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

PrimeTesterByLastDigitAccelerator integrated into Pow2MontgomeryGpuCalculator for better performance.

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-11 13:36:21 +01:00
1 parent 0de9871529
commit cf5f78c2ef
11 files changed
+222 -241
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/Accelerators/Pow2MontgomeryAccelerator.cs
+121 -7
View File
@@ -12,20 +12,21 @@ public sealed class Pow2MontgomeryAccelerator
[ThreadStatic]
private static Queue<Pow2MontgomeryAccelerator>? _pool;
private static AcceleratorPool _accelerators = new(PerfectNumberConstants.RollingAccelerators);
private static AcceleratorPool _accelerators = new(PerfectNumberConstants.RollingAccelerators << 1);
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static Pow2MontgomeryAccelerator Rent()
public static Pow2MontgomeryAccelerator Rent(int primeTesterCapacity)
{
var pool = _pool ??= [];
if (!pool.TryDequeue(out var gpu))
{
Accelerator accelerator = _accelerators.Rent();
gpu = new(accelerator, PerfectNumberConstants.DefaultFactorsBuffer);
gpu = new(accelerator, PerfectNumberConstants.DefaultFactorsBuffer, primeTesterCapacity);
}
else
{
gpu.Stream = gpu.Accelerator.CreateStream();
gpu.EnsureCapacity(PerfectNumberConstants.DefaultFactorsBuffer,primeTesterCapacity);
}
return gpu;
@@ -39,13 +40,69 @@ public sealed class Pow2MontgomeryAccelerator
_pool!.Enqueue(gpu);
}
internal static void Clear()
{
Queue<Pow2MontgomeryAccelerator>? pool = _pool;
if (pool != null)
{
while (pool.TryDequeue(out var lease))
{
lease.Dispose();
}
}
}
internal static void DisposeAll()
{
Queue<Pow2MontgomeryAccelerator>? pool = _pool;
if (pool != null)
{
while (pool.TryDequeue(out var lease))
{
lease.Dispose();
}
}
}
internal static void WarmUp()
{
Accelerator[]? accelerators = _accelerators.Accelerators;
int acceleratorCount = accelerators.Length;
for (var i = 0; i < acceleratorCount; i++)
{
Console.WriteLine($"Preparing accelerator {i}...");
var accelerator = accelerators[i];
AcceleratorStream stream = accelerator.CreateStream();
LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
// _ = GpuKernelPool.GetOrAddKernels(accelerator, stream, KernelType.None);
// KernelContainer kernels = GpuKernelPool.GetOrAddKernels(accelerator, stream);
// GpuStaticTableInitializer.EnsureStaticTables(accelerator, kernels, stream);
stream.Synchronize();
stream.Dispose();
}
}
#endregion
public readonly Accelerator Accelerator;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Input;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> Output;
public MemoryBuffer1D<ulong, Stride1D.Dense> PrimeTestInput;
public MemoryBuffer1D<byte, Stride1D.Dense> PrimeTestOutput;
public MemoryBuffer1D<int, Stride1D.Dense> HeuristicFlag;
public MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> Pow2ModEntriesToTestOnDevice;
public Dictionary<ulong, int> Pow2ModEntriesToTestOnHost = [];
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree;
public readonly MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree;
public readonly MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine;
public readonly Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> SmallPrimeSieveKernel;
public AcceleratorStream? Stream;
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernel;
@@ -53,7 +110,23 @@ public sealed class Pow2MontgomeryAccelerator
public readonly Kernel CheckFactorsKernel;
public void EnsureCapacity(int factorsCount)
[ThreadStatic]
private static Dictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>? SmallPrimeSieveKernelPool;
internal static Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> GetSmallPrimeSieveKernel(Accelerator accelerator)
{
var pool = SmallPrimeSieveKernelPool ??= [];
if (pool.TryGetValue(accelerator, out var cached))
{
return cached;
}
cached = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>();
pool[accelerator] = cached;
return cached;
}
public void EnsureCapacity(int factorsCount, int primeTesterCapacity)
{
if (factorsCount > Pow2ModEntriesToTestOnDevice.Length)
{
@@ -61,16 +134,45 @@ public sealed class Pow2MontgomeryAccelerator
Pow2ModEntriesToTestOnDevice.Dispose();
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
}
if (primeTesterCapacity > PrimeTestInput.Length)
{
PrimeTestInput?.Dispose();
PrimeTestOutput?.Dispose();
HeuristicFlag?.Dispose();
// lock (Accelerator)
{
PrimeTestInput = Accelerator.Allocate1D<ulong>(primeTesterCapacity);
PrimeTestOutput = Accelerator.Allocate1D<byte>(primeTesterCapacity);
HeuristicFlag = Accelerator.Allocate1D<int>(primeTesterCapacity);
}
}
}
public Pow2MontgomeryAccelerator(Accelerator accelerator, int factorsCount)
public Pow2MontgomeryAccelerator(Accelerator accelerator, int factorsCount, int primeTesterCapacity)
{
Accelerator = accelerator;
Stream = accelerator.CreateStream();
AcceleratorStream stream = accelerator.CreateStream();
Stream = stream;
Input = accelerator.Allocate1D<ulong>(1);
Output = accelerator.Allocate1D<ulong>(1);
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
PrimeTestInput = Accelerator.Allocate1D<ulong>(primeTesterCapacity);
PrimeTestOutput = Accelerator.Allocate1D<byte>(primeTesterCapacity);
HeuristicFlag = Accelerator.Allocate1D<int>(primeTesterCapacity);
var sharedTables = LastDigitGpuTables.EnsureStaticTables(accelerator, stream);
DevicePrimesLastOne = sharedTables.DevicePrimesLastOne;
DevicePrimesLastSeven = sharedTables.DevicePrimesLastSeven;
DevicePrimesLastThree = sharedTables.DevicePrimesLastThree;
DevicePrimesLastNine = sharedTables.DevicePrimesLastNine;
DevicePrimesPow2LastOne = sharedTables.DevicePrimesPow2LastOne;
DevicePrimesPow2LastSeven = sharedTables.DevicePrimesPow2LastSeven;
DevicePrimesPow2LastThree = sharedTables.DevicePrimesPow2LastThree;
DevicePrimesPow2LastNine = sharedTables.DevicePrimesPow2LastNine;
var keepKernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery);
var keepLauncher = KernelUtil.GetKernel(keepKernel);
@@ -84,6 +186,8 @@ public sealed class Pow2MontgomeryAccelerator
var checkFactorsKernel = accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel);
CheckFactorsKernel = KernelUtil.GetKernel(checkFactorsKernel);
SmallPrimeSieveKernel = (Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>)KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>>().Clone();
}
public void Dispose()
@@ -92,7 +196,17 @@ public sealed class Pow2MontgomeryAccelerator
Output.Dispose();
Stream?.Dispose();
Pow2ModEntriesToTestOnDevice.Dispose();
// Accelerator is shared with other threads
PrimeTestInput?.Dispose();
PrimeTestOutput?.Dispose();
HeuristicFlag?.Dispose();
// Stream is disposed upon return to the queue, so we don't need to dispose it here.
// Stream.Dispose();
// These resources are shared between GPU leases
// ReleaseSharedTables(_sharedTables);
// Accelerator.Dispose();
// Context.Dispose();
// Accelerator.Dispose();
}
}