IsPrime with cancellation token removed.

This commit is contained in:
mlemanczyk committed 2025-10-29 22:31:28 +01:00
1 parent f9504a6a19
commit fc5700f82c
11 files changed
+287 -314

No files matched your search

+251 -265
View File
@@ -1,7 +1,5 @@
using System;
using System.Buffers;
using System.Collections.Concurrent;
using System.Collections.Generic;
using System.Numerics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
@@ -22,7 +20,7 @@ public sealed class PrimeTester
public static PrimeTester Exclusive => _tester ??= new();
public static bool IsPrime(ulong n, CancellationToken ct)
public static bool IsPrime(ulong n)
{
if (n <= 1UL)
{
@@ -92,310 +90,298 @@ public sealed class PrimeTester
return result;
}
public static bool IsPrimeGpu(ulong n)
{
return Exclusive.IsPrimeGpu(n, CancellationToken.None);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public bool IsPrimeGpu(ulong n, CancellationToken ct)
{
bool forceCpu = GpuContextPool.ForceCpu;
bool belowGpuRange = n < 31UL;
if (forceCpu || belowGpuRange)
{
return IsPrimeCpu(n, ct);
}
var limiter = GpuPrimeWorkLimiter.Acquire();
var gpu = PrimeTesterGpuContextPool.Rent(1);
var accelerator = gpu.Accelerator;
var kernel = gpu.Kernel;
ulong value = n;
byte flag = 0;
var input = gpu.Input;
var output = gpu.Output;
var inputView = input.View;
var outputView = output.View;
inputView.CopyFromCPU(ref value, 1);
kernel(
1,
inputView,
gpu.DevicePrimesDefault.View,
gpu.DevicePrimesLastOne.View,
gpu.DevicePrimesLastSeven.View,
gpu.DevicePrimesLastThree.View,
gpu.DevicePrimesLastNine.View,
gpu.DevicePrimesPow2Default.View,
gpu.DevicePrimesPow2LastOne.View,
gpu.DevicePrimesPow2LastSeven.View,
gpu.DevicePrimesPow2LastThree.View,
gpu.DevicePrimesPow2LastNine.View,
outputView);
accelerator.Synchronize();
outputView.CopyToCPU(ref flag, 1);
gpu.Dispose();
limiter.Dispose();
return flag != 0;
}
public static bool IsPrimeCpu(ulong n, CancellationToken ct)
public bool IsPrimeGpu(ulong n)
{
// Preserve the legacy entry point for callers that bypass the thread-local caches.
return IsPrime(n, ct);
}
bool forceCpu = GpuContextPool.ForceCpu;
bool belowGpuRange = n < 31UL;
if (forceCpu || belowGpuRange)
{
return IsPrime(n);
}
var limiter = GpuPrimeWorkLimiter.Acquire();
var gpu = PrimeTesterGpuContextPool.Rent(1);
var accelerator = gpu.Accelerator;
var kernel = gpu.Kernel;
ulong value = n;
byte flag = 0;
var input = gpu.Input;
var output = gpu.Output;
var inputView = input.View;
var outputView = output.View;
inputView.CopyFromCPU(ref value, 1);
kernel(
1,
inputView,
gpu.DevicePrimesDefault.View,
gpu.DevicePrimesLastOne.View,
gpu.DevicePrimesLastSeven.View,
gpu.DevicePrimesLastThree.View,
gpu.DevicePrimesLastNine.View,
gpu.DevicePrimesPow2Default.View,
gpu.DevicePrimesPow2LastOne.View,
gpu.DevicePrimesPow2LastSeven.View,
gpu.DevicePrimesPow2LastThree.View,
gpu.DevicePrimesPow2LastNine.View,
outputView);
accelerator.Synchronize();
outputView.CopyToCPU(ref flag, 1);
gpu.Dispose();
limiter.Dispose();
return flag != 0;
}
public static int GpuBatchSize { get; set; } = 262_144;
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static void IsPrimeBatchGpu(ReadOnlySpan<ulong> values, Span<byte> results)
{
var limiter = GpuPrimeWorkLimiter.Acquire();
var gpu = PrimeTesterGpuContextPool.Rent(GpuBatchSize);
var accelerator = gpu.Accelerator;
var kernel = gpu.Kernel;
int totalLength = values.Length;
int batchSize = GpuBatchSize;
{
var limiter = GpuPrimeWorkLimiter.Acquire();
var gpu = PrimeTesterGpuContextPool.Rent(GpuBatchSize);
var accelerator = gpu.Accelerator;
var kernel = gpu.Kernel;
int totalLength = values.Length;
int batchSize = GpuBatchSize;
var input = gpu.Input;
var output = gpu.Output;
var inputView = input.View;
var outputView = output.View;
var input = gpu.Input;
var output = gpu.Output;
var inputView = input.View;
var outputView = output.View;
int pos = 0;
while (pos < totalLength)
{
int remaining = totalLength - pos;
int count = remaining > batchSize ? batchSize : remaining;
int pos = 0;
while (pos < totalLength)
{
int remaining = totalLength - pos;
int count = remaining > batchSize ? batchSize : remaining;
var valueSlice = values.Slice(pos, count);
ref ulong valueRef = ref MemoryMarshal.GetReference(valueSlice);
inputView.CopyFromCPU(ref valueRef, count);
var valueSlice = values.Slice(pos, count);
ref ulong valueRef = ref MemoryMarshal.GetReference(valueSlice);
inputView.CopyFromCPU(ref valueRef, count);
kernel(
count,
inputView,
gpu.DevicePrimesDefault.View,
gpu.DevicePrimesLastOne.View,
gpu.DevicePrimesLastSeven.View,
gpu.DevicePrimesLastThree.View,
gpu.DevicePrimesLastNine.View,
gpu.DevicePrimesPow2Default.View,
gpu.DevicePrimesPow2LastOne.View,
gpu.DevicePrimesPow2LastSeven.View,
gpu.DevicePrimesPow2LastThree.View,
gpu.DevicePrimesPow2LastNine.View,
outputView);
accelerator.Synchronize();
var resultSlice = results.Slice(pos, count);
ref byte resultRef = ref MemoryMarshal.GetReference(resultSlice);
outputView.CopyToCPU(ref resultRef, count);
kernel(
count,
inputView,
gpu.DevicePrimesDefault.View,
gpu.DevicePrimesLastOne.View,
gpu.DevicePrimesLastSeven.View,
gpu.DevicePrimesLastThree.View,
gpu.DevicePrimesLastNine.View,
gpu.DevicePrimesPow2Default.View,
gpu.DevicePrimesPow2LastOne.View,
gpu.DevicePrimesPow2LastSeven.View,
gpu.DevicePrimesPow2LastThree.View,
gpu.DevicePrimesPow2LastNine.View,
outputView);
accelerator.Synchronize();
var resultSlice = results.Slice(pos, count);
ref byte resultRef = ref MemoryMarshal.GetReference(resultSlice);
outputView.CopyToCPU(ref resultRef, count);
pos += count;
}
pos += count;
}
gpu.Dispose();
limiter.Dispose();
}
gpu.Dispose();
limiter.Dispose();
}
internal static class PrimeTesterGpuContextPool
{
private static readonly ConcurrentQueue<PrimeTesterGpuContextLease> Pool = new();
internal static class PrimeTesterGpuContextPool
{
private static readonly ConcurrentQueue<PrimeTesterGpuContextLease> Pool = new();
internal static PrimeTesterGpuContextLease Rent(int minBufferCapacity = 1)
{
if (!Pool.TryDequeue(out var lease))
{
lease = new PrimeTesterGpuContextLease(minBufferCapacity);
}
else
{
lease.EnsureCapacity(minBufferCapacity);
}
internal static PrimeTesterGpuContextLease Rent(int minBufferCapacity = 1)
{
if (!Pool.TryDequeue(out var lease))
{
lease = new PrimeTesterGpuContextLease(minBufferCapacity);
}
else
{
lease.EnsureCapacity(minBufferCapacity);
}
lease.ResetReturnFlag();
return lease;
}
lease.ResetReturnFlag();
return lease;
}
internal static void Return(PrimeTesterGpuContextLease lease)
{
lease.Accelerator.Synchronize();
Pool.Enqueue(lease);
}
internal static void Return(PrimeTesterGpuContextLease lease)
{
lease.Accelerator.Synchronize();
Pool.Enqueue(lease);
}
internal static void DisposeAll()
{
while (Pool.TryDequeue(out var lease))
{
lease.DisposeResources();
}
}
internal static void DisposeAll()
{
while (Pool.TryDequeue(out var lease))
{
lease.DisposeResources();
}
}
internal static void Clear(Accelerator accelerator)
{
if (accelerator is null)
{
return;
}
internal static void Clear(Accelerator accelerator)
{
if (accelerator is null)
{
return;
}
var retained = new List<PrimeTesterGpuContextLease>();
while (Pool.TryDequeue(out var lease))
{
if (lease.Accelerator == accelerator)
{
lease.DisposeResources();
}
else
{
retained.Add(lease);
}
}
var retained = new List<PrimeTesterGpuContextLease>();
while (Pool.TryDequeue(out var lease))
{
if (lease.Accelerator == accelerator)
{
lease.DisposeResources();
}
else
{
retained.Add(lease);
}
}
int retainedCount = retained.Count;
for (int i = 0; i < retainedCount; i++)
{
Pool.Enqueue(retained[i]);
}
}
int retainedCount = retained.Count;
for (int i = 0; i < retainedCount; i++)
{
Pool.Enqueue(retained[i]);
}
}
internal sealed class PrimeTesterGpuContextLease : IDisposable
{
private bool _returned;
internal sealed class PrimeTesterGpuContextLease : IDisposable
{
private bool _returned;
internal readonly Context AcceleratorContext;
public readonly Accelerator Accelerator;
public readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> Kernel;
public readonly Action<Index1D, ArrayView<ulong>, ulong, ArrayView<int>> HeuristicTrialDivisionKernel;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesDefault;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2Default;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
public MemoryBuffer1D<ulong, Stride1D.Dense> Input;
public MemoryBuffer1D<byte, Stride1D.Dense> Output;
public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag;
public int BufferCapacity;
internal readonly Context AcceleratorContext;
public readonly Accelerator Accelerator;
public readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> Kernel;
public readonly Action<Index1D, ArrayView<ulong>, ulong, ArrayView<int>> HeuristicTrialDivisionKernel;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesDefault;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2Default;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
public MemoryBuffer1D<ulong, Stride1D.Dense> Input;
public MemoryBuffer1D<byte, Stride1D.Dense> Output;
public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag;
public int BufferCapacity;
internal PrimeTesterGpuContextLease(int minBufferCapacity)
{
AcceleratorContext = Context.CreateDefault();
Accelerator = AcceleratorContext.GetPreferredDevice(false).CreateAccelerator(AcceleratorContext);
Kernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel);
HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ulong, ArrayView<int>>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
internal PrimeTesterGpuContextLease(int minBufferCapacity)
{
AcceleratorContext = Context.CreateDefault();
Accelerator = AcceleratorContext.GetPreferredDevice(false).CreateAccelerator(AcceleratorContext);
Kernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel);
HeuristicTrialDivisionKernel = Accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ulong, ArrayView<int>>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
var primesDefault = DivisorGenerator.SmallPrimes;
DevicePrimesDefault = Accelerator.Allocate1D<uint>(primesDefault.Length);
DevicePrimesDefault.View.CopyFromCPU(primesDefault);
var primesDefault = DivisorGenerator.SmallPrimes;
DevicePrimesDefault = Accelerator.Allocate1D<uint>(primesDefault.Length);
DevicePrimesDefault.View.CopyFromCPU(primesDefault);
var primesLastOne = DivisorGenerator.SmallPrimesLastOne;
DevicePrimesLastOne = Accelerator.Allocate1D<uint>(primesLastOne.Length);
DevicePrimesLastOne.View.CopyFromCPU(primesLastOne);
var primesLastOne = DivisorGenerator.SmallPrimesLastOne;
DevicePrimesLastOne = Accelerator.Allocate1D<uint>(primesLastOne.Length);
DevicePrimesLastOne.View.CopyFromCPU(primesLastOne);
var primesLastSeven = DivisorGenerator.SmallPrimesLastSeven;
DevicePrimesLastSeven = Accelerator.Allocate1D<uint>(primesLastSeven.Length);
DevicePrimesLastSeven.View.CopyFromCPU(primesLastSeven);
var primesLastSeven = DivisorGenerator.SmallPrimesLastSeven;
DevicePrimesLastSeven = Accelerator.Allocate1D<uint>(primesLastSeven.Length);
DevicePrimesLastSeven.View.CopyFromCPU(primesLastSeven);
var primesLastThree = DivisorGenerator.SmallPrimesLastThree;
DevicePrimesLastThree = Accelerator.Allocate1D<uint>(primesLastThree.Length);
DevicePrimesLastThree.View.CopyFromCPU(primesLastThree);
var primesLastThree = DivisorGenerator.SmallPrimesLastThree;
DevicePrimesLastThree = Accelerator.Allocate1D<uint>(primesLastThree.Length);
DevicePrimesLastThree.View.CopyFromCPU(primesLastThree);
var primesLastNine = DivisorGenerator.SmallPrimesLastNine;
DevicePrimesLastNine = Accelerator.Allocate1D<uint>(primesLastNine.Length);
DevicePrimesLastNine.View.CopyFromCPU(primesLastNine);
var primesLastNine = DivisorGenerator.SmallPrimesLastNine;
DevicePrimesLastNine = Accelerator.Allocate1D<uint>(primesLastNine.Length);
DevicePrimesLastNine.View.CopyFromCPU(primesLastNine);
var primesPow2Default = DivisorGenerator.SmallPrimesPow2;
DevicePrimesPow2Default = Accelerator.Allocate1D<ulong>(primesPow2Default.Length);
DevicePrimesPow2Default.View.CopyFromCPU(primesPow2Default);
var primesPow2Default = DivisorGenerator.SmallPrimesPow2;
DevicePrimesPow2Default = Accelerator.Allocate1D<ulong>(primesPow2Default.Length);
DevicePrimesPow2Default.View.CopyFromCPU(primesPow2Default);
var primesPow2LastOne = DivisorGenerator.SmallPrimesPow2LastOne;
DevicePrimesPow2LastOne = Accelerator.Allocate1D<ulong>(primesPow2LastOne.Length);
DevicePrimesPow2LastOne.View.CopyFromCPU(primesPow2LastOne);
var primesPow2LastOne = DivisorGenerator.SmallPrimesPow2LastOne;
DevicePrimesPow2LastOne = Accelerator.Allocate1D<ulong>(primesPow2LastOne.Length);
DevicePrimesPow2LastOne.View.CopyFromCPU(primesPow2LastOne);
var primesPow2LastSeven = DivisorGenerator.SmallPrimesPow2LastSeven;
DevicePrimesPow2LastSeven = Accelerator.Allocate1D<ulong>(primesPow2LastSeven.Length);
DevicePrimesPow2LastSeven.View.CopyFromCPU(primesPow2LastSeven);
var primesPow2LastSeven = DivisorGenerator.SmallPrimesPow2LastSeven;
DevicePrimesPow2LastSeven = Accelerator.Allocate1D<ulong>(primesPow2LastSeven.Length);
DevicePrimesPow2LastSeven.View.CopyFromCPU(primesPow2LastSeven);
var primesPow2LastThree = DivisorGenerator.SmallPrimesPow2LastThree;
DevicePrimesPow2LastThree = Accelerator.Allocate1D<ulong>(primesPow2LastThree.Length);
DevicePrimesPow2LastThree.View.CopyFromCPU(primesPow2LastThree);
var primesPow2LastThree = DivisorGenerator.SmallPrimesPow2LastThree;
DevicePrimesPow2LastThree = Accelerator.Allocate1D<ulong>(primesPow2LastThree.Length);
DevicePrimesPow2LastThree.View.CopyFromCPU(primesPow2LastThree);
var primesPow2LastNine = DivisorGenerator.SmallPrimesPow2LastNine;
DevicePrimesPow2LastNine = Accelerator.Allocate1D<ulong>(primesPow2LastNine.Length);
DevicePrimesPow2LastNine.View.CopyFromCPU(primesPow2LastNine);
var primesPow2LastNine = DivisorGenerator.SmallPrimesPow2LastNine;
DevicePrimesPow2LastNine = Accelerator.Allocate1D<ulong>(primesPow2LastNine.Length);
DevicePrimesPow2LastNine.View.CopyFromCPU(primesPow2LastNine);
BufferCapacity = minBufferCapacity < 1 ? 1 : minBufferCapacity;
Input = Accelerator.Allocate1D<ulong>(BufferCapacity);
Output = Accelerator.Allocate1D<byte>(BufferCapacity);
HeuristicFlag = Accelerator.Allocate1D<int>(BufferCapacity);
_returned = false;
}
BufferCapacity = minBufferCapacity < 1 ? 1 : minBufferCapacity;
Input = Accelerator.Allocate1D<ulong>(BufferCapacity);
Output = Accelerator.Allocate1D<byte>(BufferCapacity);
HeuristicFlag = Accelerator.Allocate1D<int>(BufferCapacity);
_returned = false;
}
public void EnsureCapacity(int minCapacity)
{
int required = minCapacity < 1 ? 1 : minCapacity;
if (required <= BufferCapacity)
{
return;
}
public void EnsureCapacity(int minCapacity)
{
int required = minCapacity < 1 ? 1 : minCapacity;
if (required <= BufferCapacity)
{
return;
}
Input.Dispose();
Output.Dispose();
HeuristicFlag.Dispose();
Input = Accelerator.Allocate1D<ulong>(required);
Output = Accelerator.Allocate1D<byte>(required);
HeuristicFlag = Accelerator.Allocate1D<int>(required);
BufferCapacity = required;
}
Input.Dispose();
Output.Dispose();
HeuristicFlag.Dispose();
Input = Accelerator.Allocate1D<ulong>(required);
Output = Accelerator.Allocate1D<byte>(required);
HeuristicFlag = Accelerator.Allocate1D<int>(required);
BufferCapacity = required;
}
internal void ResetReturnFlag()
{
_returned = false;
}
internal void ResetReturnFlag()
{
_returned = false;
}
public void Dispose()
{
if (_returned)
{
return;
}
public void Dispose()
{
if (_returned)
{
return;
}
_returned = true;
PrimeTesterGpuContextPool.Return(this);
}
_returned = true;
PrimeTesterGpuContextPool.Return(this);
}
internal void DisposeResources()
{
Input.Dispose();
Output.Dispose();
HeuristicFlag.Dispose();
DevicePrimesDefault.Dispose();
DevicePrimesLastOne.Dispose();
DevicePrimesLastSeven.Dispose();
DevicePrimesLastThree.Dispose();
DevicePrimesLastNine.Dispose();
DevicePrimesPow2Default.Dispose();
DevicePrimesPow2LastOne.Dispose();
DevicePrimesPow2LastSeven.Dispose();
DevicePrimesPow2LastThree.Dispose();
DevicePrimesPow2LastNine.Dispose();
Accelerator.Dispose();
AcceleratorContext.Dispose();
}
}
}
// Expose cache clearing for accelerator disposal coordination
internal void DisposeResources()
{
Input.Dispose();
Output.Dispose();
HeuristicFlag.Dispose();
DevicePrimesDefault.Dispose();
DevicePrimesLastOne.Dispose();
DevicePrimesLastSeven.Dispose();
DevicePrimesLastThree.Dispose();
DevicePrimesLastNine.Dispose();
DevicePrimesPow2Default.Dispose();
DevicePrimesPow2LastOne.Dispose();
DevicePrimesPow2LastSeven.Dispose();
DevicePrimesPow2LastThree.Dispose();
DevicePrimesPow2LastNine.Dispose();
Accelerator.Dispose();
AcceleratorContext.Dispose();
}
}
}
// Expose cache clearing for accelerator disposal coordination
public static void ClearGpuCaches(Accelerator accelerator)
{
PrimeTesterGpuContextPool.Clear(accelerator);