Reduce GC pressure in divisor checking pipeline

This commit is contained in:
Marek Lemańczyk committed 2025-10-15 23:43:27 +02:00
1 parent 571b5a9938
commit 51730bcba3
2 files changed
+55 -14

No files matched your search

+27 -1
View File
@@ -1,4 +1,5 @@
using System.Buffers;
using System.Collections.Generic;
using System.Buffers.Binary;
using System.Diagnostics;
using System.Numerics;
@@ -56,6 +57,9 @@ internal static partial class PrimeOrderCalculator
[ThreadStatic]
private static Dictionary<ulong, int>? s_partialFactorCounts;
[ThreadStatic]
private static List<ulong>? s_heuristicCandidateList;
private static bool IsGpuHeuristicDevice => s_pow2ModeInitialized && s_deviceMode == PrimeOrderHeuristicDevice.Gpu;
private readonly struct Pow2ModeScope
@@ -402,7 +406,9 @@ internal static partial class PrimeOrderCalculator
}
int capacity = config.MaxPowChecks <= 0 ? 64 : config.MaxPowChecks * 4;
List<ulong> candidates = new(capacity);
List<ulong> candidates = AcquireHeuristicCandidateList(capacity);
try
{
FactorEntry[] factorArray = orderFactors.Factors!;
// DebugLog("Building candidates list");
BuildCandidates(order, factorArray, orderFactors.Count, candidates, capacity);
@@ -520,6 +526,11 @@ internal static partial class PrimeOrderCalculator
index += batchSize;
}
}
finally
{
candidates.Clear();
}
// DebugLog("No candidate confirmed");
return false;
}
@@ -823,6 +834,21 @@ internal static partial class PrimeOrderCalculator
return dictionary;
}
private static List<ulong> AcquireHeuristicCandidateList(int capacityHint)
{
List<ulong>? list = s_heuristicCandidateList;
if (list is null)
{
list = new List<ulong>(capacityHint);
s_heuristicCandidateList = list;
return list;
}
list.Clear();
list.EnsureCapacity(capacityHint);
return list;
}
private static PartialFactorResult PartialFactor(ulong value, in PrimeOrderSearchConfig config)
{
if (value <= 1UL)
+28 -13
View File
@@ -346,7 +346,26 @@ public static class ULongExtensions
ulong result = divisor.MontgomeryOne;
ulong nPrime = divisor.NPrime;
ulong[] oddPowers = InitializeMontgomeryOddPowersCpu(divisor, modulus, nPrime, oddPowerCount);
const int StackOddPowerThreshold = PerfectNumberConstants.PooledArrayThreshold;
Span<ulong> oddPowersSpan = oddPowerCount <= StackOddPowerThreshold
? stackalloc ulong[StackOddPowerThreshold]
: default;
Span<ulong> oddPowers;
ulong[]? oddPowersArray = null;
if (!oddPowersSpan.IsEmpty)
{
oddPowers = oddPowersSpan.Slice(0, oddPowerCount);
}
else
{
oddPowersArray = ThreadStaticPools.UlongPool.Rent(oddPowerCount);
oddPowers = oddPowersArray.AsSpan(0, oddPowerCount);
}
InitializeMontgomeryOddPowersCpu(divisor, modulus, nPrime, oddPowers);
int index = bitLength - 1;
while (index >= 0)
@@ -384,9 +403,9 @@ public static class ULongExtensions
index = windowStart - 1;
}
if (oddPowerCount >= PerfectNumberConstants.PooledArrayThreshold)
if (oddPowersArray is not null)
{
ArrayPool<ulong>.Shared.Return(oddPowers);
ThreadStaticPools.UlongPool.Return(oddPowersArray, clearArray: false);
}
if (keepMontgomery)
@@ -533,26 +552,22 @@ public static class ULongExtensions
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static ulong[] InitializeMontgomeryOddPowersCpu(in MontgomeryDivisorData divisor, ulong modulus, ulong nPrime, int oddPowerCount)
private static void InitializeMontgomeryOddPowersCpu(in MontgomeryDivisorData divisor, ulong modulus, ulong nPrime, Span<ulong> oddPowers)
{
ulong[] oddPowers = oddPowerCount < PerfectNumberConstants.PooledArrayThreshold ? new ulong[oddPowerCount] : ArrayPool<ulong>.Shared.Rent(oddPowerCount);
oddPowers[0] = divisor.MontgomeryTwo;
if (oddPowerCount == 1)
if (oddPowers.Length == 1)
{
return oddPowers;
return;
}
ulong previous,
square = divisor.MontgomeryTwoSquared;
ulong previous;
ulong square = divisor.MontgomeryTwoSquared;
for (int i = 1; i < oddPowerCount; i++)
for (int i = 1; i < oddPowers.Length; i++)
{
previous = oddPowers[i - 1];
oddPowers[i] = previous.MontgomeryMultiply(square, modulus, nPrime);
}
return oddPowers;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]