Functionality and classes separated. Kernel initialization made dependent on a flag. Cleanup.
This commit is contained in:
mlemanczyk committed
1 parent
45dfa3b724
commit
0de9871529
20 files changed
+845
-619
No files matched your search
@@ -0,0 +1,13 @@
|
||||
using System.Runtime.CompilerServices;
|
||||
using ILGPU.Runtime;
|
||||
|
||||
namespace PerfectNumbers.Core.Gpu;
|
||||
|
||||
internal sealed class AcceleratorReferenceComparer : IEqualityComparer<Accelerator>
|
||||
{
|
||||
internal static AcceleratorReferenceComparer Instance { get; } = new();
|
||||
|
||||
public bool Equals(Accelerator? x, Accelerator? y) => ReferenceEquals(x, y);
|
||||
|
||||
public int GetHashCode(Accelerator obj) => RuntimeHelpers.GetHashCode(obj);
|
||||
}
|
||||
@@ -0,0 +1,142 @@
|
||||
using System.Collections.Concurrent;
|
||||
using ILGPU;
|
||||
using ILGPU.Runtime;
|
||||
|
||||
namespace PerfectNumbers.Core.Gpu.Accelerators;
|
||||
|
||||
internal sealed class HeuristicGpuTables
|
||||
{
|
||||
private static readonly ConcurrentDictionary<Accelerator, HeuristicGpuTables> _sharedTables = new(20_480, PerfectNumberConstants.RollingAccelerators);
|
||||
|
||||
internal static HeuristicGpuTables EnsureStaticTables(Accelerator accelerator, AcceleratorStream stream)
|
||||
=> _sharedTables.GetOrAdd(accelerator, _ => new HeuristicGpuTables(accelerator, stream));
|
||||
|
||||
internal HeuristicGpuTables(Accelerator accelerator, AcceleratorStream stream)
|
||||
{
|
||||
// lock (accelerator)
|
||||
{
|
||||
var heuristicGroupA = HeuristicPrimeSieves.GroupADivisorsStorage;
|
||||
HeuristicGroupADivisors = CopySpanToDevice(accelerator, stream, heuristicGroupA);
|
||||
|
||||
var heuristicGroupASquares = HeuristicPrimeSieves.GroupADivisorSquaresStorage;
|
||||
HeuristicGroupADivisorSquares = CopySpanToDevice(accelerator, stream, heuristicGroupASquares);
|
||||
|
||||
HeuristicGroupBDivisorsEnding1 = CopyUintSpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesLastOneWithoutLastThree);
|
||||
HeuristicGroupBDivisorSquaresEnding1 = CopySpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesPow2LastOneWithoutLastThree);
|
||||
|
||||
HeuristicGroupBDivisorsEnding7 = CopyUintSpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesLastSevenWithoutLastThree);
|
||||
HeuristicGroupBDivisorSquaresEnding7 = CopySpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesPow2LastSevenWithoutLastThree);
|
||||
|
||||
HeuristicGroupBDivisorsEnding9 = CopyUintSpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesLastNineWithoutLastThree);
|
||||
HeuristicGroupBDivisorSquaresEnding9 = CopySpanToDevice(accelerator, stream, DivisorGenerator.SmallPrimesPow2LastNineWithoutLastThree);
|
||||
|
||||
var combinedEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1;
|
||||
HeuristicCombinedDivisorsEnding1 = CopySpanToDevice(accelerator, stream, combinedEnding1);
|
||||
var combinedSquaresEnding1 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding1Squares;
|
||||
HeuristicCombinedDivisorSquaresEnding1 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding1);
|
||||
|
||||
var combinedEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3;
|
||||
HeuristicCombinedDivisorsEnding3 = CopySpanToDevice(accelerator, stream, combinedEnding3);
|
||||
var combinedSquaresEnding3 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding3Squares;
|
||||
HeuristicCombinedDivisorSquaresEnding3 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding3);
|
||||
|
||||
var combinedEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7;
|
||||
HeuristicCombinedDivisorsEnding7 = CopySpanToDevice(accelerator, stream, combinedEnding7);
|
||||
var combinedSquaresEnding7 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding7Squares;
|
||||
HeuristicCombinedDivisorSquaresEnding7 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding7);
|
||||
|
||||
var combinedEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9;
|
||||
HeuristicCombinedDivisorsEnding9 = CopySpanToDevice(accelerator, stream, combinedEnding9);
|
||||
var combinedSquaresEnding9 = HeuristicCombinedPrimeTester.CombinedDivisorsEnding9Squares;
|
||||
HeuristicCombinedDivisorSquaresEnding9 = CopySpanToDevice(accelerator, stream, combinedSquaresEnding9);
|
||||
}
|
||||
}
|
||||
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisors;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisorSquares;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding1;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding1;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding7;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding7;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding9;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding9;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding1;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding1;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding3;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding3;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding7;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding7;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding9;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding9;
|
||||
|
||||
internal HeuristicGpuDivisorTables CreateHeuristicDivisorTables() => new(
|
||||
HeuristicCombinedDivisorsEnding1.View,
|
||||
HeuristicCombinedDivisorsEnding3.View,
|
||||
HeuristicCombinedDivisorsEnding7.View,
|
||||
HeuristicCombinedDivisorsEnding9.View,
|
||||
HeuristicCombinedDivisorSquaresEnding1.View,
|
||||
HeuristicCombinedDivisorSquaresEnding3.View,
|
||||
HeuristicCombinedDivisorSquaresEnding7.View,
|
||||
HeuristicCombinedDivisorSquaresEnding9.View,
|
||||
HeuristicGroupADivisors.View,
|
||||
HeuristicGroupADivisorSquares.View,
|
||||
HeuristicGroupBDivisorsEnding1.View,
|
||||
HeuristicGroupBDivisorSquaresEnding1.View,
|
||||
HeuristicGroupBDivisorsEnding7.View,
|
||||
HeuristicGroupBDivisorSquaresEnding7.View,
|
||||
HeuristicGroupBDivisorsEnding9.View,
|
||||
HeuristicGroupBDivisorSquaresEnding9.View);
|
||||
|
||||
private static MemoryBuffer1D<ulong, Stride1D.Dense> CopySpanToDevice(Accelerator accelerator, AcceleratorStream stream, in ulong[] span)
|
||||
{
|
||||
MemoryBuffer1D<ulong, Stride1D.Dense>? buffer;
|
||||
// lock (accelerator)
|
||||
{
|
||||
buffer = accelerator.Allocate1D(stream, span);
|
||||
}
|
||||
|
||||
return buffer;
|
||||
}
|
||||
|
||||
private static MemoryBuffer1D<ulong, Stride1D.Dense> CopyUintSpanToDevice(Accelerator accelerator, AcceleratorStream stream, ReadOnlySpan<uint> span)
|
||||
{
|
||||
MemoryBuffer1D<ulong, Stride1D.Dense>? buffer;
|
||||
// lock (accelerator)
|
||||
{
|
||||
buffer = accelerator.Allocate1D<ulong>(span.Length);
|
||||
}
|
||||
|
||||
if (!span.IsEmpty)
|
||||
{
|
||||
var converted = new ulong[span.Length];
|
||||
for (int i = 0; i < span.Length; i++)
|
||||
{
|
||||
converted[i] = span[i];
|
||||
}
|
||||
|
||||
buffer.View.CopyFromCPU(stream, converted);
|
||||
}
|
||||
|
||||
return buffer;
|
||||
}
|
||||
|
||||
internal void Dispose()
|
||||
{
|
||||
HeuristicGroupADivisors.Dispose();
|
||||
HeuristicGroupADivisorSquares.Dispose();
|
||||
HeuristicGroupBDivisorsEnding1.Dispose();
|
||||
HeuristicGroupBDivisorSquaresEnding1.Dispose();
|
||||
HeuristicGroupBDivisorsEnding7.Dispose();
|
||||
HeuristicGroupBDivisorSquaresEnding7.Dispose();
|
||||
HeuristicGroupBDivisorsEnding9.Dispose();
|
||||
HeuristicGroupBDivisorSquaresEnding9.Dispose();
|
||||
HeuristicCombinedDivisorsEnding1.Dispose();
|
||||
HeuristicCombinedDivisorSquaresEnding1.Dispose();
|
||||
HeuristicCombinedDivisorsEnding3.Dispose();
|
||||
HeuristicCombinedDivisorSquaresEnding3.Dispose();
|
||||
HeuristicCombinedDivisorsEnding7.Dispose();
|
||||
HeuristicCombinedDivisorSquaresEnding7.Dispose();
|
||||
HeuristicCombinedDivisorsEnding9.Dispose();
|
||||
HeuristicCombinedDivisorSquaresEnding9.Dispose();
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,208 @@
|
||||
using System.Runtime.CompilerServices;
|
||||
using ILGPU;
|
||||
using ILGPU.Runtime;
|
||||
|
||||
namespace PerfectNumbers.Core.Gpu.Accelerators;
|
||||
|
||||
internal sealed class HeuristicPrimeTesterAccelerator
|
||||
{
|
||||
#region Pool
|
||||
|
||||
[ThreadStatic]
|
||||
private static Queue<HeuristicPrimeTesterAccelerator>? Pool;
|
||||
|
||||
internal static void WarmUp()
|
||||
{
|
||||
Accelerator[]? accelerators = AcceleratorPool.Shared.Accelerators;
|
||||
int acceleratorCount = accelerators.Length;
|
||||
Span<AcceleratorStream> streams = new(new AcceleratorStream[acceleratorCount]);
|
||||
for (var i = 0; i < acceleratorCount; i++)
|
||||
{
|
||||
Console.WriteLine($"Preparing heuristic accelerator {i}...");
|
||||
var accelerator = accelerators[i];
|
||||
AcceleratorStream stream = accelerator.CreateStream();
|
||||
Accelerators.HeuristicGpuTables.EnsureStaticTables(accelerator, stream);
|
||||
// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
|
||||
// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
|
||||
// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
|
||||
streams[i] = stream;
|
||||
}
|
||||
|
||||
for (var i = 0; i < acceleratorCount; i++)
|
||||
{
|
||||
AcceleratorStream stream = streams[i];
|
||||
stream.Synchronize();
|
||||
stream.Dispose();
|
||||
Console.WriteLine($"Heuristic accelerator {i} is ready");
|
||||
}
|
||||
}
|
||||
|
||||
internal static HeuristicPrimeTesterAccelerator Rent(int minBufferCapacity = 1)
|
||||
{
|
||||
var pool = Pool ??= new();
|
||||
if (!pool.TryDequeue(out var lease))
|
||||
{
|
||||
lease = new HeuristicPrimeTesterAccelerator(minBufferCapacity);
|
||||
}
|
||||
else
|
||||
{
|
||||
lease.EnsureCapacity(minBufferCapacity);
|
||||
}
|
||||
|
||||
return lease;
|
||||
}
|
||||
|
||||
internal static void Return(HeuristicPrimeTesterAccelerator lease)
|
||||
{
|
||||
var pool = Pool ??= new();
|
||||
pool.Enqueue(lease);
|
||||
}
|
||||
|
||||
internal static void DisposeAll()
|
||||
{
|
||||
if (Pool != null)
|
||||
{
|
||||
while (Pool.TryDequeue(out var lease))
|
||||
{
|
||||
lease.DisposeResources();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
internal static void Clear()
|
||||
{
|
||||
if (Pool != null)
|
||||
{
|
||||
var retained = new List<HeuristicPrimeTesterAccelerator>();
|
||||
while (Pool.TryDequeue(out var lease))
|
||||
{
|
||||
lease.DisposeResources();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#endregion
|
||||
|
||||
private readonly HeuristicGpuDivisorTables _heuristicDivisorTables;
|
||||
internal readonly Context Context;
|
||||
public readonly Accelerator Accelerator;
|
||||
public AcceleratorStream? Stream;
|
||||
|
||||
[ThreadStatic]
|
||||
public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>? SmallPrimeSieveKernel;
|
||||
|
||||
internal static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetSmallPrimeSieveKernel(Accelerator accelerator)
|
||||
{
|
||||
var pool = SmallPrimeSieveKernel ??= [];
|
||||
if (pool.TryGetValue(accelerator, out var cached))
|
||||
{
|
||||
return cached;
|
||||
}
|
||||
|
||||
cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
|
||||
pool[accelerator] = cached;
|
||||
return cached;
|
||||
}
|
||||
|
||||
[ThreadStatic]
|
||||
public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>>? HeuristicTrialDivisionKernel;
|
||||
|
||||
internal static Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables> GetHeuristicTrialDivisionKernel(Accelerator accelerator)
|
||||
{
|
||||
var pool = HeuristicTrialDivisionKernel ??= [];
|
||||
if (pool.TryGetValue(accelerator, out var cached))
|
||||
{
|
||||
return cached;
|
||||
}
|
||||
|
||||
cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>(PrimeTesterKernels.HeuristicTrialDivisionKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>>();
|
||||
pool[accelerator] = cached;
|
||||
return cached;
|
||||
}
|
||||
|
||||
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> Input = null!;
|
||||
public MemoryBuffer1D<byte, Stride1D.Dense> Output = null!;
|
||||
public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag = null!;
|
||||
public int BufferCapacity;
|
||||
|
||||
private HeuristicGpuTables _sharedTables;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisors => _sharedTables.HeuristicGroupADivisors;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupADivisorSquares => _sharedTables.HeuristicGroupADivisorSquares;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding1 => _sharedTables.HeuristicGroupBDivisorsEnding1;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding1 => _sharedTables.HeuristicGroupBDivisorSquaresEnding1;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding7 => _sharedTables.HeuristicGroupBDivisorsEnding7;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding7 => _sharedTables.HeuristicGroupBDivisorSquaresEnding7;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorsEnding9 => _sharedTables.HeuristicGroupBDivisorsEnding9;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicGroupBDivisorSquaresEnding9 => _sharedTables.HeuristicGroupBDivisorSquaresEnding9;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding1 => _sharedTables.HeuristicCombinedDivisorsEnding1;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding1 => _sharedTables.HeuristicCombinedDivisorSquaresEnding1;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding3 => _sharedTables.HeuristicCombinedDivisorsEnding3;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding3 => _sharedTables.HeuristicCombinedDivisorSquaresEnding3;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding7 => _sharedTables.HeuristicCombinedDivisorsEnding7;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding7 => _sharedTables.HeuristicCombinedDivisorSquaresEnding7;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorsEnding9 => _sharedTables.HeuristicCombinedDivisorsEnding9;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> HeuristicCombinedDivisorSquaresEnding9 => _sharedTables.HeuristicCombinedDivisorSquaresEnding9;
|
||||
|
||||
public HeuristicGpuDivisorTables HeuristicGpuTables => _heuristicDivisorTables;
|
||||
|
||||
internal HeuristicPrimeTesterAccelerator(int minBufferCapacity)
|
||||
{
|
||||
Accelerator accelerator = AcceleratorPool.Shared.Rent();
|
||||
Context = accelerator.Context;
|
||||
|
||||
Accelerator = accelerator;
|
||||
AcceleratorStream stream = accelerator.CreateStream();
|
||||
// var kernels = _kernels[index]; //GpuKernelPool.Kernels;
|
||||
// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
|
||||
|
||||
_sharedTables = Accelerators.HeuristicGpuTables.EnsureStaticTables(accelerator, stream);
|
||||
var heuristicDivisorTables = _sharedTables.CreateHeuristicDivisorTables();
|
||||
_heuristicDivisorTables = heuristicDivisorTables;
|
||||
|
||||
EnsureCapacity(minBufferCapacity);
|
||||
stream.Synchronize();
|
||||
stream.Dispose();
|
||||
}
|
||||
|
||||
public void EnsureCapacity(int minCapacity)
|
||||
{
|
||||
if (minCapacity <= BufferCapacity)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
Input?.Dispose();
|
||||
Output?.Dispose();
|
||||
HeuristicFlag?.Dispose();
|
||||
|
||||
// lock (Accelerator)
|
||||
{
|
||||
Input = Accelerator.Allocate1D<ulong>(minCapacity);
|
||||
Output = Accelerator.Allocate1D<byte>(minCapacity);
|
||||
HeuristicFlag = Accelerator.Allocate1D<int>(minCapacity);
|
||||
}
|
||||
|
||||
BufferCapacity = minCapacity;
|
||||
}
|
||||
|
||||
public void Dispose()
|
||||
{
|
||||
Stream?.Dispose();
|
||||
Stream = null;
|
||||
Return(this);
|
||||
}
|
||||
|
||||
internal void DisposeResources()
|
||||
{
|
||||
Input?.Dispose();
|
||||
Output?.Dispose();
|
||||
HeuristicFlag?.Dispose();
|
||||
Stream?.Dispose();
|
||||
Stream = null;
|
||||
// ReleaseSharedTables(_sharedTables);
|
||||
// These resources are shared between GPU leases
|
||||
// Accelerator.Dispose();
|
||||
// AcceleratorContext.Dispose();
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,58 @@
|
||||
using System.Collections.Concurrent;
|
||||
using ILGPU;
|
||||
using ILGPU.Runtime;
|
||||
|
||||
namespace PerfectNumbers.Core.Gpu.Accelerators;
|
||||
|
||||
internal sealed class LastDigitGpuTables
|
||||
{
|
||||
private static readonly ConcurrentDictionary<Accelerator, LastDigitGpuTables> _sharedTables = new(20_480, PerfectNumberConstants.RollingAccelerators);
|
||||
|
||||
internal static LastDigitGpuTables EnsureStaticTables(Accelerator accelerator, AcceleratorStream stream)
|
||||
=> _sharedTables.GetOrAdd(accelerator, _ => new LastDigitGpuTables(accelerator, stream));
|
||||
|
||||
internal LastDigitGpuTables(Accelerator accelerator, AcceleratorStream stream)
|
||||
{
|
||||
var primesLastOne = DivisorGenerator.SmallPrimesLastOne;
|
||||
var primesLastSeven = DivisorGenerator.SmallPrimesLastSeven;
|
||||
var primesLastThree = DivisorGenerator.SmallPrimesLastThree;
|
||||
var primesLastNine = DivisorGenerator.SmallPrimesLastNine;
|
||||
var primesPow2LastOne = DivisorGenerator.SmallPrimesPow2LastOne;
|
||||
var primesPow2LastSeven = DivisorGenerator.SmallPrimesPow2LastSeven;
|
||||
var primesPow2LastThree = DivisorGenerator.SmallPrimesPow2LastThree;
|
||||
var primesPow2LastNine = DivisorGenerator.SmallPrimesPow2LastNine;
|
||||
|
||||
// lock (accelerator)
|
||||
{
|
||||
DevicePrimesLastOne = accelerator.Allocate1D(stream, primesLastOne);
|
||||
DevicePrimesLastSeven = accelerator.Allocate1D(stream, primesLastSeven);
|
||||
DevicePrimesLastThree = accelerator.Allocate1D(stream, primesLastThree);
|
||||
DevicePrimesLastNine = accelerator.Allocate1D(stream, primesLastNine);
|
||||
DevicePrimesPow2LastOne = accelerator.Allocate1D(stream, primesPow2LastOne);
|
||||
DevicePrimesPow2LastSeven = accelerator.Allocate1D(stream, primesPow2LastSeven);
|
||||
DevicePrimesPow2LastThree = accelerator.Allocate1D(stream, primesPow2LastThree);
|
||||
DevicePrimesPow2LastNine = accelerator.Allocate1D(stream, primesPow2LastNine);
|
||||
}
|
||||
}
|
||||
|
||||
internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
|
||||
internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
|
||||
internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree;
|
||||
internal readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
|
||||
internal readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
|
||||
|
||||
internal void Dispose()
|
||||
{
|
||||
DevicePrimesLastOne.Dispose();
|
||||
DevicePrimesLastSeven.Dispose();
|
||||
DevicePrimesLastThree.Dispose();
|
||||
DevicePrimesLastNine.Dispose();
|
||||
DevicePrimesPow2LastOne.Dispose();
|
||||
DevicePrimesPow2LastSeven.Dispose();
|
||||
DevicePrimesPow2LastThree.Dispose();
|
||||
DevicePrimesPow2LastNine.Dispose();
|
||||
}
|
||||
}
|
||||
@@ -44,8 +44,8 @@ public sealed class Pow2MontgomeryAccelerator
|
||||
public readonly Accelerator Accelerator;
|
||||
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Input;
|
||||
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Output;
|
||||
public MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> ToTestOnDevice;
|
||||
public Dictionary<ulong, int> ToTestOnHost = [];
|
||||
public MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> Pow2ModEntriesToTestOnDevice;
|
||||
public Dictionary<ulong, int> Pow2ModEntriesToTestOnHost = [];
|
||||
public AcceleratorStream? Stream;
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernel;
|
||||
@@ -55,11 +55,11 @@ public sealed class Pow2MontgomeryAccelerator
|
||||
|
||||
public void EnsureCapacity(int factorsCount)
|
||||
{
|
||||
if (factorsCount > ToTestOnDevice.Length)
|
||||
if (factorsCount > Pow2ModEntriesToTestOnDevice.Length)
|
||||
{
|
||||
Accelerator accelerator = Accelerator;
|
||||
ToTestOnDevice.Dispose();
|
||||
ToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
|
||||
Pow2ModEntriesToTestOnDevice.Dispose();
|
||||
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -70,7 +70,7 @@ public sealed class Pow2MontgomeryAccelerator
|
||||
|
||||
Input = accelerator.Allocate1D<ulong>(1);
|
||||
Output = accelerator.Allocate1D<ulong>(1);
|
||||
ToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
|
||||
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
|
||||
|
||||
var keepKernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery);
|
||||
var keepLauncher = KernelUtil.GetKernel(keepKernel);
|
||||
@@ -91,7 +91,7 @@ public sealed class Pow2MontgomeryAccelerator
|
||||
Input.Dispose();
|
||||
Output.Dispose();
|
||||
Stream?.Dispose();
|
||||
ToTestOnDevice.Dispose();
|
||||
Pow2ModEntriesToTestOnDevice.Dispose();
|
||||
// Accelerator is shared with other threads
|
||||
// Accelerator.Dispose();
|
||||
}
|
||||
|
||||
@@ -0,0 +1,192 @@
|
||||
using System.Runtime.CompilerServices;
|
||||
using ILGPU;
|
||||
using ILGPU.Runtime;
|
||||
|
||||
namespace PerfectNumbers.Core.Gpu.Accelerators;
|
||||
|
||||
internal sealed class PrimeTesterByLastDigitAccelerator
|
||||
{
|
||||
#region Pool
|
||||
|
||||
[ThreadStatic]
|
||||
private static Queue<PrimeTesterByLastDigitAccelerator>? Pool;
|
||||
|
||||
// internal static void EnsureStaticTables(Accelerator accelerator)
|
||||
// {
|
||||
// var tables = CreateSharedTables(accelerator);
|
||||
// }
|
||||
|
||||
internal static void WarmUp()
|
||||
{
|
||||
Accelerator[]? accelerators = AcceleratorPool.Shared.Accelerators;
|
||||
int acceleratorCount = accelerators.Length;
|
||||
Span<AcceleratorStream> streams = new(new AcceleratorStream[acceleratorCount]);
|
||||
for (var i = 0; i < acceleratorCount; i++)
|
||||
{
|
||||
Console.WriteLine($"Preparing accelerator {i}...");
|
||||
var accelerator = accelerators[i];
|
||||
AcceleratorStream stream = accelerator.CreateStream();
|
||||
LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
|
||||
// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
|
||||
// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
|
||||
// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
|
||||
// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
|
||||
streams[i] = stream;
|
||||
}
|
||||
|
||||
for (var i = 0; i < acceleratorCount; i++)
|
||||
{
|
||||
AcceleratorStream stream = streams[i];
|
||||
stream.Synchronize();
|
||||
stream.Dispose();
|
||||
Console.WriteLine($"Accelerator {i} is ready");
|
||||
}
|
||||
}
|
||||
|
||||
internal static PrimeTesterByLastDigitAccelerator Rent(int minBufferCapacity = 1)
|
||||
{
|
||||
var pool = Pool ??= new();
|
||||
if (!pool.TryDequeue(out var lease))
|
||||
{
|
||||
lease = new PrimeTesterByLastDigitAccelerator(minBufferCapacity);
|
||||
}
|
||||
else
|
||||
{
|
||||
lease.EnsureCapacity(minBufferCapacity);
|
||||
}
|
||||
|
||||
return lease;
|
||||
}
|
||||
|
||||
internal static void Return(PrimeTesterByLastDigitAccelerator lease)
|
||||
{
|
||||
var pool = Pool ??= new();
|
||||
pool.Enqueue(lease);
|
||||
}
|
||||
|
||||
internal static void DisposeAll()
|
||||
{
|
||||
if (Pool != null)
|
||||
{
|
||||
while (Pool.TryDequeue(out var lease))
|
||||
{
|
||||
lease.DisposeResources();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
internal static void Clear()
|
||||
{
|
||||
if (Pool != null)
|
||||
{
|
||||
var retained = new List<PrimeTesterByLastDigitAccelerator>();
|
||||
while (Pool.TryDequeue(out var lease))
|
||||
{
|
||||
lease.DisposeResources();
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#endregion
|
||||
|
||||
private readonly LastDigitGpuTables _sharedTables;
|
||||
internal readonly Context Context;
|
||||
public readonly Accelerator Accelerator;
|
||||
private AcceleratorStream? _stream;
|
||||
public AcceleratorStream Stream
|
||||
{
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
get => _stream ?? Accelerator.CreateStream();
|
||||
}
|
||||
|
||||
[ThreadStatic]
|
||||
public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>? SmallPrimeSieveKernel;
|
||||
|
||||
internal static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetSmallPrimeSieveKernel(Accelerator accelerator)
|
||||
{
|
||||
var pool = SmallPrimeSieveKernel ??= [];
|
||||
if (pool.TryGetValue(accelerator, out var cached))
|
||||
{
|
||||
return cached;
|
||||
}
|
||||
|
||||
cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
|
||||
pool[accelerator] = cached;
|
||||
return cached;
|
||||
}
|
||||
|
||||
[ThreadStatic]
|
||||
public static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<int>, ulong, ulong, HeuristicGpuDivisorTableKind, HeuristicGpuDivisorTables>>? HeuristicTrialDivisionKernel;
|
||||
|
||||
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> Input = null!;
|
||||
public MemoryBuffer1D<byte, Stride1D.Dense> Output = null!;
|
||||
public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag = null!;
|
||||
public int BufferCapacity;
|
||||
|
||||
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne => _sharedTables.DevicePrimesLastOne;
|
||||
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven => _sharedTables.DevicePrimesLastSeven;
|
||||
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree => _sharedTables.DevicePrimesLastThree;
|
||||
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine => _sharedTables.DevicePrimesLastNine;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne => _sharedTables.DevicePrimesPow2LastOne;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven => _sharedTables.DevicePrimesPow2LastSeven;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree => _sharedTables.DevicePrimesPow2LastThree;
|
||||
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine => _sharedTables.DevicePrimesPow2LastNine;
|
||||
|
||||
internal PrimeTesterByLastDigitAccelerator(int minBufferCapacity)
|
||||
{
|
||||
Accelerator accelerator = AcceleratorPool.Shared.Rent();
|
||||
Context = accelerator.Context;
|
||||
|
||||
Accelerator = accelerator;
|
||||
AcceleratorStream stream = accelerator.CreateStream();
|
||||
|
||||
var sharedTables = LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
|
||||
_sharedTables = sharedTables;
|
||||
|
||||
EnsureCapacity(minBufferCapacity);
|
||||
stream.Synchronize();
|
||||
stream.Dispose();
|
||||
}
|
||||
|
||||
public void EnsureCapacity(int minCapacity)
|
||||
{
|
||||
if (minCapacity <= BufferCapacity)
|
||||
{
|
||||
return;
|
||||
}
|
||||
|
||||
Input?.Dispose();
|
||||
Output?.Dispose();
|
||||
HeuristicFlag?.Dispose();
|
||||
|
||||
// lock (Accelerator)
|
||||
{
|
||||
Input = Accelerator.Allocate1D<ulong>(minCapacity);
|
||||
Output = Accelerator.Allocate1D<byte>(minCapacity);
|
||||
HeuristicFlag = Accelerator.Allocate1D<int>(minCapacity);
|
||||
}
|
||||
|
||||
BufferCapacity = minCapacity;
|
||||
}
|
||||
|
||||
public void Dispose()
|
||||
{
|
||||
_stream?.Dispose();
|
||||
_stream = null;
|
||||
Return(this);
|
||||
}
|
||||
|
||||
internal void DisposeResources()
|
||||
{
|
||||
Input?.Dispose();
|
||||
Output?.Dispose();
|
||||
HeuristicFlag?.Dispose();
|
||||
_stream?.Dispose();
|
||||
_stream = null;
|
||||
// ReleaseSharedTables(_sharedTables);
|
||||
// These resources are shared between GPU leases
|
||||
// Accelerator.Dispose();
|
||||
// AcceleratorContext.Dispose();
|
||||
}
|
||||
}
|
||||
@@ -19,91 +19,156 @@ public class GpuKernelPool
|
||||
return Interlocked.CompareExchange(ref target, newValue, null) ?? newValue;
|
||||
}
|
||||
|
||||
internal static KernelContainer GetOrAddKernels(Accelerator accelerator, AcceleratorStream stream)
|
||||
internal static KernelContainer GetOrAddKernels(Accelerator accelerator, AcceleratorStream stream, KernelType kernelType)
|
||||
{
|
||||
var pool = _pool ??= [];
|
||||
if (pool.TryGetValue(accelerator, out var kernelContainer))
|
||||
if (!pool.TryGetValue(accelerator, out var kernels))
|
||||
{
|
||||
return kernelContainer;
|
||||
kernels = new KernelContainer();
|
||||
}
|
||||
|
||||
var kernels = new KernelContainer();
|
||||
|
||||
InitOnce(ref kernels.Order, () =>
|
||||
if (kernelType.HasFlag(KernelType.OrderKernelScan))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
|
||||
InitOrderKernelScan(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>>();
|
||||
});
|
||||
|
||||
InitOnce(ref kernels.Pow2Mod, () =>
|
||||
if (kernelType.HasFlag(KernelType.Pow2ModKernelScan))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModKernelScan);
|
||||
InitPow2ModKernelScan(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
|
||||
InitOnce(ref kernels.Incremental, () =>
|
||||
if (kernelType.HasFlag(KernelType.IncrementalKernelScan))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
|
||||
InitIncrementalKernelScan(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
|
||||
InitOnce(ref kernels.IncrementalOrder, () =>
|
||||
if (kernelType.HasFlag(KernelType.IncrementalOrderKernelScan))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
|
||||
IncrementalOrderKernelScan(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte,
|
||||
ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
|
||||
InitOnce(ref kernels.Pow2ModOrder, () =>
|
||||
if (kernelType.HasFlag(KernelType.Pow2ModOrderKernelScan))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
|
||||
InitPow2ModOrderKernelScan(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
|
||||
InitOnce(ref kernels.SpecialMax, () =>
|
||||
if (kernelType.HasFlag(KernelType.EvaluateSpecialMaxCandidatesKernel))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel);
|
||||
InitEvaluateSpecialMaxCandidatesKernel(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
|
||||
InitOnce(ref kernels.SmallPrimeFactor, () =>
|
||||
if (kernelType.HasFlag(KernelType.SmallPrimeFactorKernelScan))
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan);
|
||||
InitSmallPrimeFactorKernelScan(accelerator, kernels);
|
||||
}
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
|
||||
PreloadStaticTables(accelerator, kernels, stream);
|
||||
PreloadStaticTables(accelerator, kernels, stream, kernelType);
|
||||
|
||||
pool[accelerator] = kernels;
|
||||
return kernels;
|
||||
|
||||
static void InitOrderKernelScan(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.Order, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>>();
|
||||
});
|
||||
}
|
||||
|
||||
static void InitPow2ModKernelScan(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.Pow2Mod, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModKernelScan);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
}
|
||||
|
||||
static void InitIncrementalKernelScan(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.Incremental, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
}
|
||||
|
||||
static void IncrementalOrderKernelScan(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.IncrementalOrder, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte,
|
||||
ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
}
|
||||
|
||||
static void InitPow2ModOrderKernelScan(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.Pow2ModOrder, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
}
|
||||
|
||||
static void InitEvaluateSpecialMaxCandidatesKernel(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.SpecialMax, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
}
|
||||
|
||||
static void InitSmallPrimeFactorKernelScan(Accelerator accelerator, KernelContainer kernels)
|
||||
{
|
||||
InitOnce(ref kernels.SmallPrimeFactor, () =>
|
||||
{
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
|
||||
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
internal static void PreloadStaticTables(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream)
|
||||
internal static void PreloadStaticTables(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream, KernelType kernelType)
|
||||
{
|
||||
EnsureSmallCyclesOnDevice(accelerator, kernels, stream);
|
||||
EnsureSmallPrimesOnDevice(accelerator, kernels, stream);
|
||||
EnsureSmallPrimeFactorTables(accelerator, kernels, stream);
|
||||
if (kernelType.HasFlag(KernelType.SmallCycles))
|
||||
{
|
||||
EnsureSmallCyclesOnDevice(accelerator, kernels, stream);
|
||||
}
|
||||
|
||||
if (kernelType.HasFlag(KernelType.SmallPrimes))
|
||||
{
|
||||
EnsureSmallPrimesOnDevice(accelerator, kernels, stream);
|
||||
}
|
||||
|
||||
if (kernelType.HasFlag(KernelType.SmallPrimeFactorKernelScan))
|
||||
{
|
||||
EnsureSmallPrimeFactorTables(accelerator, kernels, stream);
|
||||
}
|
||||
}
|
||||
|
||||
public static ArrayView1D<ulong, Stride1D.Dense> GetSmallCyclesOnDevice(KernelContainer kernels) => kernels.SmallCycles!;
|
||||
@@ -166,12 +231,12 @@ public class GpuKernelPool
|
||||
|
||||
// lock (accelerator)
|
||||
// {
|
||||
deviceLastOne = accelerator.Allocate1D<uint>(hostLastOne.Length);
|
||||
deviceLastSeven = accelerator.Allocate1D<uint>(hostLastSeven.Length);
|
||||
deviceLastOnePow2 = accelerator.Allocate1D<ulong>(hostLastOnePow2.Length);
|
||||
deviceLastSevenPow2 = accelerator.Allocate1D<ulong>(hostLastSevenPow2.Length);
|
||||
deviceLastOne = accelerator.Allocate1D<uint>(hostLastOne.Length);
|
||||
deviceLastSeven = accelerator.Allocate1D<uint>(hostLastSeven.Length);
|
||||
deviceLastOnePow2 = accelerator.Allocate1D<ulong>(hostLastOnePow2.Length);
|
||||
deviceLastSevenPow2 = accelerator.Allocate1D<ulong>(hostLastSevenPow2.Length);
|
||||
// }
|
||||
|
||||
|
||||
deviceLastOne.View.CopyFromCPU(stream, hostLastOne);
|
||||
deviceLastSeven.View.CopyFromCPU(stream, hostLastSeven);
|
||||
deviceLastOnePow2.View.CopyFromCPU(stream, hostLastOnePow2);
|
||||
@@ -186,7 +251,7 @@ public class GpuKernelPool
|
||||
kernels.SmallPrimeFactorsPrimes!,
|
||||
kernels.SmallPrimeFactorsSquares!
|
||||
);
|
||||
|
||||
|
||||
private static void EnsureSmallPrimeFactorTables(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream)
|
||||
{
|
||||
if (kernels.SmallPrimeFactorsPrimes is { } && kernels.SmallPrimeFactorsSquares is { })
|
||||
@@ -206,7 +271,7 @@ public class GpuKernelPool
|
||||
devicePrimes = accelerator.Allocate1D<uint>(length);
|
||||
deviceSquares = accelerator.Allocate1D<ulong>(length);
|
||||
}
|
||||
|
||||
|
||||
devicePrimes.View.CopyFromCPU(stream, hostPrimes);
|
||||
deviceSquares.View.CopyFromCPU(stream, hostSquares);
|
||||
|
||||
|
||||
@@ -8,7 +8,7 @@ internal static class GpuStaticTableInitializer
|
||||
internal static void EnsureStaticTables(Accelerator accelerator, KernelContainer kernels, AcceleratorStream stream)
|
||||
{
|
||||
// PrimeTester.PrimeTesterGpuContextPool.EnsureStaticTables(accelerator);
|
||||
GpuKernelPool.PreloadStaticTables(accelerator, kernels, stream);
|
||||
GpuKernelPool.PreloadStaticTables(accelerator, kernels, stream, KernelType.None);
|
||||
// PrimeOrderGpuHeuristics.PreloadStaticTables(accelerator, stream);
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,16 @@
|
||||
namespace PerfectNumbers.Core.Gpu;
|
||||
|
||||
[Flags]
|
||||
public enum KernelType
|
||||
{
|
||||
None,
|
||||
SmallCycles,
|
||||
SmallPrimes,
|
||||
Pow2ModKernelScan,
|
||||
OrderKernelScan,
|
||||
IncrementalKernelScan,
|
||||
IncrementalOrderKernelScan,
|
||||
Pow2ModOrderKernelScan,
|
||||
EvaluateSpecialMaxCandidatesKernel,
|
||||
SmallPrimeFactorKernelScan
|
||||
}
|
||||
@@ -20,12 +20,10 @@ internal static class PrimeTesterKernels
|
||||
public static void SmallPrimeSieveKernel(
|
||||
Index1D index,
|
||||
ArrayView<ulong> numbers,
|
||||
ArrayView<uint> smallPrimes,
|
||||
ArrayView<uint> smallPrimesLastOne,
|
||||
ArrayView<uint> smallPrimesLastSeven,
|
||||
ArrayView<uint> smallPrimesLastThree,
|
||||
ArrayView<uint> smallPrimesLastNine,
|
||||
ArrayView<ulong> smallPrimesPow2,
|
||||
ArrayView<ulong> smallPrimesPow2LastOne,
|
||||
ArrayView<ulong> smallPrimesPow2LastSeven,
|
||||
ArrayView<ulong> smallPrimesPow2LastThree,
|
||||
@@ -35,8 +33,8 @@ internal static class PrimeTesterKernels
|
||||
ulong n = numbers[index];
|
||||
byte result = 1;
|
||||
|
||||
ArrayView<uint> primes = smallPrimes;
|
||||
ArrayView<ulong> primeSquares = smallPrimesPow2;
|
||||
ArrayView<uint> primes;
|
||||
ArrayView<ulong> primeSquares;
|
||||
ulong lastDigit = n % 10UL;
|
||||
switch (lastDigit)
|
||||
{
|
||||
@@ -55,7 +53,9 @@ internal static class PrimeTesterKernels
|
||||
case 9UL:
|
||||
primes = smallPrimesLastNine;
|
||||
primeSquares = smallPrimesPow2LastNine;
|
||||
break;
|
||||
break;
|
||||
default:
|
||||
return;
|
||||
}
|
||||
|
||||
int length = (int)primes.Length;
|
||||
|
||||
@@ -24,7 +24,7 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
|
||||
ulong divMul = (ulong)((((UInt128)1 << 64) - UInt128.One) / exponent) + 1UL;
|
||||
byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0; // ILGPU kernels do not support bool parameters
|
||||
|
||||
KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
|
||||
KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.Pow2ModKernelScan | KernelType.IncrementalKernelScan);
|
||||
var pow2Kernel = kernels.Pow2Mod!;
|
||||
var incKernel = kernels.Incremental!;
|
||||
ulong step10 = (exponent.Mod10() << 1).Mod10();
|
||||
|
||||
@@ -18,7 +18,7 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
|
||||
ulong divMul = (ulong)((((UInt128)1 << 64) - UInt128.One) / exponent) + 1UL;
|
||||
byte last = lastDigit == LastDigit.Seven ? (byte)1 : (byte)0; // ILGPU kernels do not support bool parameters
|
||||
|
||||
KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
|
||||
KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.Pow2ModOrderKernelScan | KernelType.IncrementalOrderKernelScan);
|
||||
var kernel = _kernelType switch
|
||||
{
|
||||
GpuKernelType.Pow2Mod => kernels.Pow2ModOrder!,
|
||||
|
||||
@@ -33,7 +33,7 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
|
||||
var stream = accelerator.CreateStream();
|
||||
var kernel = GetPow2ModKernel(accelerator);
|
||||
var resources = RentResources(accelerator, GpuConstants.ScanBatchSize);
|
||||
var gpuKernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
|
||||
var gpuKernels = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.SmallCycles | KernelType.SmallPrimes);
|
||||
|
||||
var orderBuffer = resources.OrderBuffer;
|
||||
ulong[] orderArray = resources.OrderArray;
|
||||
|
||||
Reference in new issue
Block a user