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

PrimeOrderCalculatorAccelerator.Rent is now created outside of the loops and re-used for all calculations.

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-24 22:51:53 +01:00
1 parent 57a50e2a37
commit b9ca95bf8c
53 files changed
+4639 -4229
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs
+38 -53
View File
@@ -20,7 +20,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
private ulong _divisorLimit;
// private bool _isConfigured;
private readonly ConcurrentBag<MersenneNumberDivisorByDivisorAccelerator>[] _resourcePool = [..AcceleratorPool.Shared.Accelerators.Select(x => new ConcurrentBag<MersenneNumberDivisorByDivisorAccelerator>())];
private readonly ConcurrentBag<MersenneNumberDivisorByDivisorAccelerator>[] _resourcePool = [.. AcceleratorPool.Shared.Accelerators.Select(x => new ConcurrentBag<MersenneNumberDivisorByDivisorAccelerator>())];
private readonly ConcurrentBag<DivisorScanSession> _sessionPool = [];
@@ -50,9 +50,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
// _isConfigured = true;
}
private static readonly Accelerator[] _accelerators = AcceleratorPool.Shared.Accelerators;
public bool IsPrime(ulong prime, out bool divisorsExhausted, out ulong divisor)
public bool IsPrime(PrimeOrderCalculatorAccelerator gpu, ulong prime, out bool divisorsExhausted, out ulong divisor)
{
ulong allowedMax;
int batchCapacity;
@@ -87,7 +85,6 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
// GpuPrimeWorkLimiter();
// int acceleratorIndex = AcceleratorPool.Shared.Rent();
// var accelerator = _accelerators[acceleratorIndex];
var gpu = PrimeOrderCalculatorAccelerator.Rent(batchCapacity);
var acceleratorIndex = gpu.AcceleratorIndex;
// Monitor.Enter(gpuLease.ExecutionLock);
@@ -117,7 +114,6 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
out processedCount);
ReturnBatchResources(acceleratorIndex, resources);
PrimeOrderCalculatorAccelerator.Return(gpu);
// Monitor.Exit(gpuLease.ExecutionLock);
// GpuPrimeWorkLimiter.Release();
@@ -364,7 +360,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
{
if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(gpu, divisor, prime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
{
return MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
return MersenneDivisorCycles.CalculateCycleLength(gpu, divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
}
return computedCycle;
@@ -481,8 +477,8 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
public sealed class DivisorScanSession : IMersenneNumberDivisorByDivisorTester.IDivisorScanSession
{
private readonly MersenneNumberDivisorByDivisorGpuTester _owner;
private readonly PrimeOrderCalculatorAccelerator _primeTesterAccelerator;
private readonly Accelerator _accelerator;
private readonly PrimeOrderCalculatorAccelerator _primeOrderCalculatorAccelerator;
private readonly AcceleratorStream _stream;
private Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> _kernel = null!;
private MemoryBuffer1D<GpuDivisorPartialData, Stride1D.Dense> _divisorBuffer = null!;
@@ -496,15 +492,36 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
private int _capacity;
internal DivisorScanSession(MersenneNumberDivisorByDivisorGpuTester owner)
internal DivisorScanSession(PrimeOrderCalculatorAccelerator gpu, MersenneNumberDivisorByDivisorGpuTester owner)
{
_owner = owner;
// GpuPrimeWorkLimiter.Acquire();
_primeTesterAccelerator = PrimeOrderCalculatorAccelerator.Rent(1);
Accelerator accelerator = _primeTesterAccelerator.Accelerator;
_primeOrderCalculatorAccelerator = gpu;
Accelerator accelerator = gpu.Accelerator;
_accelerator = accelerator;
_stream = accelerator.CreateStream();
_capacity = Math.Max(1, owner._gpuBatchSize);
// lock (accelerator)
{
_divisorBuffer = accelerator.Allocate1D<GpuDivisorPartialData>(1);
_offsetBuffer = accelerator.Allocate1D<int>(1);
_countBuffer = accelerator.Allocate1D<int>(1);
_cycleBuffer = accelerator.Allocate1D<ulong>(1);
_firstHitBuffer = accelerator.Allocate1D<int>(1);
}
int desiredCapacity = PerfectNumberConstants.DefaultSmallPrimeFactorSlotCount;
_capacity = desiredCapacity;
// lock (accelerator)
{
_exponentsBuffer = accelerator.Allocate1D<ulong>(desiredCapacity);
_hitBuffer = accelerator.Allocate1D<byte>(desiredCapacity);
}
_hostBuffer = ThreadStaticPools.UlongPool.Rent(desiredCapacity);
_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>>>();
}
internal void Reset()
@@ -513,45 +530,16 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
private void EnsureExecutionResourcesLocked(int requiredCapacity)
{
Accelerator accelerator = _accelerator;
if (_kernel == null)
if (requiredCapacity > _capacity)
{
_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>>>();
}
_exponentsBuffer.Dispose();
_hitBuffer.Dispose();
ThreadStaticPools.UlongPool.Return(_hostBuffer, clearArray: false);
if (_divisorBuffer == null)
{
// lock (accelerator)
{
_divisorBuffer = accelerator.Allocate1D<GpuDivisorPartialData>(1);
_offsetBuffer = accelerator.Allocate1D<int>(1);
_countBuffer = accelerator.Allocate1D<int>(1);
_cycleBuffer = accelerator.Allocate1D<ulong>(1);
_firstHitBuffer = accelerator.Allocate1D<int>(1);
}
}
bool allocateBuffers = _exponentsBuffer == null;
if (!allocateBuffers && _capacity < requiredCapacity)
{
_exponentsBuffer?.Dispose();
_hitBuffer?.Dispose();
if (_hostBuffer != null)
{
ThreadStaticPools.UlongPool.Return(_hostBuffer, clearArray: false);
_hostBuffer = null!;
}
_exponentsBuffer = null!;
_hitBuffer = null!;
allocateBuffers = true;
}
if (allocateBuffers)
{
Accelerator accelerator = _accelerator;
int desiredCapacity = requiredCapacity > _capacity ? requiredCapacity : _capacity;
_capacity = desiredCapacity;
// lock (accelerator)
{
_exponentsBuffer = accelerator.Allocate1D<ulong>(desiredCapacity);
@@ -581,13 +569,12 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
ulong cycle = divisorCycle;
ulong firstPrime = primes[0];
var gpu = _primeOrderCalculatorAccelerator;
if (cycle == 0UL)
{
var gpu = _primeTesterAccelerator;
if (!MersenneDivisorCycles.TryCalculateCycleLengthForExponentCpu(gpu, divisor, firstPrime, divisorData, out ulong computedCycle, out bool primeOrderFailed) || computedCycle == 0UL)
{
cycle = MersenneDivisorCycles.CalculateCycleLength(divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
cycle = MersenneDivisorCycles.CalculateCycleLength(gpu, divisor, divisorData, skipPrimeOrderHeuristic: primeOrderFailed);
}
else
{
@@ -714,7 +701,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
return ComputeAllowedMaxDivisorGpu(prime, _divisorLimit);
}
public DivisorScanSession CreateDivisorSession()
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.
// lock (_sync)
@@ -739,8 +726,6 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
return session;
}
return new DivisorScanSession(this);
return new DivisorScanSession(gpu, this);
}
IMersenneNumberDivisorByDivisorTester.IDivisorScanSession IMersenneNumberDivisorByDivisorTester.CreateDivisorSession() => CreateDivisorSession();
}
PerfectNumbers.Core/Gpu/MersenneNumberDivisorGpuTester.cs
+3 -2
View File
@@ -1,5 +1,6 @@
using ILGPU;
using ILGPU.Runtime;
using PerfectNumbers.Core.Gpu.Accelerators;
namespace PerfectNumbers.Core.Gpu;
@@ -86,7 +87,7 @@ public sealed class MersenneNumberDivisorGpuTester
private static (ulong divisor, uint cycle)[]? _divisorCandidates = [];
public bool IsPrime(ulong p, UInt128 d, ulong divisorCyclesSearchLimit, out bool divisorsExhausted)
public bool IsPrime(PrimeOrderCalculatorAccelerator gpu, ulong p, UInt128 d, ulong divisorCyclesSearchLimit, out bool divisorsExhausted)
{
ReadOnlyGpuUInt128 readOnlyDivisor;
@@ -142,7 +143,7 @@ public sealed class MersenneNumberDivisorGpuTester
continue;
}
UInt128 cycle128 = MersenneDivisorCycles.GetCycle(d);
UInt128 cycle128 = MersenneDivisorCycles.GetCycle(gpu, d);
if ((UInt128)p % cycle128 != UInt128.Zero)
{
continue;
PerfectNumbers.Core/Gpu/MersenneNumberLucasLehmerGpuTester.cs
+3 -2
View File
@@ -2,6 +2,7 @@ using System.Buffers;
using System.Collections.Concurrent;
using ILGPU;
using ILGPU.Runtime;
using PerfectNumbers.Core.Gpu.Accelerators;
using UInt128 = System.UInt128;
namespace PerfectNumbers.Core.Gpu;
@@ -131,7 +132,7 @@ public class MersenneNumberLucasLehmerGpuTester
// Configurable LL slice size to keep kernels short. Default 32.
public int SliceSize = 32;
public bool IsPrime(ulong exponent, bool runOnGpu)
public bool IsPrime(PrimeOrderCalculatorAccelerator gpu, ulong exponent, bool runOnGpu)
{
// Early rejections aligned with incremental/order sieves, but safe for small p:
// - If 3 | p and p != 3, then 7 | M_p -> composite.
@@ -142,7 +143,7 @@ public class MersenneNumberLucasLehmerGpuTester
return false;
}
if ((exponent & 3UL) == 1UL && exponent.SharesFactorWithExponentMinusOne())
if ((exponent & 3UL) == 1UL && exponent.SharesFactorWithExponentMinusOne(gpu))
{
return false;
}
PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs
+1 -33
View File
@@ -8,12 +8,11 @@ namespace PerfectNumbers.Core.Gpu;
public static class Pow2MontgomeryGpuCalculator
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static ulong CalculateKeep(ulong exponent, in MontgomeryDivisorData divisor)
public static ulong CalculateKeep(PrimeOrderCalculatorAccelerator gpu, ulong exponent, in MontgomeryDivisorData divisor)
{
// GpuPrimeWorkLimiter.Acquire();
ulong result = 0UL;
var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
int acceleratorIndex = gpu.AcceleratorIndex;
Accelerator? accelerator = gpu.Accelerator;
MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
@@ -32,37 +31,6 @@ public static class Pow2MontgomeryGpuCalculator
AcceleratorStreamPool.Return(acceleratorIndex, stream);
// GpuPrimeWorkLimiter.Release();
PrimeOrderCalculatorAccelerator.Return(gpu);
return result;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static ulong CalculateConvert(ulong exponent, in MontgomeryDivisorData divisor)
{
// GpuPrimeWorkLimiter.Acquire();
ulong result = 0UL;
var gpu = PrimeOrderCalculatorAccelerator.Rent(1);
MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
int acceleratorIndex = gpu.AcceleratorIndex;
AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
exponentBuffer.View.CopyFromCPU(stream, ref exponent, 1);
// We don't need to worry about any left-overs here.
// resultBuffer.MemSetToZero(stream);
var kernelLauncher = gpu.ConvertToStandardKernelLauncher!;
kernelLauncher(stream, 1, exponentBuffer.View, divisor.Modulus, divisor.NPrime, divisor.MontgomeryOne, divisor.MontgomeryTwo, divisor.MontgomeryTwoSquared, resultBuffer.View);
resultBuffer.View.CopyToCPU(stream, ref result, 1);
stream.Synchronize();
AcceleratorStreamPool.Return(acceleratorIndex, stream);
// GpuPrimeWorkLimiter.Release();
PrimeOrderCalculatorAccelerator.Return(gpu);
return result;
}
PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs
+15 -13
View File
@@ -243,7 +243,7 @@ internal static partial class PrimeOrderGpuHeuristics
}
}
public static GpuPow2ModStatus TryPow2Mod(ulong exponent, ulong prime, out ulong remainder, in MontgomeryDivisorData divisorData)
public static GpuPow2ModStatus TryPow2Mod(ulong exponent, ulong prime, out ulong remainder, in MontgomeryDivisorData? divisorData)
{
Span<ulong> exponents = stackalloc ulong[1];
Span<ulong> remainders = stackalloc ulong[1];
@@ -254,7 +254,7 @@ internal static partial class PrimeOrderGpuHeuristics
return status;
}
public static GpuPow2ModStatus TryPow2ModBatch(ReadOnlySpan<ulong> exponents, ulong prime, Span<ulong> remainders, in MontgomeryDivisorData divisorData)
public static GpuPow2ModStatus TryPow2ModBatch(ReadOnlySpan<ulong> exponents, ulong prime, Span<ulong> remainders, in MontgomeryDivisorData? divisorData)
{
// This will never occur in production code
// if (exponents.Length == 0)
@@ -309,20 +309,20 @@ internal static partial class PrimeOrderGpuHeuristics
return computed ? GpuPow2ModStatus.Success : GpuPow2ModStatus.Unavailable;
}
public static GpuPow2ModStatus TryPow2Mod(in UInt128 exponent, in UInt128 prime, out UInt128 remainder)
public static GpuPow2ModStatus TryPow2Mod(PrimeOrderCalculatorAccelerator gpu, in UInt128 exponent, in UInt128 prime, out UInt128 remainder)
{
Span<UInt128> exponents = stackalloc UInt128[1];
Span<UInt128> remainders = stackalloc UInt128[1];
exponents[0] = exponent;
GpuPow2ModStatus status = TryPow2ModBatch(exponents, prime, remainders);
GpuPow2ModStatus status = TryPow2ModBatch(gpu, exponents, prime, remainders);
remainder = remainders[0];
return status;
}
public static GpuPow2ModStatus TryPow2ModBatch(ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> remainders)
public static GpuPow2ModStatus TryPow2ModBatch(PrimeOrderCalculatorAccelerator gpu, ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> remainders)
{
return TryPow2ModBatchInternal(exponents, prime, remainders);
return TryPow2ModBatchInternal(gpu, exponents, prime, remainders);
}
internal static bool TryCalculateOrder(
@@ -471,7 +471,7 @@ internal static partial class PrimeOrderGpuHeuristics
return kernel;
}
private static GpuPow2ModStatus TryPow2ModBatchInternal(ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> remainders)
private static GpuPow2ModStatus TryPow2ModBatchInternal(PrimeOrderCalculatorAccelerator gpu, ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> remainders)
{
if (exponents.Length == 0)
{
@@ -509,11 +509,11 @@ internal static partial class PrimeOrderGpuHeuristics
}
}
ComputeOnGpuWide(exponents, prime, target);
ComputeOnGpuWide(gpu, exponents, prime, target);
return GpuPow2ModStatus.Success;
}
private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, MontgomeryDivisorData? divisorData, Span<ulong> results)
{
// GpuPrimeWorkLimiter.Acquire();
int acceleratorIndex = AcceleratorPool.Shared.Rent();
@@ -540,6 +540,11 @@ internal static partial class PrimeOrderGpuHeuristics
try
{
if (divisorData is null)
{
divisorData = MontgomeryDivisorData.Empty;
}
kernelLauncher(stream, exponents.Length, exponentBuffer.View, divisorData.Modulus, divisorData.NPrime, divisorData.MontgomeryOne, divisorData.MontgomeryTwo, divisorData.MontgomeryTwoSquared, remainderBuffer.View);
remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(results), exponents.Length);
@@ -573,14 +578,13 @@ internal static partial class PrimeOrderGpuHeuristics
}
}
private static void ComputeOnGpuWide(ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> results)
private static void ComputeOnGpuWide(PrimeOrderCalculatorAccelerator gpu, ReadOnlySpan<UInt128> exponents, UInt128 prime, Span<UInt128> results)
{
int length = exponents.Length;
GpuUInt128[]? rentedExponents = null;
GpuUInt128[]? rentedResults = null;
ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
PrimeOrderCalculatorAccelerator? gpu = null;
bool poolingRequired = length > WideStackThreshold;
Span<GpuUInt128> exponentSpan = poolingRequired
? stackalloc GpuUInt128[length]
@@ -596,7 +600,6 @@ internal static partial class PrimeOrderGpuHeuristics
}
GpuUInt128 modulus = (GpuUInt128)prime;
gpu = PrimeOrderCalculatorAccelerator.Rent(1);
int acceleratorIndex = gpu.AcceleratorIndex;
Accelerator accelerator = gpu.Accelerator;
@@ -625,7 +628,6 @@ internal static partial class PrimeOrderGpuHeuristics
remainderBuffer.Dispose();
exponentBuffer.Dispose();
PrimeOrderCalculatorAccelerator.Return(gpu);
if (rentedExponents is not null)
{