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

Inline GPU divisor locking adjustments

Browse Source
This commit is contained in:
Marek Lemańczyk committed 2025-10-21 21:19:57 +02:00
1 parent 9bcf1bd9a1
commit 5dd9eb636d
1 file changed
+53 -40
Download Patch File Download Diff File Expand all files Collapse all files
PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs
+53 -40
View File
@@ -2,6 +2,7 @@ using System.Buffers;
using System.Collections.Concurrent; using System.Collections.Concurrent;
using System.Runtime.CompilerServices; using System.Runtime.CompilerServices;
using System.Runtime.InteropServices; using System.Runtime.InteropServices;
using System.Threading;
using ILGPU; using ILGPU;
using ILGPU.Runtime; using ILGPU.Runtime;
@@ -96,42 +97,41 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
var gpuLease = GpuContextPool.RentPreferred(preferCpu: false); var gpuLease = GpuContextPool.RentPreferred(preferCpu: false);
var accelerator = gpuLease.Accelerator; var accelerator = gpuLease.Accelerator;
var kernel = GetKernel(accelerator);
BatchResources resources = RentBatchResources(accelerator, batchCapacity);
try Monitor.Enter(gpuLease.ExecutionLock);
{
composite = CheckDivisors( var kernel = GetKernel(accelerator);
prime, var resources = RentBatchResources(accelerator, batchCapacity);
allowedMax,
accelerator, composite = CheckDivisors(
kernel, prime,
resources.DivisorsBuffer, allowedMax,
resources.ExponentBuffer, accelerator,
resources.HitsBuffer, kernel,
resources.DivisorDeltaBuffer, resources.DivisorsBuffer,
resources.RemainderDeltaBuffer, resources.ExponentBuffer,
resources.RemainderBuffer, resources.HitsBuffer,
resources.RemainderDeltaKernel, resources.DivisorDeltaBuffer,
resources.RemainderScanKernel, resources.RemainderDeltaBuffer,
resources.CandidateMaskKernel, resources.RemainderBuffer,
resources.Divisors, resources.RemainderDeltaKernel,
resources.Exponents, resources.RemainderScanKernel,
resources.Hits, resources.CandidateMaskKernel,
resources.DivisorData, resources.Divisors,
resources.CycleCandidates, resources.Exponents,
resources.CycleLengths, resources.Hits,
resources.CycleCapacity, resources.DivisorData,
out lastProcessed, resources.CycleCandidates,
out coveredRange, resources.CycleLengths,
out processedCount resources.CycleCapacity,
); out lastProcessed,
} out coveredRange,
finally out processedCount
{ );
ReturnBatchResources(accelerator, resources);
gpuLease.Dispose(); ReturnBatchResources(accelerator, resources);
} Monitor.Exit(gpuLease.ExecutionLock);
gpuLease.Dispose();
if (composite) if (composite)
{ {
@@ -540,7 +540,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
private readonly MersenneNumberDivisorByDivisorGpuTester _owner; private readonly MersenneNumberDivisorByDivisorGpuTester _owner;
private readonly GpuContextPool.GpuContextLease _lease; private readonly GpuContextPool.GpuContextLease _lease;
private readonly Accelerator _accelerator; private readonly Accelerator _accelerator;
private readonly Action<Index1D, MontgomeryDivisorData, ArrayView<ulong>, ArrayView<ulong>> _kernel; private Action<Index1D, MontgomeryDivisorData, ArrayView<ulong>, ArrayView<ulong>> _kernel;
private MemoryBuffer1D<ulong, Stride1D.Dense> _exponentsBuffer; private MemoryBuffer1D<ulong, Stride1D.Dense> _exponentsBuffer;
private MemoryBuffer1D<ulong, Stride1D.Dense> _resultsBuffer; private MemoryBuffer1D<ulong, Stride1D.Dense> _resultsBuffer;
private ulong[] _hostBuffer; private ulong[] _hostBuffer;
@@ -552,11 +552,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
_owner = owner; _owner = owner;
_lease = GpuContextPool.RentPreferred(preferCpu: false); _lease = GpuContextPool.RentPreferred(preferCpu: false);
_accelerator = _lease.Accelerator; _accelerator = _lease.Accelerator;
_kernel = owner.GetKernelByPrimeExponent(_accelerator);
_capacity = Math.Max(1, owner._gpuBatchSize); _capacity = Math.Max(1, owner._gpuBatchSize);
_exponentsBuffer = _accelerator.Allocate1D<ulong>(_capacity);
_resultsBuffer = _accelerator.Allocate1D<ulong>(_capacity);
_hostBuffer = ArrayPool<ulong>.Shared.Rent(_capacity);
} }
internal void Reset() internal void Reset()
@@ -564,6 +560,14 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
_disposed = false; _disposed = false;
} }
private void EnsureExecutionResourcesLocked()
{
_kernel = _owner.GetKernelByPrimeExponent(_accelerator);
_exponentsBuffer = _accelerator.Allocate1D<ulong>(_capacity);
_resultsBuffer = _accelerator.Allocate1D<ulong>(_capacity);
_hostBuffer = ArrayPool<ulong>.Shared.Rent(_capacity);
}
public void CheckDivisor(ulong divisor, in MontgomeryDivisorData divisorData, ulong divisorCycle, in ReadOnlySpan<ulong> primes, Span<byte> hits) public void CheckDivisor(ulong divisor, in MontgomeryDivisorData divisorData, ulong divisorCycle, in ReadOnlySpan<ulong> primes, Span<byte> hits)
{ {
if (_disposed) if (_disposed)
@@ -583,6 +587,13 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
return; return;
} }
Monitor.Enter(_lease.ExecutionLock);
if (_kernel == null)
{
EnsureExecutionResourcesLocked();
}
ArrayView1D<ulong, Stride1D.Dense> exponentsView = _exponentsBuffer.View; ArrayView1D<ulong, Stride1D.Dense> exponentsView = _exponentsBuffer.View;
ArrayView1D<ulong, Stride1D.Dense> resultsView = _resultsBuffer.View; ArrayView1D<ulong, Stride1D.Dense> resultsView = _resultsBuffer.View;
Span<ulong> hostSpan = _hostBuffer.AsSpan(0, _capacity); Span<ulong> hostSpan = _hostBuffer.AsSpan(0, _capacity);
@@ -609,6 +620,8 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
offset += batchSize; offset += batchSize;
} }
Monitor.Exit(_lease.ExecutionLock);
} }
public void Dispose() public void Dispose()