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

GpuKernelLease removed

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-07 15:18:17 +01:00
1 parent 0410fd2c91
commit da3093a9ba
12 files changed
+58 -105
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core.Tests/Gpu/GpuKernelCompilationTests.cs
+15 -15
View File
@@ -330,7 +330,7 @@ public class GpuKernelCompilationTests
private static void CompileMersenneDivisorCyclesKernel(Accelerator accelerator)
{
var kernel = accelerator.LoadAutoGroupedStreamKernel(MersenneDivisorCyclesKernel);
var kernel = accelerator.LoadAutoGroupedStreamKernel(MersenneDivisorCyclesKernel);
_ = kernel;
}
@@ -349,32 +349,32 @@ public class GpuKernelCompilationTests
private static void CompileGpuKernelPoolKernels(Accelerator accelerator)
{
_ = accelerator;
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
try
{
_ = GpuKernelLease.OrderKernel;
_ = GpuKernelLease.IncrementalKernel;
_ = GpuKernelLease.Pow2ModKernel;
_ = GpuKernelLease.IncrementalOrderKernel;
_ = GpuKernelLease.Pow2ModOrderKernel;
_ = GpuKernelPool.Kernels.Order;
_ = GpuKernelPool.Kernels.Incremental;
_ = GpuKernelPool.Kernels.Pow2Mod;
_ = GpuKernelPool.Kernels.IncrementalOrder;
_ = GpuKernelPool.Kernels.Pow2ModOrder;
}
finally
{
lease.Dispose();
GpuPrimeWorkLimiter.Release();
}
var orderLease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
try
{
_ = GpuKernelLease.OrderKernel;
_ = GpuKernelLease.IncrementalKernel;
_ = GpuKernelLease.Pow2ModKernel;
_ = GpuKernelLease.IncrementalOrderKernel;
_ = GpuKernelLease.Pow2ModOrderKernel;
_ = GpuKernelPool.Kernels.Order;
_ = GpuKernelPool.Kernels.Incremental;
_ = GpuKernelPool.Kernels.Pow2Mod;
_ = GpuKernelPool.Kernels.IncrementalOrder;
_ = GpuKernelPool.Kernels.Pow2ModOrder;
}
finally
{
orderLease.Dispose();
GpuPrimeWorkLimiter.Release();
}
}
}
PerfectNumbers.Core/DivisorCycleCache.cs
+2 -2
View File
@@ -267,7 +267,7 @@ public sealed class DivisorCycleCache
private void ComputeCyclesGpuCore(ReadOnlySpan<ulong> divisors, Span<ulong> destination)
{
int length = divisors.Length;
var gpuLease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
@@ -349,7 +349,7 @@ public sealed class DivisorCycleCache
ThreadStaticPools.BytePool.Return(rentedStatus!, clearArray: false);
}
gpuLease.Dispose();
GpuPrimeWorkLimiter.Release();
}
private static Action<AcceleratorStream, Index1D, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>> LoadKernel(Accelerator accelerator)
PerfectNumbers.Core/Gpu/GpuKernelLease.cs
-40
View File
@@ -1,40 +0,0 @@
using System.Collections.Concurrent;
using ILGPU;
using ILGPU.Runtime;
namespace PerfectNumbers.Core.Gpu;
public sealed class GpuKernelLease
{
private static readonly ConcurrentQueue<GpuKernelLease> Pool = new();
internal static GpuKernelLease Rent()
{
if (!Pool.TryDequeue(out var lease))
{
lease = new GpuKernelLease();
}
return lease;
}
public static readonly Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>> OrderKernel = GpuKernelPool.Kernels.Order!;
public static readonly 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>> Pow2ModKernel = GpuKernelPool.Kernels.Pow2Mod!;
public static readonly Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>> IncrementalKernel = GpuKernelPool.Kernels.Incremental!;
public static readonly Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>> IncrementalOrderKernel = GpuKernelPool.Kernels.IncrementalOrder!;
public static readonly Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>> Pow2ModOrderKernel = GpuKernelPool.Kernels.Pow2ModOrder!;
public static readonly Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SpecialMaxKernel = GpuKernelPool.Kernels.SpecialMax!;
public static readonly 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>> SmallPrimeFactorKernel = GpuKernelPool.Kernels.SmallPrimeFactor!;
public void Dispose()
{
Pool.Enqueue(this);
GpuPrimeWorkLimiter.Release();
}
}
PerfectNumbers.Core/Gpu/GpuKernelPool.cs
+2 -9
View File
@@ -201,25 +201,18 @@ public class GpuKernelPool
return new SmallPrimeFactorViews(devicePrimes, deviceSquares);
}
public static GpuKernelLease Rent()
{
GpuPrimeWorkLimiter.Acquire();
var stream = SharedGpuContext.Accelerator.CreateStream();
return GpuKernelLease.Rent();
}
/// <summary>
/// Runs a GPU action with an acquired accelerator and stream.
/// </summary>
/// <param name="action">Action to run with (Accelerator, Stream).</param>
public static void Run(Action<Accelerator, AcceleratorStream> action)
{
var lease = Rent();
GpuPrimeWorkLimiter.Acquire();
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
action(accelerator, stream);
stream.Synchronize();
stream.Dispose();
lease.Dispose();
GpuPrimeWorkLimiter.Release();
}
}
PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs
+3 -4
View File
@@ -95,7 +95,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
ulong processedCount;
ulong lastProcessed;
var gpuLease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
@@ -129,7 +129,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
stream.Dispose();
ReturnBatchResources(resources);
// Monitor.Exit(gpuLease.ExecutionLock);
gpuLease.Dispose();
GpuPrimeWorkLimiter.Release();
if (composite)
{
@@ -485,7 +485,6 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
public sealed class DivisorScanSession : IMersenneNumberDivisorByDivisorTester.IDivisorScanSession
{
private readonly MersenneNumberDivisorByDivisorGpuTester _owner;
private readonly GpuKernelLease _lease;
private readonly Accelerator _accelerator;
private readonly AcceleratorStream _stream;
private Action<AcceleratorStream, Index1D, ArrayView<GpuDivisorPartialData>, ArrayView<int>, ArrayView<int>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>, ArrayView<int>> _kernel = null!;
@@ -502,7 +501,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
internal DivisorScanSession(MersenneNumberDivisorByDivisorGpuTester owner)
{
_owner = owner;
_lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
_accelerator = accelerator;
_stream = accelerator.CreateStream();
PerfectNumbers.Core/Gpu/MersenneNumberIncrementalGpuTester.cs
+5 -4
View File
@@ -16,7 +16,7 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
{
throw new NotImplementedException($"GPU incremental scanning requires the device cycle heuristics implementation (kernel: {_kernelType}, GPU order: {_useGpuOrder}).");
var gpuLease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy
@@ -24,8 +24,9 @@ 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
var pow2Kernel = GpuKernelLease.Pow2ModKernel;
var incKernel = GpuKernelLease.IncrementalKernel;
KernelContainer kernels = GpuKernelPool.Kernels;
var pow2Kernel = kernels.Pow2Mod!;
var incKernel = kernels.Incremental!;
ulong step10 = (exponent.Mod10() << 1).Mod10();
ulong step8 = ((exponent & 7UL) << 1) & 7UL;
ulong step3 = ((exponent % 3UL) << 1) % 3UL;
@@ -124,6 +125,6 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
stream.Dispose();
ulongPool.Return(orderArray);
orderBuffer.Dispose();
gpuLease.Dispose();
GpuPrimeWorkLimiter.Release();
}
}
PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs
+9 -6
View File
@@ -9,7 +9,7 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
{
var gpuLease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy and avoids TDR
@@ -17,10 +17,11 @@ 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.Kernels;
var kernel = _kernelType switch
{
GpuKernelType.Pow2Mod => GpuKernelLease.Pow2ModOrderKernel,
_ => GpuKernelLease.IncrementalOrderKernel,
GpuKernelType.Pow2Mod => kernels.Pow2ModOrder!,
_ => kernels.IncrementalOrder!,
};
var foundBuffer = accelerator.Allocate1D<int>(1); // TODO: Replace this allocation with the pooled device buffer from GpuOrderKernelBenchmarks so repeated scans reuse the pinned staging memory instead of allocating per run.
@@ -46,8 +47,10 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
var kernelArgs = new ResidueAutomatonArgs(q0m10, step10, q0m8, step8, q0m3, step3, q0m5, step5);
// Ensure device has small cycles table for early rejections in order kernels
kernel(stream, currentSize, exponent, (GpuUInt128)twoP, (GpuUInt128)kStart, last, divMul,
kernelArgs, foundBuffer.View, smallCyclesView);
kernel(
stream, currentSize, exponent, (GpuUInt128)twoP, (GpuUInt128)kStart,
last, divMul, kernelArgs, foundBuffer.View, smallCyclesView
);
foundBuffer.View.CopyToCPU(stream, ref found, 1);
stream.Synchronize();
@@ -62,6 +65,6 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
stream.Dispose();
foundBuffer.Dispose();
gpuLease.Dispose();
GpuPrimeWorkLimiter.Release();
}
}
PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs
+9 -12
View File
@@ -2,10 +2,7 @@ using System.Buffers;
using System.Collections.Concurrent;
using ILGPU;
using ILGPU.Runtime;
using System.Numerics;
using PerfectNumbers.Core;
using System.Runtime.InteropServices;
using System.Runtime.CompilerServices;
using ILGPU.Runtime.OpenCL;
namespace PerfectNumbers.Core.Gpu;
@@ -198,7 +195,7 @@ internal static partial class PrimeOrderGpuHeuristics
try
{
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = accelerator.CreateStream();
@@ -254,7 +251,7 @@ internal static partial class PrimeOrderGpuHeuristics
countBuffer.Dispose();
remainingBuffer.Dispose();
fullyFactoredBuffer.Dispose();
lease.Dispose();
GpuPrimeWorkLimiter.Release();
return true;
}
catch (CLException ex)
@@ -366,7 +363,7 @@ internal static partial class PrimeOrderGpuHeuristics
{
order = 0UL;
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = accelerator.CreateStream();
@@ -444,7 +441,7 @@ internal static partial class PrimeOrderGpuHeuristics
stackProductBuffer.Dispose();
resultBuffer.Dispose();
statusBuffer.Dispose();
lease.Dispose();
GpuPrimeWorkLimiter.Release();
PrimeOrderKernelStatus kernelStatus = (PrimeOrderKernelStatus)status;
if (kernelStatus == PrimeOrderKernelStatus.Fallback)
@@ -519,7 +516,7 @@ internal static partial class PrimeOrderGpuHeuristics
private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
{
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream? stream = accelerator.CreateStream();
var kernel = GetPow2ModKernel(accelerator);
@@ -544,7 +541,7 @@ internal static partial class PrimeOrderGpuHeuristics
stream = null;
exponentBuffer.Dispose();
remainderBuffer.Dispose();
lease.Dispose();
GpuPrimeWorkLimiter.Release();
}
catch (Exception)
{
@@ -552,7 +549,7 @@ internal static partial class PrimeOrderGpuHeuristics
stream?.Dispose();
exponentBuffer.Dispose();
remainderBuffer.Dispose();
lease.Dispose();
GpuPrimeWorkLimiter.Release();
throw;
}
@@ -578,7 +575,7 @@ internal static partial class PrimeOrderGpuHeuristics
GpuUInt128[]? rentedExponents = null;
GpuUInt128[]? rentedResults = null;
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
ArrayPool<GpuUInt128> gpuUInt128Pool = ThreadStaticPools.GpuUInt128Pool;
try
@@ -639,7 +636,7 @@ internal static partial class PrimeOrderGpuHeuristics
gpuUInt128Pool.Return(rentedResults!, clearArray: false);
}
lease.Dispose();
GpuPrimeWorkLimiter.Release();
}
}
PerfectNumbers.Core/MersenneDivisorCycles.cs
+2 -2
View File
@@ -804,7 +804,7 @@ public class MersenneDivisorCycles
divisors = pool.Rent(count);
outCycles = pool.Rent(count);
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var loaded = accelerator.LoadAutoGroupedStreamKernel<
@@ -898,7 +898,7 @@ public class MersenneDivisorCycles
pool.Return(divisors, clearArray: false);
pool.Return(outCycles, clearArray: false);
lease.Dispose();
GpuPrimeWorkLimiter.Release();
}
PerfectNumbers.Core/MersenneNumberTester.cs
+3 -3
View File
@@ -147,10 +147,10 @@ public sealed class MersenneNumberTester(
UInt128[] qs = qsBuffer;
var gpuLease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var orderKernel = GpuKernelLease.OrderKernel!;
var orderKernel = GpuKernelPool.Kernels.Order!;
// Guard long-running kernels by chunking the warm-up across batches.
// Reuse the same ScanBatchSize knob used for scanning.
@@ -197,7 +197,7 @@ public sealed class MersenneNumberTester(
}
stream.Dispose();
gpuLease.Dispose();
GpuPrimeWorkLimiter.Release();
uInt128Pool.Return(qs);
}
PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs
+6 -6
View File
@@ -19,7 +19,7 @@ internal static partial class PrimeOrderCalculator
Span<int> exponentBuffer = exponentBufferArray.AsSpan(0, GpuSmallPrimeFactorSlots);
remaining = value;
var lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
KernelContainer kernels = GpuKernelPool.Kernels;
@@ -31,7 +31,7 @@ internal static partial class PrimeOrderCalculator
if (kernels.SmallPrimeFactorsPrimes is null) throw new NullReferenceException($"{nameof(kernels.SmallPrimeFactorsPrimes)} is null");
if (kernels.SmallPrimeFactorsSquares is null) throw new NullReferenceException($"{nameof(kernels.SmallPrimeFactorsSquares)} is null");
var kernel = GpuKernelLease.SmallPrimeFactorKernel;
var kernel = GpuKernelPool.Kernels.SmallPrimeFactor!;
ArrayView1D<int, Stride1D.Dense> smallPrimeFactorCountSlotView = scratch.SmallPrimeFactorCountSlot.View;
ArrayView1D<ulong, Stride1D.Dense> smallPrimeFactorRemainingSlotView = scratch.SmallPrimeFactorRemainingSlot.View;
@@ -78,7 +78,7 @@ internal static partial class PrimeOrderCalculator
}
GpuScratchBufferPool.Return(scratch);
lease.Dispose();
GpuPrimeWorkLimiter.Release();
return true;
}
@@ -101,7 +101,7 @@ internal static partial class PrimeOrderCalculator
factorSpan[i] = factors[i].Value;
}
GpuKernelLease lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
@@ -111,7 +111,7 @@ internal static partial class PrimeOrderCalculator
ArrayView1D<ulong, Stride1D.Dense> specialMaxFactorsView = scratch.SpecialMaxFactors.View;
specialMaxFactorsView.SubView(0, factorCount).CopyFromCPU(stream, ref MemoryMarshal.GetReference(factorSpan), factorCount);
var kernel = GpuKernelLease.SpecialMaxKernel;
var kernel = GpuKernelPool.Kernels.SpecialMax!;
ArrayView1D<ulong, Stride1D.Dense> specialMaxResultView = scratch.SpecialMaxResult.View;
@@ -134,7 +134,7 @@ internal static partial class PrimeOrderCalculator
// stream.Synchronize();
// return result != 0;
GpuScratchBufferPool.Return(scratch);
lease.Dispose();
GpuPrimeWorkLimiter.Release();
return result[0] != 0;
}
}
PerfectNumbers.Core/ULongExtensions.Gpu.cs
+2 -2
View File
@@ -461,7 +461,7 @@ public static partial class ULongExtensions
{
ulong result = 0UL;
GpuKernelLease lease = GpuKernelPool.Rent();
GpuPrimeWorkLimiter.Acquire();
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = accelerator.CreateStream();
@@ -484,7 +484,7 @@ public static partial class ULongExtensions
stream.Dispose();
resultBuffer.Dispose();
exponentBuffer.Dispose();
lease.Dispose();
GpuPrimeWorkLimiter.Release();
return result;
}