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

Accelerator index added for dictionary-less use-cases

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-15 02:21:45 +01:00
1 parent 6cb63b6e73
commit 0540411cad
21 files changed
+855 -778
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/Accelerators/AcceleratorStreamPool.cs
+14 -14
View File
@@ -5,38 +5,38 @@ namespace PerfectNumbers.Core.Gpu.Accelerators;
public static class AcceleratorStreamPool
{
private static readonly Dictionary<Accelerator, ConcurrentQueue<AcceleratorStream>> _streams = new(PerfectNumberConstants .RollingAccelerators << 1);
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 SemaphoreSlim CreateLock() => new(PerfectNumberConstants.ThreadsByStream);
private static SemaphoreSlim CreateLock() => new(PerfectNumberConstants.ThreadsByAccelerator);
private static readonly Dictionary<Accelerator, SemaphoreSlim> _locks = new(PerfectNumberConstants.RollingAccelerators << 1);
// private static int _rented;
public static AcceleratorStream Rent(Accelerator accelerator)
public static AcceleratorStream Rent(int acceleratorIndex)
{
var streamLock = _locks[accelerator];
var queue = _streams[accelerator];
var streamLock = _locks[acceleratorIndex];
var queue = _streams[acceleratorIndex];
streamLock.Wait();
return queue.TryDequeue(out var stream)
? stream
: accelerator.CreateStream();
: _accelerators[acceleratorIndex].CreateStream();
}
public static void Return(AcceleratorStream stream)
public static void Return(int acceleratorIndex, AcceleratorStream stream)
{
Accelerator accelerator = stream.Accelerator;
var streamLock = _locks[accelerator];
var queue = _streams[accelerator];
var streamLock = _locks[acceleratorIndex];
var queue = _streams[acceleratorIndex];
queue.Enqueue(stream);
streamLock.Release();
}
public static void WarmUp(Accelerator accelerator)
public static void WarmUp(int acceleratorIndex)
{
_streams.Add(accelerator, new());
_ = _locks.TryAdd(accelerator, CreateLock());
_streams[acceleratorIndex] = new();
_locks[acceleratorIndex] = CreateLock();
}
}
PerfectNumbers.Core/Gpu/Accelerators/HeuristicPrimeTesterAccelerator.cs
+6 -3
View File
@@ -131,9 +131,12 @@ internal sealed class HeuristicPrimeTesterAccelerator
public HeuristicGpuDivisorTables HeuristicGpuTables => _heuristicDivisorTables;
private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
internal HeuristicPrimeTesterAccelerator(int minBufferCapacity)
{
Accelerator accelerator = AcceleratorPool.Shared.Rent();
var acceleratorIndex = AcceleratorPool.Shared.Rent();
Accelerator accelerator = _accelerators[acceleratorIndex];
Context = accelerator.Context;
Accelerator = accelerator;
@@ -141,7 +144,7 @@ internal sealed class HeuristicPrimeTesterAccelerator
HeuristicTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel));
AcceleratorStream stream = AcceleratorStreamPool.Rent(accelerator);
AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
@@ -149,7 +152,7 @@ internal sealed class HeuristicPrimeTesterAccelerator
_heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
stream.Synchronize();
AcceleratorStreamPool.Return(stream);
AcceleratorStreamPool.Return(acceleratorIndex, stream);
}
public void EnsureCapacity(int minCapacity)
PerfectNumbers.Core/Gpu/Accelerators/LastDigitGpuTables.cs
+4 -3
View File
@@ -6,13 +6,14 @@ namespace PerfectNumbers.Core.Gpu.Accelerators;
internal sealed class LastDigitGpuTables
{
private static readonly Dictionary<Accelerator, LastDigitGpuTables> _sharedTables = new(PerfectNumberConstants.RollingAccelerators);
private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
private static readonly LastDigitGpuTables[] _sharedTables = new LastDigitGpuTables[PerfectNumberConstants.RollingAccelerators];
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static LastDigitGpuTables EnsureStaticTables(Accelerator accelerator) => _sharedTables[accelerator];
internal static LastDigitGpuTables EnsureStaticTables(int acceleratorIndex) => _sharedTables[acceleratorIndex];
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static void WarmUp(Accelerator accelerator, AcceleratorStream stream) => _sharedTables[accelerator] = new LastDigitGpuTables(accelerator, stream);
internal static void WarmUp(int acceleratorIndex, AcceleratorStream stream) => _sharedTables[acceleratorIndex] = new LastDigitGpuTables(_accelerators[acceleratorIndex], stream);
internal LastDigitGpuTables(Accelerator accelerator, AcceleratorStream stream)
{
PerfectNumbers.Core/Gpu/Accelerators/PrimeOrderCalculatorAccelerator.cs
+12 -7
View File
@@ -1,7 +1,6 @@
using System.Runtime.CompilerServices;
using ILGPU;
using ILGPU.Runtime;
using PerfectNumbers.Core.Gpu;
using PerfectNumbers.Core.Gpu.Kernels;
namespace PerfectNumbers.Core.Gpu.Accelerators;
@@ -19,8 +18,8 @@ public sealed class PrimeOrderCalculatorAccelerator
var pool = _pool ??= [];
if (!pool.TryDequeue(out var gpu))
{
Accelerator accelerator = AcceleratorPool.Shared.Rent();
gpu = new(accelerator, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity, PerfectNumberConstants.DefaultSmallPrimeFactorSlotCount, PerfectNumberConstants.DefaultSpecialMaxFactorCapacity);
int acceleratorIndex = AcceleratorPool.Shared.Rent();
gpu = new(acceleratorIndex, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity, PerfectNumberConstants.DefaultSmallPrimeFactorSlotCount, PerfectNumberConstants.DefaultSpecialMaxFactorCapacity);
}
else
{
@@ -65,22 +64,26 @@ public sealed class PrimeOrderCalculatorAccelerator
{
Console.WriteLine($"Preparing accelerator {i}...");
var accelerator = accelerators[i];
AcceleratorStreamPool.WarmUp(accelerator);
AcceleratorStreamPool.WarmUp(i);
// Don't take this from the pool as quick uploads of data to the accelerator consumes much of GPU's memory and throws.
AcceleratorStream stream = accelerator.CreateStream();
LastDigitGpuTables.WarmUp(accelerator, stream);
LastDigitGpuTables.WarmUp(i, stream);
// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
stream.Synchronize();
stream.Dispose();
accelerator.Synchronize();
}
}
#endregion
private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
public readonly Accelerator Accelerator;
public readonly int AcceleratorIndex;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Input;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> OutputUlong;
public MemoryBuffer1D<ulong, Stride1D.Dense> PrimeTestInput;
@@ -161,9 +164,11 @@ public sealed class PrimeOrderCalculatorAccelerator
}
}
public PrimeOrderCalculatorAccelerator(Accelerator accelerator, int factorsCount, int primeTesterCapacity, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
public PrimeOrderCalculatorAccelerator(int acceleratorIndex, int factorsCount, int primeTesterCapacity, int smallPrimeFactorSlotCount, int specialMaxFactorCapacity)
{
var accelerator = _accelerators[acceleratorIndex];
Accelerator = accelerator;
AcceleratorIndex = acceleratorIndex;
Input = accelerator.Allocate1D<ulong>(1);
OutputUlong = accelerator.Allocate1D<ulong>(1);
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
@@ -179,7 +184,7 @@ public sealed class PrimeOrderCalculatorAccelerator
SpecialMaxCandidates = accelerator.Allocate1D<ulong>(specialMaxFactorCapacity);
SpecialMaxResult = accelerator.Allocate1D<ulong>(1);
var sharedTables = LastDigitGpuTables.EnsureStaticTables(accelerator);
var sharedTables = LastDigitGpuTables.EnsureStaticTables(acceleratorIndex);
DevicePrimesLastOne = sharedTables.DevicePrimesLastOne;
DevicePrimesLastSeven = sharedTables.DevicePrimesLastSeven;