Avoid pooling small buffers in CPU divisor pipeline

This commit is contained in:
Marek Lemańczyk committed 2025-10-11 21:17:18 +02:00
1 parent 8e9390bc5e
commit ad22329814
4 files changed
+102 -41

No files matched your search

+41 -11
View File
@@ -63,7 +63,9 @@ public sealed class DivisorCycleCache
int[]? rentedMissing = null;
Span<int> missingBuffer = length <= StackBufferThreshold
? stackalloc int[length]
: new Span<int>(rentedMissing = ThreadLocalArrayPool<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;
@@ -128,12 +130,16 @@ public sealed class DivisorCycleCache
ulong[]? rentedDivisors = null;
Span<ulong> workDivisors = count <= StackBufferThreshold
? stackalloc ulong[count]
: new Span<ulong>(rentedDivisors = ThreadLocalArrayPool<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 = ThreadLocalArrayPool<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++)
{
@@ -150,7 +156,11 @@ public sealed class DivisorCycleCache
if (rentedDivisors is not null)
{
ThreadLocalArrayPool<ulong>.Shared.Return(rentedDivisors, clearArray: false);
ThreadLocalArrayPool<ulong>.Shared.Return(rentedResults!, 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 = ThreadLocalArrayPool<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 = ThreadLocalArrayPool<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 = ThreadLocalArrayPool<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 = ThreadLocalArrayPool<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);
@@ -235,9 +253,21 @@ public sealed class DivisorCycleCache
if (rentedPow is not null)
{
ThreadLocalArrayPool<ulong>.Shared.Return(rentedPow, clearArray: false);
ThreadLocalArrayPool<ulong>.Shared.Return(rentedOrder!, clearArray: false);
ThreadLocalArrayPool<ulong>.Shared.Return(rentedResult!, clearArray: false);
ThreadLocalArrayPool<byte>.Shared.Return(rentedStatus!, 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();
+17 -4
View File
@@ -567,7 +567,10 @@ public class MersenneDivisorCycles
}
int count = factorCounts.Count;
ulong[] primes = ThreadLocalArrayPool<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
{
ThreadLocalArrayPool<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 = ThreadLocalArrayPool<(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
{
ThreadLocalArrayPool<(ulong, ulong)>.Shared.Return(localCycles, clearArray: false);
if (rentLocalCycles)
{
ThreadLocalArrayPool<(ulong, ulong)>.Shared.Return(localCycles, clearArray: false);
}
}
});
}
+37 -24
View File
@@ -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 = ThreadLocalArrayPool<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 = ThreadLocalArrayPool<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)
return;
}
FactorEntry[]? factors = Factors;
if (factors is not null)
{
Factors = null;
if (_factorsFromPool)
{
Factors = null;
ThreadLocalArrayPool<FactorEntry>.Shared.Return(factors, clearArray: false);
}
_next = s_poolHead;
s_poolHead = this;
}
_factorsFromPool = false;
_next = s_poolHead;
s_poolHead = this;
}
}
}
+7 -2
View File
@@ -1058,7 +1058,10 @@ internal static partial class PrimeOrderCalculator
return PartialFactorResult.Rent(null, cofactor, cofactor == 1UL, 0);
}
FactorEntry[] factors = ThreadLocalArrayPool<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 = ThreadLocalArrayPool<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++)
{