mlemanczyk/even-perfect-numbers-scanner
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Pools renamed to FixedCapacityPools. Other fixed capacity collections added. Queues, concurrent queues and standard stacks replaced with fixed capacity stacks. MethodImplOptions.AggressiveInlining and .AggressiveOptimization added to simple methods and getters / setters. Variables optimized in few places.

Browse Source
This commit is contained in:
mlemanczyk committed 2025-12-11 21:10:19 +01:00
1 parent 547d43de84
commit 260aa19e44
25 files changed
+1372 -1300
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/Accelerators/AcceleratorStreamPool.cs
+7 -5
View File
@@ -1,4 +1,4 @@
using System.Collections.Concurrent;
using System.Runtime.CompilerServices;
using ILGPU.Runtime;
namespace PerfectNumbers.Core.Gpu.Accelerators;
@@ -7,13 +7,14 @@ public static class AcceleratorStreamPool
{
private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
private static readonly SemaphoreSlim[] _locks = new SemaphoreSlim[PerfectNumberConstants.RollingAccelerators];
private static readonly ConcurrentQueue<AcceleratorStream>[] _streams = new ConcurrentQueue<AcceleratorStream> [PerfectNumberConstants.RollingAccelerators];
private static readonly ConcurrentFixedCapacityStack<AcceleratorStream>[] _streams = new ConcurrentFixedCapacityStack<AcceleratorStream>[PerfectNumberConstants.RollingAccelerators];
private static SemaphoreSlim CreateLock() => new(PerfectNumberConstants.ThreadsByAccelerator);
// private static int _rented;
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public static AcceleratorStream Rent(int acceleratorIndex)
{
var streamLock = _locks[acceleratorIndex];
@@ -21,22 +22,23 @@ public static class AcceleratorStreamPool
streamLock.Wait();
return queue.TryDequeue(out var stream)
return queue.Pop() is { } stream
? stream
: _accelerators[acceleratorIndex].CreateStream();
}
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public static void Return(int acceleratorIndex, AcceleratorStream stream)
{
var streamLock = _locks[acceleratorIndex];
var queue = _streams[acceleratorIndex];
queue.Enqueue(stream);
queue.Push(stream);
streamLock.Release();
}
public static void WarmUp(int acceleratorIndex)
{
_streams[acceleratorIndex] = new();
_streams[acceleratorIndex] = new(PerfectNumberConstants.DefaultPoolCapacity);
_locks[acceleratorIndex] = CreateLock();
}
}
PerfectNumbers.Core/Gpu/Accelerators/PrimeOrderCalculatorAccelerator.cs
+9 -13
View File
@@ -1,4 +1,3 @@
using System.Collections.Concurrent;
using System.Runtime.CompilerServices;
using ILGPU;
using ILGPU.Runtime;
@@ -11,34 +10,31 @@ public sealed class PrimeOrderCalculatorAccelerator
#region Pool
// [ThreadStatic]
private static ConcurrentQueue<PrimeOrderCalculatorAccelerator>? _pool;
private static ConcurrentFixedCapacityStack<PrimeOrderCalculatorAccelerator>? _pool;
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static PrimeOrderCalculatorAccelerator Rent(int primeTesterCapacity)
{
var pool = _pool ??= [];
if (!pool.TryDequeue(out var gpu))
{
int acceleratorIndex = AcceleratorPool.Shared.Rent();
gpu = new(acceleratorIndex, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity, PerfectNumberConstants.DefaultSmallPrimeFactorSlotCount, PerfectNumberConstants.DefaultSpecialMaxFactorCapacity);
}
else
var pool = _pool ??= new(PerfectNumberConstants.DefaultThreadPoolCapacity);
if (pool.Pop() is { } gpu)
{
gpu.EnsureCapacity(PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity);
return gpu;
}
return gpu;
int acceleratorIndex = AcceleratorPool.Shared.Rent();
return new(acceleratorIndex, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity, PerfectNumberConstants.DefaultSmallPrimeFactorSlotCount, PerfectNumberConstants.DefaultSpecialMaxFactorCapacity);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void Return(PrimeOrderCalculatorAccelerator gpu) => _pool!.Enqueue(gpu);
public static void Return(PrimeOrderCalculatorAccelerator gpu) => _pool!.Push(gpu);
internal static void Clear()
{
var pool = _pool;
if (pool != null)
{
while (pool.TryDequeue(out var gpu))
while (pool.Pop() is { } gpu)
{
gpu.Dispose();
}
@@ -50,7 +46,7 @@ public sealed class PrimeOrderCalculatorAccelerator
var pool = _pool;
if (pool != null)
{
while (pool.TryDequeue(out var gpu))
while (pool.Pop() is { } gpu)
{
gpu.Dispose();
}
PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs
+19 -18
View File
@@ -22,9 +22,9 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
private BigInteger _minK = BigInteger.One;
// private bool _isConfigured;
private readonly ConcurrentBag<MersenneNumberDivisorByDivisorAccelerator>[] _resourcePool = [.. AcceleratorPool.Shared.Accelerators.Select(x => new ConcurrentBag<MersenneNumberDivisorByDivisorAccelerator>())];
private readonly ConcurrentFixedCapacityStack<MersenneNumberDivisorByDivisorAccelerator>[] _resourcePool = [.. AcceleratorPool.Shared.Accelerators.Select(x => new ConcurrentFixedCapacityStack<MersenneNumberDivisorByDivisorAccelerator>(PerfectNumberConstants.DefaultPoolCapacity))];
private readonly ConcurrentBag<DivisorScanSession> _sessionPool = [];
private readonly ConcurrentFixedCapacityStack<DivisorScanSession> _sessionPool = new(PerfectNumberConstants.DefaultPoolCapacity);
public int GpuBatchSize
{
@@ -272,7 +272,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
ArrayView1D<ulong, Stride1D.Dense> exponentViewDevice = exponentBuffer.View;
ArrayView1D<byte, Stride1D.Dense> hitsView = hitsBuffer.View;
ArrayView1D<int, Stride1D.Dense> hitIndexView = hitIndexBuffer.View;
Queue<MontgomeryDivisorData> divisorPool = MontgomeryDivisorDataPool.Shared;
FixedCapacityStack<MontgomeryDivisorData> divisorPool = MontgomeryDivisorDataPool.Shared;
static bool ProcessRange(
PrimeOrderCalculatorAccelerator gpu,
ulong prime,
@@ -291,7 +291,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
Span<int> offsetSpan,
Span<int> countSpan,
Span<ulong> cycleSpan,
Queue<MontgomeryDivisorData> divisorPool,
FixedCapacityStack<MontgomeryDivisorData> divisorPool,
int chunkCountBaseline,
int acceleratorIndex,
ref bool composite,
@@ -598,6 +598,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
_kernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>(DivisorByDivisorKernels.CheckKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>>>();
}
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
internal void Reset()
{
}
@@ -613,7 +614,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
Accelerator accelerator = _accelerator;
int desiredCapacity = requiredCapacity > _capacity ? requiredCapacity : _capacity;
_capacity = desiredCapacity;
// lock (accelerator)
{
_exponentsBuffer = accelerator.Allocate1D<ulong>(desiredCapacity);
@@ -706,15 +707,13 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
return sentinel == 0;
}
public void Return()
{
_owner._sessionPool.Add(this);
}
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public void Return() => _owner._sessionPool.Push(this);
}
public ulong DivisorLimit
{
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
get
{
// The EvenPerfectBitScanner driver configures the GPU tester before exposing it to callers, so the previous synchronization guard remains commented out.
@@ -732,22 +731,22 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
private MersenneNumberDivisorByDivisorAccelerator RentBatchResources(int acceleratorIndex, int capacity)
{
var queue = _resourcePool[acceleratorIndex];
var accelerator = AcceleratorPool.Shared.Accelerators[acceleratorIndex];
if (!queue.TryTake(out MersenneNumberDivisorByDivisorAccelerator? resources))
if (queue.Pop() is { } resources)
{
return new MersenneNumberDivisorByDivisorAccelerator(accelerator, capacity);
resources.EnsureCapacity(accelerator, capacity);
return resources;
}
resources.EnsureCapacity(accelerator, capacity);
return resources;
return new MersenneNumberDivisorByDivisorAccelerator(accelerator, capacity);
}
private void ReturnBatchResources(int acceleratorIndex, MersenneNumberDivisorByDivisorAccelerator resources) => _resourcePool[acceleratorIndex].Add(resources);
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
private void ReturnBatchResources(int acceleratorIndex, MersenneNumberDivisorByDivisorAccelerator resources) => _resourcePool[acceleratorIndex].Push(resources);
private sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
{
@@ -758,6 +757,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
}
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public ulong GetAllowedMaxDivisor(ulong prime)
{
// The configuration is immutable after setup, so the cached limit can be read without locking.
@@ -774,6 +774,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
return ComputeAllowedMaxDivisorGpu(prime, _divisorLimit);
}
[MethodImpl(MethodImplOptions.AggressiveInlining | MethodImplOptions.AggressiveOptimization)]
public IMersenneNumberDivisorByDivisorTester.IDivisorScanSession CreateDivisorSession(PrimeOrderCalculatorAccelerator gpu)
{
// The tester is configured once at startup and the session pool relies on thread-safe collections, so the previous synchronization guard stays commented out.
@@ -793,7 +794,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
// return new DivisorScanSession(this);
// }
if (_sessionPool.TryTake(out DivisorScanSession? session))
if (_sessionPool.Pop() is { } session)
{
session.Reset();
return session;