables(existingPrimes, existingSquares, (int)existingPrimes.Length);
+            }
+
+            var hostPrimes = PrimesGenerator.SmallPrimes;
+            var hostSquares = PrimesGenerator.SmallPrimesPow2;
+
+            var devicePrimes = accelerator.Allocate1D<uint>(hostPrimes.Length);
+            devicePrimes.View.CopyFromCPU(hostPrimes);
+            var deviceSquares = accelerator.Allocate1D<ulong>(hostSquares.Length);
+            deviceSquares.View.CopyFromCPU(hostSquares);
+
+            kernels.SmallPrimeFactorsPrimes = devicePrimes;
+            kernels.SmallPrimeFactorsSquares = deviceSquares;
+
+            return new SmallPrimeFactorTables(devicePrimes, deviceSquares, hostPrimes.Length);
+        }
+    }
+
+    public static SmallPrimeFactorScratch EnsureSmallPrimeFactorScratch(Accelerator accelerator, int slotCount)
+    {
+        var kernels = GetKernels(accelerator);
+        lock (kernels)
+        {
+            MemoryBuffer1D<ulong, Stride1D.Dense>? primeSlots = kernels.SmallPrimeFactorPrimeSlots;
+            if (primeSlots is null || primeSlots.Length < slotCount)
+            {
+                primeSlots?.Dispose();
+                primeSlots = accelerator.Allocate1D<ulong>(slotCount);
+                kernels.SmallPrimeFactorPrimeSlots = primeSlots;
+            }
+
+            MemoryBuffer1D<int, Stride1D.Dense>? exponentSlots = kernels.SmallPrimeFactorExponentSlots;
+            if (exponentSlots is null || exponentSlots.Length < slotCount)
+            {
+                exponentSlots?.Dispose();
+                exponentSlots = accelerator.Allocate1D<int>(slotCount);
+                kernels.SmallPrimeFactorExponentSlots = exponentSlots;
+            }
+
+            MemoryBuffer1D<int, Stride1D.Dense>? countSlot = kernels.SmallPrimeFactorCountSlot;
+            if (countSlot is null)
+            {
+                countSlot = accelerator.Allocate1D<int>(1);
+                kernels.SmallPrimeFactorCountSlot = countSlot;
+            }
+
+            MemoryBuffer1D<ulong, Stride1D.Dense>? remainingSlot = kernels.SmallPrimeFactorRemainingSlot;
+            if (remainingSlot is null)
+            {
+                remainingSlot = accelerator.Allocate1D<ulong>(1);
+                kernels.SmallPrimeFactorRemainingSlot = remainingSlot;
+            }
+
+            return new SmallPrimeFactorScratch(primeSlots!, exponentSlots!, countSlot!, remainingSlot!);
+        }
+    }
+
     public static GpuKernelLease GetKernel(bool useGpuOrder)
     {
         var limiter = GpuPrimeWorkLimiter.Acquire();

PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs

@@ -546,7 +546,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
         }
 
         ulong exponent = exponents[index];
-        hits[index] = exponent.Pow2MontgomeryModWindowedGpu(divisor, keepMontgomery: false) == 1UL ? (byte)1 : (byte)0;
+        hits[index] = ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisor, keepMontgomery: false) == 1UL ? (byte)1 : (byte)0;
     }
 
     public sealed class DivisorScanSession : IMersenneNumberDivisorByDivisorTester.IDivisorScanSession
@@ -765,7 +765,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
 
         ulong exponent = exponents[index];
 
-        results[index] = exponent.Pow2MontgomeryModWindowedGpu(divisor, keepMontgomery: true);
+        results[index] = ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisor, keepMontgomery: true);
     }
 
     public ulong DivisorLimit

PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs

@@ -809,7 +809,7 @@ internal static partial class PrimeOrderGpuHeuristics
             }
 
             ulong reduced = phi / factor;
-            if (reduced.Pow2MontgomeryModWindowedGpu(divisor, keepMontgomery: false) == 1UL)
+            if (ULongExtensions.Pow2MontgomeryModWindowedKernel(reduced, divisor, keepMontgomery: false) == 1UL)
             {
                 return false;
             }
@@ -1593,7 +1593,7 @@ internal static partial class PrimeOrderGpuHeuristics
 
     private static bool Pow2EqualsOneKernel(ulong exponent, in MontgomeryDivisorData divisor)
     {
-        return exponent.Pow2MontgomeryModWindowedGpu(divisor, keepMontgomery: false) == 1UL;
+        return ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisor, keepMontgomery: false) == 1UL;
     }
 
     private static ulong CalculateByDoublingKernel(ulong prime)
@@ -1793,7 +1793,7 @@ internal static partial class PrimeOrderGpuHeuristics
             return 0UL;
         }
 
-        return exponent.Pow2MontgomeryModWindowedGpu(divisorData, keepMontgomery: false);
+        return ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisorData, keepMontgomery: false);
     }
 
     private static Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> GetPow2ModKernel(Accelerator accelerator)
@@ -1809,7 +1809,7 @@ internal static partial class PrimeOrderGpuHeuristics
     private static void Pow2ModKernel(Index1D index, ArrayView1D<ulong, Stride1D.Dense> exponents, MontgomeryDivisorData divisor, ArrayView1D<ulong, Stride1D.Dense> remainders)
     {
         ulong exponent = exponents[index];
-        remainders[index] = exponent.Pow2MontgomeryModWindowedGpu(divisor, keepMontgomery: false);
+        remainders[index] = ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisor, keepMontgomery: false);
     }
 
     private static Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>> GetPow2ModWideKernel(Accelerator accelerator)

PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs

@@ -1,3 +1,8 @@
+using System;
+using System.Runtime.InteropServices;
+using ILGPU;
+using ILGPU.Runtime;
+using ILGPU.Runtime.OpenCL;
 using PerfectNumbers.Core.Gpu;
 
 namespace PerfectNumbers.Core;
@@ -16,30 +21,88 @@ internal static partial class PrimeOrderCalculator
                 Span<int> exponentBuffer = exponentBufferArray.AsSpan(0, GpuSmallPrimeFactorSlots);
                 primeBuffer.Clear();
                 exponentBuffer.Clear();
+                remaining = value;
 
-                if (!PrimeOrderGpuHeuristics.TryPartialFactor(value, limit, primeBuffer, exponentBuffer, out int factorCount, out ulong gpuRemaining, out _))
+                try
+                {
+                        var lease = GpuKernelPool.GetKernel(useGpuOrder: true);
+                        try
+                        {
+                                var execution = lease.EnterExecutionScope();
+                                try
+                                {
+                                        Accelerator accelerator = lease.Accelerator;
+                                        AcceleratorStream stream = lease.Stream;
+
+                                        SmallPrimeFactorTables tables = GpuKernelPool.EnsureSmallPrimeFactorTables(accelerator);
+                                        SmallPrimeFactorScratch scratch = GpuKernelPool.EnsureSmallPrimeFactorScratch(accelerator, GpuSmallPrimeFactorSlots);
+                                        scratch.Clear();
+
+                                        var kernel = lease.SmallPrimeFactorKernel;
+                                        kernel(
+                                            stream,
+                                            1,
+                                            value,
+                                            limit,
+                                            tables.PrimesView,
+                                            tables.SquaresView,
+                                            tables.Count,
+                                            scratch.PrimeSlots.View,
+                                            scratch.ExponentSlots.View,
+                                            scratch.CountSlot.View,
+                                            scratch.RemainingSlot.View);
+
+                                        stream.Synchronize();
+
+                                        int factorCount = 0;
+                                        scratch.CountSlot.View.CopyToCPU(ref factorCount, 1);
+                                        factorCount = Math.Min(factorCount, GpuSmallPrimeFactorSlots);
+
+                                        if (factorCount > 0)
+                                        {
+                                                scratch.PrimeSlots.View.CopyToCPU(ref MemoryMarshal.GetReference(primeBuffer), factorCount);
+                                                scratch.ExponentSlots.View.CopyToCPU(ref MemoryMarshal.GetReference(exponentBuffer), factorCount);
+                                        }
+                                        scratch.RemainingSlot.View.CopyToCPU(ref remaining, 1);
+
+                                        for (int i = 0; i < factorCount; i++)
+                                        {
+                                                ulong primeValue = primeBuffer[i];
+                                                int exponent = exponentBuffer[i];
+                                                if (primeValue == 0UL || exponent == 0)
+                                                {
+                                                        continue;
+                                                }
+
+                                                counts[primeValue] = exponent;
+                                        }
+
+                                        return true;
+                                }
+                                finally
+                                {
+                                        execution.Dispose();
+                                }
+                        }
+                        finally
+                        {
+                                lease.Dispose();
+                        }
+                }
+                catch (CLException)
                 {
                         remaining = value;
-                        ThreadStaticPools.UlongPool.Return(primeBufferArray);
-                        ThreadStaticPools.IntPool.Return(exponentBufferArray);
                         return false;
                 }
-
-                remaining = gpuRemaining;
-                for (int i = 0; i < factorCount; i++)
+                catch (Exception ex) when (ex is AcceleratorException or InternalCompilerException or NotSupportedException or InvalidOperationException or AggregateException)
                 {
-                        ulong primeValue = primeBuffer[i];
-                        int exponent = exponentBuffer[i];
-                        if (primeValue == 0UL || exponent == 0)
-                        {
-                                continue;
-                        }
-
-                        counts[primeValue] = exponent;
+                        remaining = value;
+                        return false;
+                }
+                finally
+                {
+                        ThreadStaticPools.UlongPool.Return(primeBufferArray);
+                        ThreadStaticPools.IntPool.Return(exponentBufferArray);
                 }
-
-                ThreadStaticPools.UlongPool.Return(primeBufferArray);
-                ThreadStaticPools.IntPool.Return(exponentBufferArray);
-                return true;
         }
 }

PerfectNumbers.Core/ULongExtensions.cs

@@ -1,6 +1,9 @@
+using System.Collections.Concurrent;
 using System.Numerics;
 using System.Runtime.CompilerServices;
+using ILGPU;
 using ILGPU.Algorithms;
+using ILGPU.Runtime;
 using PerfectNumbers.Core.Gpu;
 
 namespace PerfectNumbers.Core;
@@ -409,13 +412,32 @@ public static class ULongExtensions
             return Pow2MontgomeryModSingleBit(exponent, divisor, keepMontgomery);
         }
 
+        if (Pow2MontgomeryGpuExecutor.TryExecute(exponent, divisor, keepMontgomery, out ulong gpuResult))
+        {
+            return gpuResult;
+        }
+
+        return Pow2MontgomeryModWindowedCpu(exponent, divisor, keepMontgomery);
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    internal static ulong Pow2MontgomeryModWindowedKernel(ulong exponent, in MontgomeryDivisorData divisor, bool keepMontgomery)
+    {
+        ulong modulus = divisor.Modulus;
+        if (exponent == 0UL)
+        {
+            return keepMontgomery ? divisor.MontgomeryOne : 1UL % modulus;
+        }
+
+        if (exponent <= Pow2WindowFallbackThreshold)
+        {
+            return Pow2MontgomeryModSingleBit(exponent, divisor, keepMontgomery);
+        }
+
         int bitLength = GetPortableBitLength(exponent);
         int windowSize = GetWindowSize(bitLength);
-        int oddPowerCount = 1 << (windowSize - 1);
-
         ulong result = divisor.MontgomeryOne;
         ulong nPrime = divisor.NPrime;
-        ulong[] oddPowers = InitializeMontgomeryOddPowersGpu(divisor, modulus, nPrime, oddPowerCount);
 
         int index = bitLength - 1;
         while (index >= 0)
@@ -446,19 +468,113 @@ public static class ULongExtensions
 
             ulong mask = (1UL << windowLength) - 1UL;
             ulong windowValue = (exponent >> windowStart) & mask;
-            int tableIndex = (int)((windowValue - 1UL) >> 1);
-            ulong multiplier = oddPowers[tableIndex];
+            ulong multiplier = ComputeMontgomeryOddPower(windowValue, divisor, modulus, nPrime);
             result = result.MontgomeryMultiply(multiplier, modulus, nPrime);
 
             index = windowStart - 1;
         }
 
-        if (keepMontgomery)
+        if (!keepMontgomery)
         {
-            return result;
+            result = result.MontgomeryMultiply(1UL, modulus, nPrime);
         }
 
-        return result.MontgomeryMultiply(1UL, modulus, nPrime);
+        return result;
+    }
+
+    [MethodImpl(MethodImplOptions.AggressiveInlining)]
+    private static ulong ComputeMontgomeryOddPower(ulong exponent, in MontgomeryDivisorData divisor, ulong modulus, ulong nPrime)
+    {
+        ulong baseValue = divisor.MontgomeryTwo;
+        ulong power = divisor.MontgomeryOne;
+        ulong remaining = exponent;
+
+        while (remaining != 0UL)
+        {
+            if ((remaining & 1UL) != 0UL)
+            {
+                power = power.MontgomeryMultiply(baseValue, modulus, nPrime);
+            }
+
+            remaining >>= 1;
+            if (remaining == 0UL)
+            {
+                break;
+            }
+
+            baseValue = baseValue.MontgomeryMultiply(baseValue, modulus, nPrime);
+        }
+
+        return power;
+    }
+
+    private static class Pow2MontgomeryGpuExecutor
+    {
+        private static readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, byte, ArrayView1D<ulong, Stride1D.Dense>>> KernelCache = new();
+
+        public static bool TryExecute(ulong exponent, in MontgomeryDivisorData divisor, bool keepMontgomery, out ulong result)
+        {
+            result = 0UL;
+
+            GpuKernelLease lease;
+            try
+            {
+                lease = GpuKernelPool.GetKernel(useGpuOrder: true);
+            }
+            catch (Exception)
+            {
+                return false;
+            }
+
+            var execution = lease.EnterExecutionScope();
+            try
+            {
+                Accelerator accelerator = lease.Accelerator;
+                if (accelerator.AcceleratorType == AcceleratorType.CPU)
+                {
+                    return false;
+                }
+
+                var kernel = KernelCache.GetOrAdd(accelerator, static accel =>
+                {
+                    var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, byte, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernel);
+                    var launcher = KernelUtil.GetKernel(loaded);
+                    return launcher.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, byte, ArrayView1D<ulong, Stride1D.Dense>>>();
+                });
+
+                using var exponentBuffer = accelerator.Allocate1D<ulong>(1);
+                using var resultBuffer = accelerator.Allocate1D<ulong>(1);
+
+                exponentBuffer.View.CopyFromCPU(ref exponent, 1);
+                resultBuffer.MemSetToZero();
+
+                byte keepFlag = keepMontgomery ? (byte)1 : (byte)0;
+                AcceleratorStream stream = lease.Stream;
+                kernel(stream, 1, exponentBuffer.View, divisor, keepFlag, resultBuffer.View);
+                stream.Synchronize();
+
+                resultBuffer.View.CopyToCPU(ref result, 1);
+                return true;
+            }
+            catch (AcceleratorException)
+            {
+                return false;
+            }
+            catch (NotSupportedException)
+            {
+                return false;
+            }
+            finally
+            {
+                execution.Dispose();
+                lease.Dispose();
+            }
+        }
+
+        private static void Pow2MontgomeryKernel(Index1D index, ArrayView1D<ulong, Stride1D.Dense> exponents, MontgomeryDivisorData divisor, byte keepMontgomery, ArrayView1D<ulong, Stride1D.Dense> results)
+        {
+            results[index] = Pow2MontgomeryModWindowedKernel(exponents[index], divisor, keepMontgomery != 0);
+        }
     }
 
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -549,26 +665,6 @@ public static class ULongExtensions
         }
     }
 
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    private static ulong[] InitializeMontgomeryOddPowersGpu(in MontgomeryDivisorData divisor, ulong modulus, ulong nPrime, int oddPowerCount)
-    {
-        ulong[] oddPowers = new ulong[PerfectNumberConstants.MaxOddPowersCount];
-        oddPowers[0] = divisor.MontgomeryTwo;
-        if (oddPowerCount == 1)
-        {
-            return oddPowers;
-        }
-
-        ulong square = divisor.MontgomeryTwoSquared;
-        for (int i = 1; i < oddPowerCount; i++)
-        {
-            ulong previous = oddPowers[i - 1];
-            oddPowers[i] = previous.MontgomeryMultiply(square, modulus, nPrime);
-        }
-
-        return oddPowers;
-    }
-
     [MethodImpl(MethodImplOptions.AggressiveInlining)]
     public static ulong Pow2MontgomeryModWithCycleCpu(this ulong exponent, ulong cycleLength, in MontgomeryDivisorData divisor)
     {