Merge pull request #153 from mlemanczyk/codex/analyze-caching-efficiency-in-evenperfectbitscanner

Remove Montgomery divisor cache

PerfectNumbers.Core/Cpu/MersenneNumberDivisorByDivisorCpuTester.cs

@@ -192,6 +192,10 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
         Dictionary<ulong, MersenneDivisorCycles.FactorCacheEntry>? factorCache = null;
         DivisorCycleCache cycleCache = DivisorCycleCache.Shared;
 
+        ulong cachedModulus = 0UL;
+        MontgomeryDivisorData cachedDivisorData = default;
+        bool hasCachedDivisorData = false;
+
         byte step10 = (byte)(step % 10UL);
         byte step8 = (byte)(step % 8UL);
         byte step5 = (byte)(step % 5UL);
@@ -222,7 +226,14 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
 
             if (admissible && (remainder8 == 1 || remainder8 == 7) && remainder3 != 0 && remainder5 != 0 && remainder7 != 0 && remainder11 != 0)
             {
-                MontgomeryDivisorData divisorData = MontgomeryDivisorDataCache.Get(candidate);
+                if (!hasCachedDivisorData || cachedModulus != candidate)
+                {
+                    cachedDivisorData = MontgomeryDivisorData.FromModulus(candidate);
+                    cachedModulus = candidate;
+                    hasCachedDivisorData = true;
+                }
+
+                MontgomeryDivisorData divisorData = cachedDivisorData;
                 ulong divisorCycle;
 
                 if (candidate <= PerfectNumberConstants.MaxQForDivisorCycles)
@@ -340,6 +351,9 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
     {
         private readonly MersenneNumberDivisorByDivisorCpuTester _owner;
         private bool _disposed;
+        private bool _hasCachedDivisorData;
+        private ulong _cachedDivisor;
+        private MontgomeryDivisorData _cachedDivisorData;
 
         internal DivisorScanSession(MersenneNumberDivisorByDivisorCpuTester owner)
         {
@@ -349,6 +363,9 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
         internal void Reset()
         {
             _disposed = false;
+            _hasCachedDivisorData = false;
+            _cachedDivisor = 0UL;
+            _cachedDivisorData = default;
         }
 
         public void CheckDivisor(ulong divisor, in MontgomeryDivisorData divisorData, ulong divisorCycle, ReadOnlySpan<ulong> primes, Span<byte> hits)
@@ -367,7 +384,14 @@ public sealed class MersenneNumberDivisorByDivisorCpuTester : IMersenneNumberDiv
             MontgomeryDivisorData cachedData = divisorData;
             if (cachedData.Modulus != divisor)
             {
-                cachedData = MontgomeryDivisorDataCache.Get(divisor);
+                if (!_hasCachedDivisorData || _cachedDivisor != divisor)
+                {
+                    _cachedDivisorData = MontgomeryDivisorData.FromModulus(divisor);
+                    _cachedDivisor = divisor;
+                    _hasCachedDivisorData = true;
+                }
+
+                cachedData = _cachedDivisorData;
             }
 
             if (divisorCycle == 0UL)

PerfectNumbers.Core/DivisorCycleCache.cs

@@ -63,7 +63,9 @@ public sealed class DivisorCycleCache
         int[]? rentedMissing = null;
         Span<int> missingBuffer = length <= StackBufferThreshold
             ? stackalloc int[length]
-            : new Span<int>(rentedMissing = ArrayPool<int>.Shared.Rent(length), 0, length);
+            : length < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<int>(new int[length])
+                : new Span<int>(rentedMissing = ThreadLocalArrayPool<int>.Shared.Rent(length), 0, length);
 
         int missingCount = 0;
 
@@ -106,7 +108,7 @@ public sealed class DivisorCycleCache
 
         if (rentedMissing is not null)
         {
-            ArrayPool<int>.Shared.Return(rentedMissing, clearArray: false);
+            ThreadLocalArrayPool<int>.Shared.Return(rentedMissing, clearArray: false);
         }
     }
 
@@ -128,12 +130,16 @@ public sealed class DivisorCycleCache
         ulong[]? rentedDivisors = null;
         Span<ulong> workDivisors = count <= StackBufferThreshold
             ? stackalloc ulong[count]
-            : new Span<ulong>(rentedDivisors = ArrayPool<ulong>.Shared.Rent(count), 0, count);
+            : count < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<ulong>(new ulong[count])
+                : new Span<ulong>(rentedDivisors = ThreadLocalArrayPool<ulong>.Shared.Rent(count), 0, count);
 
         ulong[]? rentedResults = null;
         Span<ulong> workResults = count <= StackBufferThreshold
             ? stackalloc ulong[count]
-            : new Span<ulong>(rentedResults = ArrayPool<ulong>.Shared.Rent(count), 0, count);
+            : count < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<ulong>(new ulong[count])
+                : new Span<ulong>(rentedResults = ThreadLocalArrayPool<ulong>.Shared.Rent(count), 0, count);
 
         for (int i = 0; i < count; i++)
         {
@@ -149,8 +155,12 @@ public sealed class DivisorCycleCache
 
         if (rentedDivisors is not null)
         {
-            ArrayPool<ulong>.Shared.Return(rentedDivisors, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedResults!, clearArray: false);
+            ThreadLocalArrayPool<ulong>.Shared.Return(rentedDivisors, clearArray: false);
+        }
+
+        if (rentedResults is not null)
+        {
+            ThreadLocalArrayPool<ulong>.Shared.Return(rentedResults, clearArray: false);
         }
     }
 
@@ -172,22 +182,30 @@ public sealed class DivisorCycleCache
         ulong[]? rentedPow = null;
         Span<ulong> powSpan = length <= StackBufferThreshold
             ? stackalloc ulong[length]
-            : new Span<ulong>(rentedPow = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+            : length < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<ulong>(new ulong[length])
+                : new Span<ulong>(rentedPow = ThreadLocalArrayPool<ulong>.Shared.Rent(length), 0, length);
 
         ulong[]? rentedOrder = null;
         Span<ulong> orderSpan = length <= StackBufferThreshold
             ? stackalloc ulong[length]
-            : new Span<ulong>(rentedOrder = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+            : length < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<ulong>(new ulong[length])
+                : new Span<ulong>(rentedOrder = ThreadLocalArrayPool<ulong>.Shared.Rent(length), 0, length);
 
         ulong[]? rentedResult = null;
         Span<ulong> resultSpan = length <= StackBufferThreshold
             ? stackalloc ulong[length]
-            : new Span<ulong>(rentedResult = ArrayPool<ulong>.Shared.Rent(length), 0, length);
+            : length < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<ulong>(new ulong[length])
+                : new Span<ulong>(rentedResult = ThreadLocalArrayPool<ulong>.Shared.Rent(length), 0, length);
 
         byte[]? rentedStatus = null;
         Span<byte> statusSpan = length <= StackBufferThreshold
             ? stackalloc byte[length]
-            : new Span<byte>(rentedStatus = ArrayPool<byte>.Shared.Rent(length), 0, length);
+            : length < PerfectNumberConstants.PooledArrayThreshold
+                ? new Span<byte>(new byte[length])
+                : new Span<byte>(rentedStatus = ThreadLocalArrayPool<byte>.Shared.Rent(length), 0, length);
 
         ref ulong divisorRef = ref MemoryMarshal.GetReference(divisors);
         divisorBuffer.View.CopyFromCPU(ref divisorRef, length);
@@ -234,10 +252,22 @@ public sealed class DivisorCycleCache
         resultSpan.CopyTo(destination);
         if (rentedPow is not null)
         {
-            ArrayPool<ulong>.Shared.Return(rentedPow, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedOrder!, clearArray: false);
-            ArrayPool<ulong>.Shared.Return(rentedResult!, clearArray: false);
-            ArrayPool<byte>.Shared.Return(rentedStatus!, clearArray: false);
+            ThreadLocalArrayPool<ulong>.Shared.Return(rentedPow, clearArray: false);
+        }
+
+        if (rentedOrder is not null)
+        {
+            ThreadLocalArrayPool<ulong>.Shared.Return(rentedOrder, clearArray: false);
+        }
+
+        if (rentedResult is not null)
+        {
+            ThreadLocalArrayPool<ulong>.Shared.Return(rentedResult, clearArray: false);
+        }
+
+        if (rentedStatus is not null)
+        {
+            ThreadLocalArrayPool<byte>.Shared.Return(rentedStatus, clearArray: false);
         }
 
         execution.Dispose();

PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs

@@ -7,7 +7,6 @@ using System.Numerics;
 using ILGPU;
 using ILGPU.Runtime;
 using MontgomeryDivisorData = PerfectNumbers.Core.MontgomeryDivisorData;
-using MontgomeryDivisorDataCache = PerfectNumbers.Core.MontgomeryDivisorDataCache;
 using CycleRemainderStepper = PerfectNumbers.Core.CycleRemainderStepper;
 
 namespace PerfectNumbers.Core.Gpu;

PerfectNumbers.Core/MersenneDivisorCycles.cs

@@ -567,7 +567,10 @@ public class MersenneDivisorCycles
         }
 
         int count = factorCounts.Count;
-        ulong[] primes = ArrayPool<ulong>.Shared.Rent(count);
+        bool rentFromPool = count >= PerfectNumberConstants.PooledArrayThreshold;
+        ulong[] primes = rentFromPool
+            ? ThreadLocalArrayPool<ulong>.Shared.Rent(count)
+            : new ulong[count];
         try
         {
             int index = 0;
@@ -607,7 +610,10 @@ public class MersenneDivisorCycles
         }
         finally
         {
-            ArrayPool<ulong>.Shared.Return(primes, clearArray: false);
+            if (rentFromPool)
+            {
+                ThreadLocalArrayPool<ulong>.Shared.Return(primes, clearArray: false);
+            }
         }
     }
 
@@ -793,7 +799,11 @@ public class MersenneDivisorCycles
             tasks[taskIndex] = Task.Run(() =>
             {
                 ulong rangeLength = localEnd - localStart + 1UL;
-                (ulong divisor, ulong cycle)[] localCycles = ArrayPool<(ulong, ulong)>.Shared.Rent(checked((int)rangeLength));
+                int localLength = checked((int)rangeLength);
+                bool rentLocalCycles = localLength >= PerfectNumberConstants.PooledArrayThreshold;
+                (ulong divisor, ulong cycle)[] localCycles = rentLocalCycles
+                    ? ThreadLocalArrayPool<(ulong, ulong)>.Shared.Rent(localLength)
+                    : new (ulong divisor, ulong cycle)[localLength];
                 int localCycleIndex = 0;
 
                 try
@@ -841,7 +851,10 @@ public class MersenneDivisorCycles
                 }
                 finally
                 {
-                    ArrayPool<(ulong, ulong)>.Shared.Return(localCycles, clearArray: false);
+                    if (rentLocalCycles)
+                    {
+                        ThreadLocalArrayPool<(ulong, ulong)>.Shared.Return(localCycles, clearArray: false);
+                    }
                 }
             });
         }
@@ -856,7 +869,7 @@ public class MersenneDivisorCycles
         ulong start = 3UL;
         var locker = new object();
 
-        var pool = ArrayPool<ulong>.Shared;
+        var pool = ThreadLocalArrayPool<ulong>.Shared;
         ulong batchSizeUL = (ulong)batchSize, d, end;
         int count = (int)Math.Min(batchSizeUL, maxDivisor), i, idx;
         ulong[] divisors, outCycles, validDivisors;

PerfectNumbers.Core/MontgomeryDivisorData.cs

@@ -45,32 +45,4 @@ public readonly struct MontgomeryDivisorData(ulong modulus, ulong nPrime, ulong
             montgomeryTwo,
             montgomeryTwoSquared);
     }
-
-}
-
-internal static class MontgomeryDivisorDataCache
-{
-    [ThreadStatic]
-    private static bool s_hasCachedValue;
-
-    [ThreadStatic]
-    private static ulong s_cachedModulus;
-
-    [ThreadStatic]
-    private static MontgomeryDivisorData s_cachedData;
-
-    [MethodImpl(MethodImplOptions.AggressiveInlining)]
-    public static MontgomeryDivisorData Get(ulong modulus)
-    {
-        if (s_hasCachedValue && s_cachedModulus == modulus)
-        {
-            return s_cachedData;
-        }
-
-        MontgomeryDivisorData computed = MontgomeryDivisorData.FromModulus(modulus);
-        s_cachedModulus = modulus;
-        s_cachedData = computed;
-        s_hasCachedValue = true;
-        return computed;
-    }
 }

PerfectNumbers.Core/PartialFactorResult.cs

@@ -1,19 +1,26 @@
 using System;
-using System.Buffers;
 
 namespace PerfectNumbers.Core
 {
     public sealed class PartialFactorResult
     {
-		[ThreadStatic]
-		private static PartialFactorResult? s_poolHead;
+        [ThreadStatic]
+        private static PartialFactorResult? s_poolHead;
+
+        public static readonly PartialFactorResult Empty = new()
+        {
+            Cofactor = 1UL,
+            FullyFactored = true,
+        };
 
         public FactorEntry[]? Factors;
-		private PartialFactorResult? _next;
         public ulong Cofactor;
         public bool FullyFactored;
         public int Count;
 
+        private PartialFactorResult? _next;
+        private bool _factorsFromPool;
+
         private PartialFactorResult()
         {
         }
@@ -30,52 +37,58 @@ namespace PerfectNumbers.Core
                 s_poolHead = instance._next;
             }
 
+            FactorEntry[]? factorArray = factors;
             instance._next = null;
-            instance.Factors = factors;
+            instance.Factors = factorArray;
             instance.Cofactor = cofactor;
             instance.FullyFactored = fullyFactored;
             instance.Count = count;
+            instance._factorsFromPool = factorArray is not null && factorArray.Length >= PerfectNumberConstants.PooledArrayThreshold;
             return instance;
         }
 
-        public static readonly PartialFactorResult Empty = new()
-		{
-            Cofactor = 1UL,
-            FullyFactored = true,
-        };
-
         public PartialFactorResult WithAdditionalPrime(ulong prime)
         {
             FactorEntry[]? source = Factors;
             if (source is null || Count == 0)
             {
-                FactorEntry[] local = ArrayPool<FactorEntry>.Shared.Rent(1);
-                local[0] = new FactorEntry(prime);
-                return Rent(local, 1UL, true, 1);
+                FactorEntry[] single = new FactorEntry[1];
+                single[0] = new FactorEntry(prime);
+                return Rent(single, 1UL, true, 1);
             }
 
-            FactorEntry[] extended = ArrayPool<FactorEntry>.Shared.Rent(Count + 1);
+            int newCount = Count + 1;
+            FactorEntry[] extended = newCount >= PerfectNumberConstants.PooledArrayThreshold
+                ? ThreadLocalArrayPool<FactorEntry>.Shared.Rent(newCount)
+                : new FactorEntry[newCount];
+
             Array.Copy(source, 0, extended, 0, Count);
             extended[Count] = new FactorEntry(prime);
-            Span<FactorEntry> span = extended.AsSpan(0, Count + 1);
+            Span<FactorEntry> span = extended.AsSpan(0, newCount);
             span.Sort(static (a, b) => a.Value.CompareTo(b.Value));
-            return Rent(extended, 1UL, true, Count + 1);
+            return Rent(extended, 1UL, true, newCount);
         }
 
         public void Dispose()
         {
-            if (this != Empty)
+            if (this == Empty)
             {
-                FactorEntry[]? factors = Factors;
-                if (factors is not null)
-                {
-                    Factors = null;
-                    ArrayPool<FactorEntry>.Shared.Return(factors, clearArray: false);
-                }
-
-                _next = s_poolHead;
-                s_poolHead = this;
+                return;
             }
+
+            FactorEntry[]? factors = Factors;
+            if (factors is not null)
+            {
+                Factors = null;
+                if (_factorsFromPool)
+                {
+                    ThreadLocalArrayPool<FactorEntry>.Shared.Return(factors, clearArray: false);
+                }
+            }
+
+            _factorsFromPool = false;
+            _next = s_poolHead;
+            s_poolHead = this;
         }
     }
 }

PerfectNumbers.Core/PrimeFactorPartialResult.cs

@@ -81,12 +81,12 @@ namespace PerfectNumbers.Core
             {
                 if (_primes is not null)
                 {
-                    ArrayPool<ulong>.Shared.Return(_primes, clearArray: false);
+                    ThreadLocalArrayPool<ulong>.Shared.Return(_primes, clearArray: false);
                 }
 
                 if (_exponents is not null)
                 {
-                    ArrayPool<byte>.Shared.Return(_exponents, clearArray: false);
+                    ThreadLocalArrayPool<byte>.Shared.Return(_exponents, clearArray: false);
                 }
 
                 _arraysFromPool = false;
@@ -119,12 +119,12 @@ namespace PerfectNumbers.Core
             {
                 if (_primes is not null)
                 {
-                    ArrayPool<ulong>.Shared.Return(_primes, clearArray: false);
+                    ThreadLocalArrayPool<ulong>.Shared.Return(_primes, clearArray: false);
                 }
 
                 if (_exponents is not null)
                 {
-                    ArrayPool<byte>.Shared.Return(_exponents, clearArray: false);
+                    ThreadLocalArrayPool<byte>.Shared.Return(_exponents, clearArray: false);
                 }
 
                 _arraysFromPool = false;
@@ -134,8 +134,8 @@ namespace PerfectNumbers.Core
 
             if (capacity >= PerfectNumberConstants.PooledArrayThreshold)
             {
-                _primes = ArrayPool<ulong>.Shared.Rent(capacity);
-                _exponents = ArrayPool<byte>.Shared.Rent(capacity);
+                _primes = ThreadLocalArrayPool<ulong>.Shared.Rent(capacity);
+                _exponents = ThreadLocalArrayPool<byte>.Shared.Rent(capacity);
                 _arraysFromPool = true;
             }
             else

PerfectNumbers.Core/PrimeOrderCalculator.Wide.cs

@@ -162,7 +162,7 @@ internal static partial class PrimeOrderCalculator
                 return order;
             }
 
-            tempArray = ArrayPool<FactorEntry128>.Shared.Rent(length + 1);
+            tempArray = ThreadLocalArrayPool<FactorEntry128>.Shared.Rent(length + 1);
             Span<FactorEntry128> buffer = tempArray.AsSpan(0, length);
             factorSpan.CopyTo(buffer);
 
@@ -200,7 +200,7 @@ internal static partial class PrimeOrderCalculator
         {
             if (tempArray is not null)
             {
-                ArrayPool<FactorEntry128>.Shared.Return(tempArray, clearArray: false);
+                ThreadLocalArrayPool<FactorEntry128>.Shared.Return(tempArray, clearArray: false);
             }
         }
     }
@@ -664,7 +664,7 @@ internal static partial class PrimeOrderCalculator
             return new PartialFactorResult128(null, value, false, 0);
         }
 
-        FactorEntry128[]? rented = counts.Count > 0 ? ArrayPool<FactorEntry128>.Shared.Rent(counts.Count) : null;
+        FactorEntry128[]? rented = counts.Count > 0 ? ThreadLocalArrayPool<FactorEntry128>.Shared.Rent(counts.Count) : null;
         int index = 0;
         if (rented is not null)
         {
@@ -682,7 +682,7 @@ internal static partial class PrimeOrderCalculator
             span.Sort(static (a, b) => a.Value.CompareTo(b.Value));
             resultArray = new FactorEntry128[index];
             span.CopyTo(resultArray);
-            ArrayPool<FactorEntry128>.Shared.Return(rented, clearArray: false);
+            ThreadLocalArrayPool<FactorEntry128>.Shared.Return(rented, clearArray: false);
         }
 
         bool fullyFactored = cofactor == UInt128.One;

PerfectNumbers.Core/PrimeOrderCalculator.cs

@@ -1058,7 +1058,10 @@ internal static partial class PrimeOrderCalculator
                     return PartialFactorResult.Rent(null, cofactor, cofactor == 1UL, 0);
                 }
 
-                FactorEntry[] factors = ArrayPool<FactorEntry>.Shared.Rent(counts.Count);
+                int dictionaryCount = counts.Count;
+                FactorEntry[] factors = dictionaryCount >= PerfectNumberConstants.PooledArrayThreshold
+                    ? ThreadLocalArrayPool<FactorEntry>.Shared.Rent(dictionaryCount)
+                    : new FactorEntry[dictionaryCount];
                 int index = 0;
                 foreach (KeyValuePair<ulong, int> entry in counts)
                 {
@@ -1092,7 +1095,9 @@ internal static partial class PrimeOrderCalculator
                 return PartialFactorResult.Rent(null, cofactor, cofactor == 1UL, 0);
             }
 
-            FactorEntry[] array = ArrayPool<FactorEntry>.Shared.Rent(actualCount);
+            FactorEntry[] array = actualCount >= PerfectNumberConstants.PooledArrayThreshold
+                ? ThreadLocalArrayPool<FactorEntry>.Shared.Rent(actualCount)
+                : new FactorEntry[actualCount];
             int arrayIndex = 0;
             for (int i = 0; i < factorCount; i++)
             {

PerfectNumbers.Core/ThreadLocalArrayPool.cs

@@ -0,0 +1,27 @@
+using System;
+using System.Buffers;
+using System.Runtime.CompilerServices;
+
+namespace PerfectNumbers.Core;
+
+internal static class ThreadLocalArrayPool<T>
+{
+    [ThreadStatic]
+    private static ArrayPool<T>? s_threadPool;
+
+    public static ArrayPool<T> Shared
+    {
+        [MethodImpl(MethodImplOptions.AggressiveInlining)]
+        get
+        {
+            ArrayPool<T>? pool = s_threadPool;
+            if (pool is null)
+            {
+                pool = ArrayPool<T>.Create();
+                s_threadPool = pool;
+            }
+
+            return pool;
+        }
+    }
+}