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

GpuContextPool removed, because we share context and accelerator between threads, using streams to separate.

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-07 12:52:20 +01:00
1 parent 7c035239b4
commit f4e365194d
19 files changed
+92 -232
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/DivisorCycleCache.cs
+1 -1
View File
@@ -269,7 +269,7 @@ public sealed class DivisorCycleCache
int length = divisors.Length;
var gpuLease = GpuKernelPool.Rent();
Accelerator accelerator = gpuLease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
var stream = gpuLease.Stream;
var kernel = _gpuKernel;
EnsureCapacity(length);
PerfectNumbers.Core/Gpu/GpuContextPool.cs
-107
View File
@@ -1,107 +0,0 @@
using System.Collections.Concurrent;
using ILGPU;
using ILGPU.Runtime;
namespace PerfectNumbers.Core.Gpu;
public sealed class GpuContextLease
{
private GpuContextPool.PooledContext? _ctx;
internal GpuContextLease(GpuContextPool.PooledContext ctx)
{
_ctx = ctx;
}
public Context Context => _ctx!.Context;
public Accelerator Accelerator => _ctx!.Accelerator;
public void Dispose() => GpuContextPool.Return(_ctx!);
}
public static class GpuContextPool
{
// The pool intentionally skips pre-loading ProcessEightBitWindows kernels so accelerator initialization
// stays stable; GpuKernelPool JIT-compiles them on first use alongside the small-cycle uploads.
// Default device preference for generic GPU kernels (prime scans, NTT, etc.)
internal sealed class PooledContext
{
public readonly Context Context;
public readonly Accelerator Accelerator;
public PooledContext()
{
Context = SharedGpuContext.Context;
Accelerator = SharedGpuContext.Accelerator;
// NOTE: Avoid loading/compiling any kernel here to prevent implicit
// CL stream/queue creation during accelerator construction.
// Some OpenCL drivers are fragile when a queue is created immediately
// after device init. Real kernels will be JIT-loaded at first use.
}
public void Dispose()
{
// Ensure all cached GPU buffers for this accelerator are released.
// Release PrimeTester GPU state for this accelerator (kernel/device primes)
// Clear any per-accelerator cached resources to avoid releasing
// after the accelerator is destroyed.
// These resources are shared between GPU leases
// Accelerator.Dispose();
// Context.Dispose();
}
}
private static readonly ConcurrentQueue<PooledContext> GpuPool = new();
private static int WarmedGpuContextCount;
// Allows callers to choose CPU/GPU per use-case, decoupled from ForceCpu.
public static GpuContextLease Rent()
{
if (GpuPool.TryDequeue(out var gpu))
{
return new GpuContextLease(gpu);
}
// Create a new accelerator when the pool does not have one available.
return new GpuContextLease(new PooledContext());
}
public static void WarmUpPool(int threadCount)
{
if (threadCount <= WarmedGpuContextCount)
{
return;
}
int toCreate = threadCount - WarmedGpuContextCount;
Console.WriteLine($"Warming up GPU context pool to {toCreate} contexts...");
var contexts = new PooledContext[toCreate];
var pool = GpuPool;
PooledContext context;
for (int i = 0; i < toCreate; i++)
{
context = new PooledContext();
contexts[i] = context;
pool.Enqueue(context);
}
WarmedGpuContextCount = threadCount;
}
public static void DisposeAll()
{
while (GpuPool.TryDequeue(out var gpu))
{
gpu.Dispose();
}
WarmedGpuContextCount = 0;
}
internal static void Return(PooledContext ctx)
{
GpuPool.Enqueue(ctx);
}
}
PerfectNumbers.Core/Gpu/GpuKernelLease.cs
+8 -22
View File
@@ -1,6 +1,5 @@
using System.Collections.Concurrent;
using ILGPU;
using ILGPU.Algorithms;
using ILGPU.Runtime;
namespace PerfectNumbers.Core.Gpu;
@@ -9,7 +8,6 @@ public sealed class GpuKernelLease
{
private static readonly ConcurrentQueue<GpuKernelLease> Pool = new();
private GpuContextLease _gpu;
public KernelContainer Kernels;
public AcceleratorStream Stream;
@@ -20,29 +18,25 @@ public sealed class GpuKernelLease
Kernels = kernels;
}
internal static GpuKernelLease Rent(GpuContextLease gpu, AcceleratorStream stream, KernelContainer kernels)
internal static GpuKernelLease Rent(AcceleratorStream stream, KernelContainer kernels)
{
if (!Pool.TryDequeue(out var lease))
{
lease = new GpuKernelLease(kernels);
}
lease._gpu = gpu;
lease.Kernels = kernels;
lease.Stream = stream;
return lease;
}
public Accelerator Accelerator => _gpu?.Accelerator ?? throw new NullReferenceException("GPU context is already released");
public Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>> OrderKernel
{
get
{
var accel = Accelerator; // avoid capturing 'this' in lambda
return GpuKernelPool.InitOnce(ref Kernels!.Order, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>(OrderKernels.OrderKernelScan);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ulong, ArrayView<GpuUInt128>, ArrayView<ulong>>>();
});
@@ -53,10 +47,9 @@ public sealed class GpuKernelLease
{
get
{
var accel = Accelerator;
return GpuKernelPool.InitOnce(ref Kernels!.Pow2Mod, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<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>>(Pow2ModKernels.Pow2ModKernelScan);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<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>>(Pow2ModKernels.Pow2ModKernelScan);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<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>>>();
});
@@ -67,10 +60,9 @@ public sealed class GpuKernelLease
{
get
{
var accel = Accelerator; // avoid capturing 'this' in lambda
return GpuKernelPool.InitOnce(ref Kernels!.Incremental, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalKernelScan);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ulong, ulong, ulong, ulong, ArrayView<ulong>, ArrayView1D<ulong, Stride1D.Dense>>>();
});
@@ -81,10 +73,9 @@ public sealed class GpuKernelLease
{
get
{
var accel = Accelerator;
return GpuKernelPool.InitOnce(ref Kernels!.IncrementalOrder, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(IncrementalKernels.IncrementalOrderKernelScan);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
});
@@ -95,10 +86,9 @@ public sealed class GpuKernelLease
{
get
{
var accel = Accelerator;
return GpuKernelPool.InitOnce(ref Kernels!.Pow2ModOrder, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernels.Pow2ModOrderKernelScan);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, GpuUInt128, GpuUInt128, byte, ulong, ResidueAutomatonArgs, ArrayView<int>, ArrayView1D<ulong, Stride1D.Dense>>>();
});
@@ -109,10 +99,9 @@ public sealed class GpuKernelLease
{
get
{
var accel = Accelerator;
return GpuKernelPool.InitOnce(ref Kernels!.SpecialMax, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
});
@@ -123,10 +112,9 @@ public sealed class GpuKernelLease
{
get
{
var accel = Accelerator;
return GpuKernelPool.InitOnce(ref Kernels!.SmallPrimeFactor, () =>
{
var loaded = accel.LoadAutoGroupedStreamKernel<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>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan);
var loaded = SharedGpuContext.Accelerator.LoadAutoGroupedStreamKernel<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>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan);
var kernel = KernelUtil.GetKernel(loaded);
return kernel.CreateLauncherDelegate<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>>>();
});
@@ -136,9 +124,7 @@ public sealed class GpuKernelLease
public void Dispose()
{
Stream.Dispose();
_gpu.Dispose();
Pool.Enqueue(this);
GpuPrimeWorkLimiter.Release();
}
}
PerfectNumbers.Core/Gpu/GpuKernelPool.cs
+45 -24
View File
@@ -1,13 +1,12 @@
using System.Collections.Concurrent;
using ILGPU;
using ILGPU.Runtime;
using static PerfectNumbers.Core.Gpu.GpuContextPool;
namespace PerfectNumbers.Core.Gpu;
public class GpuKernelPool
{
private static readonly ConcurrentDictionary<Accelerator, KernelContainer> _kernels = new();
public static readonly KernelContainer Kernels = CreateKernels();
public static T InitOnce<T>(ref T? target, Func<T> valueFactory) where T : class
{
@@ -20,26 +19,35 @@ public class GpuKernelPool
return Interlocked.CompareExchange(ref target, newValue, null) ?? newValue;
}
internal static KernelContainer GetKernels(Accelerator accelerator, AcceleratorStream stream)
{
return _kernels.GetOrAdd(accelerator, (accelerator) =>
private static KernelContainer CreateKernels()
{
var kernels = new KernelContainer();
PreloadStaticTables(kernels, accelerator, stream);
var stream = SharedGpuContext.Accelerator.CreateStream();
PreloadStaticTables(kernels, stream);
stream.Synchronize();
stream.Dispose();
return kernels;
});
}
internal static void PreloadStaticTables(KernelContainer kernels, Accelerator accelerator, AcceleratorStream stream)
private static KernelContainer CreateKernels(AcceleratorStream stream)
{
EnsureSmallCyclesOnDevice(kernels, accelerator, stream);
EnsureSmallPrimesOnDevice(kernels, accelerator, stream);
EnsureSmallPrimeFactorTables(kernels, accelerator, stream);
KernelContainer kernels = new();
PreloadStaticTables(kernels, stream);
return Kernels;
}
internal static void PreloadStaticTables(KernelContainer kernels, AcceleratorStream stream)
{
_ = EnsureSmallCyclesOnDevice(kernels, stream);
_ = EnsureSmallPrimesOnDevice(kernels, stream);
_ = EnsureSmallPrimeFactorTables(kernels, stream);
}
public static ArrayView1D<ulong, Stride1D.Dense> GetSmallCyclesOnDevice() => Kernels.SmallCycles!;
// Ensures the small cycles table is uploaded to the device for the given accelerator.
// Returns the ArrayView to pass into kernels (when kernels are extended to accept it).
public static ArrayView1D<ulong, Stride1D.Dense> EnsureSmallCyclesOnDevice(KernelContainer kernels, Accelerator accelerator, AcceleratorStream stream)
private static ArrayView1D<ulong, Stride1D.Dense> EnsureSmallCyclesOnDevice(KernelContainer kernels, AcceleratorStream stream)
{
if (kernels.SmallCycles is { } buffer)
{
@@ -47,13 +55,20 @@ public class GpuKernelPool
}
var host = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot(); // TODO: Preload this device buffer during startup and keep it immutable so we can delete the lock above in favor of the preloaded snapshot.
var device = accelerator.Allocate1D<ulong>(host.Length);
var device = SharedGpuContext.Accelerator.Allocate1D<ulong>(host.Length);
device.View.CopyFromCPU(stream, host);
kernels.SmallCycles = device;
return device.View;
}
public static ResiduePrimeViews EnsureSmallPrimesOnDevice(KernelContainer kernels, Accelerator accelerator, AcceleratorStream stream)
public static ResiduePrimeViews GetSmallPrimesOnDevice() => new(
Kernels.SmallPrimesLastOne!,
Kernels.SmallPrimesLastSeven!,
Kernels.SmallPrimesPow2LastOne!,
Kernels.SmallPrimesPow2LastSeven!
);
private static ResiduePrimeViews EnsureSmallPrimesOnDevice(KernelContainer kernels, AcceleratorStream stream)
{
if (kernels.SmallPrimesLastOne is { } lastOne &&
kernels.SmallPrimesLastSeven is { } lastSeven &&
@@ -71,6 +86,7 @@ public class GpuKernelPool
return new ResiduePrimeViews(existingLastOne.View, existingLastSeven.View, existingLastOnePow2.View, existingLastSevenPow2.View);
}
Accelerator accelerator = SharedGpuContext.Accelerator;
var hostLastOne = PrimesGenerator.SmallPrimesLastOne;
var hostLastSeven = PrimesGenerator.SmallPrimesLastSeven;
var hostLastOnePow2 = PrimesGenerator.SmallPrimesPow2LastOne;
@@ -93,34 +109,39 @@ public class GpuKernelPool
return new ResiduePrimeViews(deviceLastOne.View, deviceLastSeven.View, deviceLastOnePow2.View, deviceLastSevenPow2.View);
}
public static SmallPrimeFactorViews EnsureSmallPrimeFactorTables(KernelContainer kernels, Accelerator accelerator, AcceleratorStream stream)
public static SmallPrimeFactorViews GetSmallPrimeFactorTables() => new(
Kernels.SmallPrimeFactorsPrimes!,
Kernels.SmallPrimeFactorsSquares!
);
private static SmallPrimeFactorViews EnsureSmallPrimeFactorTables(KernelContainer kernels, AcceleratorStream stream)
{
if (kernels.SmallPrimeFactorsPrimes is { } primeBuffer && kernels.SmallPrimeFactorsSquares is { } squareBuffer)
{
return new SmallPrimeFactorViews(primeBuffer, squareBuffer, (int)primeBuffer.Length);
return new SmallPrimeFactorViews(primeBuffer, squareBuffer);
}
Accelerator accelerator = SharedGpuContext.Accelerator;
var hostPrimes = PrimesGenerator.SmallPrimes;
var hostSquares = PrimesGenerator.SmallPrimesPow2;
var devicePrimes = accelerator.Allocate1D<uint>(hostPrimes.Length);
int length = hostPrimes.Length;
var devicePrimes = accelerator.Allocate1D<uint>(length);
devicePrimes.View.CopyFromCPU(stream, hostPrimes);
var deviceSquares = accelerator.Allocate1D<ulong>(hostSquares.Length);
var deviceSquares = accelerator.Allocate1D<ulong>(length);
deviceSquares.View.CopyFromCPU(stream, hostSquares);
kernels.SmallPrimeFactorsPrimes = devicePrimes;
kernels.SmallPrimeFactorsSquares = deviceSquares;
return new SmallPrimeFactorViews(devicePrimes, deviceSquares, hostPrimes.Length);
return new SmallPrimeFactorViews(devicePrimes, deviceSquares);
}
public static GpuKernelLease Rent()
{
GpuPrimeWorkLimiter.Acquire();
var gpu = GpuContextPool.Rent();
var stream = gpu.Accelerator.CreateStream();
var kernels = GetKernels(gpu.Accelerator, stream);
return GpuKernelLease.Rent(gpu, stream, kernels);
var stream = SharedGpuContext.Accelerator.CreateStream();
return GpuKernelLease.Rent(stream, Kernels);
}
/// <summary>
@@ -130,7 +151,7 @@ public class GpuKernelPool
public static void Run(Action<Accelerator, AcceleratorStream> action)
{
var lease = Rent();
var accelerator = lease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = lease.Stream;
action(accelerator, stream);
lease.Dispose();
PerfectNumbers.Core/Gpu/GpuStaticTableInitializer.cs
+1 -1
View File
@@ -16,7 +16,7 @@ internal static class GpuStaticTableInitializer
Initialized.GetOrAdd(accelerator, acc =>
{
PrimeTester.PrimeTesterGpuContextPool.EnsureStaticTables(kernels, acc, stream);
GpuKernelPool.PreloadStaticTables(kernels, acc, stream);
GpuKernelPool.PreloadStaticTables(kernels, stream);
PrimeOrderGpuHeuristics.PreloadStaticTables( acc, stream);
return 0;
});
PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs
+2 -2
View File
@@ -96,7 +96,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
ulong lastProcessed;
var gpuLease = GpuKernelPool.Rent();
var accelerator = gpuLease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = gpuLease.Stream;
// Monitor.Enter(gpuLease.ExecutionLock);
@@ -502,7 +502,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
{
_owner = owner;
_lease = GpuKernelPool.Rent();
_accelerator = _lease.Accelerator;
_accelerator = SharedGpuContext.Accelerator;
_stream = _lease.Stream;
_capacity = Math.Max(1, owner._gpuBatchSize);
}
PerfectNumbers.Core/Gpu/MersenneNumberDivisorGpuTester.cs
+2 -4
View File
@@ -43,8 +43,7 @@ public sealed class MersenneNumberDivisorGpuTester
public bool IsDivisible(ulong exponent, in ReadOnlyGpuUInt128 divisor)
{
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var kernel = GetKernel(accelerator);
var resultBuffer = _resultBuffers.GetOrAdd(accelerator, acc => acc.Allocate1D<byte>(1));
@@ -56,11 +55,10 @@ public sealed class MersenneNumberDivisorGpuTester
stream.Synchronize();
bool divisible = result[0] != 0;
stream.Dispose();
gpu.Dispose();
return divisible;
}
private static (ulong divisor, uint cycle)[]? _divisorCandidates = Array.Empty<(ulong divisor, uint cycle)>();
private static (ulong divisor, uint cycle)[]? _divisorCandidates = [];
public bool IsPrime(ulong p, UInt128 d, ulong divisorCyclesSearchLimit, out bool divisorsExhausted)
{
PerfectNumbers.Core/Gpu/MersenneNumberIncrementalGpuTester.cs
+3 -3
View File
@@ -17,7 +17,7 @@ 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();
var accelerator = gpuLease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = gpuLease.Stream;
int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy
UInt128 kStart = 1UL;
@@ -31,11 +31,11 @@ public class MersenneNumberIncrementalGpuTester(GpuKernelType kernelType, bool u
ulong step3 = ((exponent % 3UL) << 1) % 3UL;
ulong step5 = ((exponent % 5UL) << 1) % 5UL;
GpuUInt128 twoPGpu = (GpuUInt128)twoP;
var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(gpuLease.Kernels, accelerator, stream);
var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice();
ResiduePrimeViews primeViews = default;
if (_kernelType == GpuKernelType.Pow2Mod)
{
primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(gpuLease.Kernels, accelerator, stream);
primeViews = GpuKernelPool.GetSmallPrimesOnDevice();
}
var orderBuffer = accelerator.Allocate1D<ulong>(batchSize);
PerfectNumbers.Core/Gpu/MersenneNumberLucasLehmerGpuTester.cs
+4 -9
View File
@@ -97,8 +97,7 @@ public class MersenneNumberLucasLehmerGpuTester
}
else
{
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var kernel = GetKernel(accelerator);
var modulus = new GpuUInt128(((UInt128)1 << (int)exponent) - 1UL); // TODO: Cache these Mersenne moduli per exponent so LL GPU runs skip rebuilding them every launch.
@@ -109,7 +108,6 @@ public class MersenneNumberLucasLehmerGpuTester
result = buffer.GetAsArray1D()[0].IsZero;
buffer.Dispose();
stream.Dispose();
gpu.Dispose();
}
GpuPrimeWorkLimiter.Release();
@@ -132,8 +130,7 @@ public class MersenneNumberLucasLehmerGpuTester
}
}
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var kernel = GetBatchKernel(accelerator);
@@ -166,7 +163,6 @@ public class MersenneNumberLucasLehmerGpuTester
gpuUInt128Pool.Return(modulusArray);
ulongPool.Return(expArray);
stream.Dispose();
gpu.Dispose();
}
public void IsMersennePrimeBatch(ReadOnlySpan<ulong> exponents, Span<bool> results)
@@ -189,9 +185,9 @@ public class MersenneNumberLucasLehmerGpuTester
private bool IsMersennePrimeNtt(ulong exponent, GpuUInt128 nttMod, GpuUInt128 primitiveRoot, bool runOnGpu)
{
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var addKernel = GetAddSmallKernel(accelerator);
var subKernel = GetSubSmallKernel(accelerator);
var reduceKernel = GetReduceKernel(accelerator);
@@ -240,7 +236,6 @@ public class MersenneNumberLucasLehmerGpuTester
resultBuffer.Dispose();
stateBuffer.Dispose();
stream.Dispose();
gpu.Dispose();
return isPrime;
}
PerfectNumbers.Core/Gpu/MersenneNumberOrderGpuTester.cs
+2 -2
View File
@@ -10,7 +10,7 @@ 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();
var accelerator = gpuLease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = gpuLease.Stream;
int batchSize = GpuConstants.ScanBatchSize; // large batch improves GPU occupancy and avoids TDR
UInt128 kStart = UInt128.One;
@@ -30,7 +30,7 @@ public class MersenneNumberOrderGpuTester(GpuKernelType kernelType, bool useGpuO
int currentSize;
int found = 0;
var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(gpuLease.Kernels, accelerator, stream);
var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice();
while (kStart <= maxK && Volatile.Read(ref isPrime))
{
remaining = maxK - kStart + 1UL;
PerfectNumbers.Core/Gpu/MersenneNumberResidueGpuTester.cs
+4 -19
View File
@@ -14,7 +14,6 @@ namespace PerfectNumbers.Core.Gpu;
public class MersenneNumberResidueGpuTester(bool useGpuOrder)
{
private readonly bool _useGpuOrder = useGpuOrder;
private readonly ConcurrentBag<GpuContextLease> _acceleratorPool = new();
private readonly ConcurrentDictionary<Accelerator, 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>>> _pow2ModKernelCache = new(AcceleratorReferenceComparer.Instance);
private readonly ConcurrentDictionary<Accelerator, ConcurrentBag<ResidueResources>> _resourcePools = new(AcceleratorReferenceComparer.Instance);
@@ -28,15 +27,14 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
public void Scan(ulong exponent, UInt128 twoP, LastDigit lastDigit, UInt128 maxK, ref bool isPrime)
{
GpuPrimeWorkLimiter.Acquire();
var gpuLease = RentAccelerator();
var accelerator = gpuLease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
// Monitor.Enter(gpuLease.ExecutionLock);
var stream = accelerator.CreateStream();
var kernel = GetPow2ModKernel(accelerator);
var resources = RentResources(accelerator, GpuConstants.ScanBatchSize);
var gpuKernels = GpuKernelPool.GetKernels(accelerator, stream);
var gpuKernels = GpuKernelPool.Kernels;
var orderBuffer = resources.OrderBuffer;
ulong[] orderArray = resources.OrderArray;
@@ -44,8 +42,8 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
// Ensure device has small cycles and primes tables for in-kernel lookup
// TODO: Pass-in the stream down the path
var smallCyclesView = GpuKernelPool.EnsureSmallCyclesOnDevice(gpuKernels, accelerator, stream);
ResiduePrimeViews primeViews = GpuKernelPool.EnsureSmallPrimesOnDevice(gpuKernels, accelerator, stream);
var smallCyclesView = GpuKernelPool.GetSmallCyclesOnDevice();
ResiduePrimeViews primeViews = GpuKernelPool.GetSmallPrimesOnDevice();
UInt128 kStart = 1UL;
UInt128 limit = maxK + UInt128.One;
@@ -114,7 +112,6 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
stream.Dispose();
// Monitor.Exit(gpuLease.ExecutionLock);
gpuKernels.Dispose();
ReturnAccelerator(gpuLease);
GpuPrimeWorkLimiter.Release();
}
@@ -130,18 +127,6 @@ public class MersenneNumberResidueGpuTester(bool useGpuOrder)
loaded.Method);
});
private GpuContextLease RentAccelerator()
{
if (_acceleratorPool.TryTake(out var lease))
{
return lease;
}
return GpuContextPool.Rent();
}
private void ReturnAccelerator(GpuContextLease lease) => _acceleratorPool.Add(lease);
private ResidueResources RentResources(Accelerator accelerator, int capacity)
{
var bag = _resourcePools.GetOrAdd(accelerator, static _ => []);
PerfectNumbers.Core/Gpu/NttGpuMath.cs
+6 -23
View File
@@ -504,8 +504,7 @@ public static class NttGpuMath
{
int n = values.Length;
int bits = (int)Math.Log2(n);
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var kernel = GetBitReverseKernel(accelerator);
var pool = ThreadStaticPools.GpuUInt128Pool;
var array = pool.Rent(n);
@@ -520,7 +519,6 @@ public static class NttGpuMath
pool.Return(array, clearArray: true);
buffer.Dispose();
stream.Dispose();
gpu.Dispose();
}
@@ -568,8 +566,7 @@ public static class NttGpuMath
private static void ForwardGpuReference(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
{
int n = (int)values.Length;
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var kernel = GetForwardKernel(accelerator);
var pool = ThreadStaticPools.GpuUInt128Pool;
@@ -592,8 +589,6 @@ public static class NttGpuMath
outputBuffer.Dispose();
inputBuffer.Dispose();
stream.Dispose();
gpu.Dispose();
}
public static void Inverse(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
@@ -667,8 +662,7 @@ public static class NttGpuMath
private static void InverseGpuReference(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
{
int n = values.Length;
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var kernel = GetInverseKernel(accelerator);
var pool = ThreadStaticPools.GpuUInt128Pool;
@@ -710,15 +704,12 @@ public static class NttGpuMath
outputBuffer.Dispose();
inputBuffer.Dispose();
stream.Dispose();
gpu.Dispose();
}
private static void ForwardGpuStaged(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
{
int n = values.Length;
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var buffer = accelerator.Allocate1D<GpuUInt128>(n);
buffer.View.CopyFromCPU(stream, ref values[0], n);
@@ -728,15 +719,12 @@ public static class NttGpuMath
stream.Synchronize();
stream.Dispose();
buffer.Dispose();
gpu.Dispose();
}
private static void InverseGpuStaged(Span<GpuUInt128> values, GpuUInt128 modulus, GpuUInt128 primitiveRoot)
{
int n = values.Length;
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var buffer = accelerator.Allocate1D<GpuUInt128>(n);
buffer.View.CopyFromCPU(stream, ref values[0], n);
@@ -746,8 +734,6 @@ public static class NttGpuMath
stream.Synchronize();
stream.Dispose();
buffer.Dispose();
gpu.Dispose();
}
private static void ForwardDevice(Accelerator accelerator, AcceleratorStream stream, ArrayView<GpuUInt128> data, int length, GpuUInt128 modulus, SquareCacheEntry cache)
@@ -939,8 +925,7 @@ public static class NttGpuMath
public static void PointwiseMultiply(Span<GpuUInt128> left, Span<GpuUInt128> right, GpuUInt128 modulus)
{
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = accelerator.CreateStream();
var kernel = GetMulKernel(accelerator);
var pool = ThreadStaticPools.GpuUInt128Pool;
@@ -963,8 +948,6 @@ public static class NttGpuMath
rightBuffer.Dispose();
leftBuffer.Dispose();
stream.Dispose();
gpu.Dispose();
}
private static ulong[] GenerateSmallPrimes(int limit)
PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs
+4 -4
View File
@@ -199,7 +199,7 @@ internal static partial class PrimeOrderGpuHeuristics
try
{
var lease = GpuKernelPool.Rent();
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = lease.Stream;
var kernel = GetPartialFactorKernel(accelerator);
@@ -366,7 +366,7 @@ internal static partial class PrimeOrderGpuHeuristics
order = 0UL;
var lease = GpuKernelPool.Rent();
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = lease.Stream;
var kernel = GetOrderKernel(accelerator);
@@ -527,7 +527,7 @@ internal static partial class PrimeOrderGpuHeuristics
private static bool TryComputeOnGpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
{
var lease = GpuKernelPool.Rent();
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = lease.Stream;
var kernel = GetPow2ModKernel(accelerator);
// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
@@ -586,7 +586,7 @@ internal static partial class PrimeOrderGpuHeuristics
try
{
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = lease.Stream;
var kernel = GetPow2ModWideKernel(accelerator);
// TODO: These buffers should be allocated once per accelerator and only reallocated when the new length exceeds capacity.
PerfectNumbers.Core/Gpu/SmallPrimeFactorViews.cs
+2 -3
View File
@@ -6,12 +6,11 @@ namespace PerfectNumbers.Core.Gpu;
public readonly struct SmallPrimeFactorViews(
MemoryBuffer1D<uint, Stride1D.Dense> primes,
MemoryBuffer1D<ulong, Stride1D.Dense> squares,
int count)
MemoryBuffer1D<ulong, Stride1D.Dense> squares)
{
public readonly MemoryBuffer1D<uint, Stride1D.Dense> Primes = primes;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Squares = squares;
public readonly int Count = count;
public readonly int Count = (int)primes.Length;
public ArrayView1D<uint, Stride1D.Dense> PrimesView
{
PerfectNumbers.Core/MersenneDivisorCycles.cs
+1 -1
View File
@@ -805,7 +805,7 @@ public class MersenneDivisorCycles
divisors = pool.Rent(count);
outCycles = pool.Rent(count);
var lease = GpuKernelPool.Rent();
var accelerator = lease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = lease.Stream;
var loaded = accelerator.LoadAutoGroupedStreamKernel<
Index1D,
PerfectNumbers.Core/MersenneNumberTester.cs
+1 -1
View File
@@ -148,7 +148,7 @@ public sealed class MersenneNumberTester(
UInt128[] qs = qsBuffer;
var gpuLease = GpuKernelPool.Rent();
var accelerator = gpuLease.Accelerator;
var accelerator = SharedGpuContext.Accelerator;
var stream = gpuLease.Stream;
var orderKernel = gpuLease.OrderKernel!;
PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs
+3 -3
View File
@@ -20,13 +20,13 @@ internal static partial class PrimeOrderCalculator
remaining = value;
var lease = GpuKernelPool.Rent();
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
var stream = lease.Stream;
KernelContainer kernels = lease.Kernels;
ScratchBuffer scratch = GpuScratchBufferPool.Rent(accelerator, GpuSmallPrimeFactorSlots, 0);
SmallPrimeFactorViews tables = GpuKernelPool.EnsureSmallPrimeFactorTables(kernels, accelerator, stream);
SmallPrimeFactorViews tables = GpuKernelPool.GetSmallPrimeFactorTables();
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");
@@ -101,7 +101,7 @@ internal static partial class PrimeOrderCalculator
}
GpuKernelLease lease = GpuKernelPool.Rent();
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
var stream = lease.Stream;
PerfectNumbers.Core/PrimeTester.cs
+2 -2
View File
@@ -273,7 +273,7 @@ public sealed class PrimeTester
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = accelerator.CreateStream();
KernelContainer kernels = GpuKernelPool.GetKernels(accelerator, stream);
KernelContainer kernels = GpuKernelPool.Kernels;
GpuStaticTableInitializer.EnsureStaticTables(kernels, accelerator, stream);
@@ -433,7 +433,7 @@ public sealed class PrimeTester
AcceleratorContext = SharedGpuContext.Context;
Accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = SharedGpuContext.Accelerator.CreateStream();
var kernels = GpuKernelPool.GetKernels(Accelerator, stream);
var kernels = GpuKernelPool.Kernels;
GpuStaticTableInitializer.EnsureStaticTables(kernels, Accelerator, stream);
_sharedTables = RentSharedTables(Accelerator, stream);
PerfectNumbers.Core/ULongExtensions.Gpu.cs
+1 -1
View File
@@ -463,7 +463,7 @@ public static partial class ULongExtensions
GpuKernelLease lease = GpuKernelPool.Rent();
Accelerator accelerator = lease.Accelerator;
Accelerator accelerator = SharedGpuContext.Accelerator;
AcceleratorStream stream = lease.Stream;
var kernelGroup = Kernel;