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

Throttle GPU cycle generation batches

Browse Source
This commit is contained in:
Marek Lemańczyk committed 2025-09-25 02:48:14 +02:00
1 parent 211df684ce
commit ad3effa5fa
1 file changed
+24 -8
Download Patch File Download Diff File Expand all files Collapse all files
PerfectNumbers.Core/DivisorCycleCache.cs
+24 -8
View File
@@ -2,6 +2,7 @@ using System;
using System.Buffers; 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.Threading.Tasks; using System.Threading.Tasks;
using ILGPU; using ILGPU;
using ILGPU.Runtime; using ILGPU.Runtime;
@@ -67,6 +68,8 @@ public sealed class DivisorCycleCache
} }
} }
private const int CycleGenerationBatchSize = 262_144;
private readonly object _sync = new(); private readonly object _sync = new();
private readonly ConcurrentDictionary<int, Task<CycleBlock>> _pending = new(); private readonly ConcurrentDictionary<int, Task<CycleBlock>> _pending = new();
private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>> _gpuKernelCache = new(AcceleratorReferenceComparer.Instance); private readonly ConcurrentDictionary<Accelerator, Action<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>> _gpuKernelCache = new(AcceleratorReferenceComparer.Instance);
@@ -328,22 +331,35 @@ public sealed class DivisorCycleCache
Action<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> kernel = _gpuKernelCache.GetOrAdd(accelerator, LoadKernel); Action<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> kernel = _gpuKernelCache.GetOrAdd(accelerator, LoadKernel);
int length = destination.Length; int length = destination.Length;
ulong[] divisors = ArrayPool<ulong>.Shared.Rent(length); int chunkCapacity = Math.Min(length, CycleGenerationBatchSize);
ulong[] divisors = ArrayPool<ulong>.Shared.Rent(chunkCapacity);
try try
{ {
for (int i = 0; i < length; i++) using MemoryBuffer1D<ulong, Stride1D.Dense> divisorBuffer = accelerator.Allocate1D<ulong>(chunkCapacity);
using MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = accelerator.Allocate1D<ulong>(chunkCapacity);
int offset = 0;
while (offset < length)
{ {
divisors[i] = start + (ulong)i; int batchSize = Math.Min(chunkCapacity, length - offset);
Span<ulong> divisorsSpan = divisors.AsSpan(0, batchSize);
for (int i = 0; i < batchSize; i++)
{
divisorsSpan[i] = start + (ulong)(offset + i);
} }
using MemoryBuffer1D<ulong, Stride1D.Dense> divisorBuffer = accelerator.Allocate1D<ulong>(length); ArrayView1D<ulong, Stride1D.Dense> divisorView = divisorBuffer.View.SubView(0, batchSize);
using MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = accelerator.Allocate1D<ulong>(length); ArrayView1D<ulong, Stride1D.Dense> resultView = resultBuffer.View.SubView(0, batchSize);
divisorBuffer.View.CopyFromCPU(ref divisors[0], length); divisorView.CopyFromCPU(ref MemoryMarshal.GetReference(divisorsSpan), batchSize);
kernel(length, divisorBuffer.View, resultBuffer.View); kernel(batchSize, divisorView, resultView);
accelerator.Synchronize(); accelerator.Synchronize();
resultBuffer.View.CopyToCPU(ref destination[0], length);
Span<ulong> destinationSpan = destination.AsSpan(offset, batchSize);
resultView.CopyToCPU(ref MemoryMarshal.GetReference(destinationSpan), batchSize);
offset += batchSize;
}
} }
finally finally
{ {