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

Merge pull request #246 from mlemanczyk/codex/compare-cpu-and-gpu-execution-paths

Browse Source 
Make GPU sieve results authoritative
This commit is contained in:
Marek Lemańczyk authored and GitHub committed 2025-10-28 20:14:33 +01:00
2 parents 2f461ee001 1792349bad
commit d149700311
3 files changed
+60 -15
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core.Tests/Gpu/GpuKernelCompilationTests.cs
+2 -2
View File
@@ -121,7 +121,7 @@ public class GpuKernelCompilationTests
private static readonly Action<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> MersenneDivisorCyclesKernel =
DivisorCycleKernels.GpuDivisorCycleKernel;
private static readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<byte>> PrimeTesterSmallPrimeKernel =
private static readonly Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> PrimeTesterSmallPrimeKernel =
PrimeTesterKernels.SmallPrimeSieveKernel;
private static readonly Action<Index1D, ArrayView<ulong>, ArrayView<byte>> PrimeTesterSharesFactorKernel =
@@ -336,7 +336,7 @@ public class GpuKernelCompilationTests
private static void CompilePrimeTesterSmallPrimeKernel(Accelerator accelerator)
{
var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<byte>>(PrimeTesterSmallPrimeKernel);
var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterSmallPrimeKernel);
_ = kernel;
}
PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs
+12 -2
View File
@@ -15,34 +15,44 @@ internal static class PrimeTesterKernels
ArrayView<uint> smallPrimesLastSeven,
ArrayView<uint> smallPrimesLastThree,
ArrayView<uint> smallPrimesLastNine,
ArrayView<ulong> smallPrimesPow2,
ArrayView<ulong> smallPrimesPow2LastOne,
ArrayView<ulong> smallPrimesPow2LastSeven,
ArrayView<ulong> smallPrimesPow2LastThree,
ArrayView<ulong> smallPrimesPow2LastNine,
ArrayView<byte> results)
{
ulong n = numbers[index];
byte result = 1;
ArrayView<uint> primes = smallPrimes;
ArrayView<ulong> primeSquares = smallPrimesPow2;
ulong lastDigit = n % 10UL;
switch (lastDigit)
{
case 1UL:
primes = smallPrimesLastOne;
primeSquares = smallPrimesPow2LastOne;
break;
case 3UL:
primes = smallPrimesLastThree;
primeSquares = smallPrimesPow2LastThree;
break;
case 7UL:
primes = smallPrimesLastSeven;
primeSquares = smallPrimesPow2LastSeven;
break;
case 9UL:
primes = smallPrimesLastNine;
primeSquares = smallPrimesPow2LastNine;
break;
}
int length = (int)primes.Length;
int length = XMath.Min((int)primes.Length, (int)primeSquares.Length);
for (int i = 0; i < length; i++)
{
ulong prime = primes[i];
ulong primeSquare = prime * prime;
ulong primeSquare = primeSquares[i];
if (primeSquare > n)
{
break;
PerfectNumbers.Core/PrimeTester.cs
+46 -11
View File
@@ -101,21 +101,21 @@ public sealed class PrimeTester
public bool IsPrimeGpu(ulong n, CancellationToken ct)
{
bool forceCpu = GpuContextPool.ForceCpu;
bool belowGpuRange = n < 31UL;
if (forceCpu || belowGpuRange)
{
return IsPrimeCpu(n, ct);
}
Span<ulong> one = stackalloc ulong[1];
Span<byte> outFlags = stackalloc byte[1];
one[0] = n;
outFlags[0] = 0;
if (!forceCpu)
{
IsPrimeBatchGpu(one, outFlags);
}
IsPrimeBatchGpu(one, outFlags);
bool belowGpuRange = n < 31UL;
bool gpuReportedPrime = !forceCpu && !belowGpuRange && outFlags[0] != 0;
bool requiresCpuFallback = forceCpu || belowGpuRange || !gpuReportedPrime;
return requiresCpuFallback ? IsPrimeCpu(n, ct) : true;
return outFlags[0] != 0;
}
public static bool IsPrimeCpu(ulong n, CancellationToken ct)
@@ -162,6 +162,11 @@ public sealed class PrimeTester
state.DevicePrimesLastSeven.View,
state.DevicePrimesLastThree.View,
state.DevicePrimesLastNine.View,
state.DevicePrimesPow2Default.View,
state.DevicePrimesPow2LastOne.View,
state.DevicePrimesPow2LastSeven.View,
state.DevicePrimesPow2LastThree.View,
state.DevicePrimesPow2LastNine.View,
output.View);
accelerator.Synchronize();
output.View.CopyToCPU(ref results[pos], count);
@@ -275,13 +280,18 @@ public sealed class PrimeTester
// Per-accelerator GPU state for prime sieve (kernel + uploaded primes).
internal sealed class KernelState
{
public Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<byte>> Kernel { get; }
public Action<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>> Kernel { get; }
public Action<Index1D, ArrayView<ulong>, ulong, ArrayView<byte>> HeuristicTrialDivisionKernel { get; }
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesDefault { get; }
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastOne { get; }
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastSeven { get; }
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastThree { get; }
public MemoryBuffer1D<uint, Stride1D.Dense> DevicePrimesLastNine { get; }
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2Default { get; }
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastOne { get; }
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastSeven { get; }
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastThree { get; }
public MemoryBuffer1D<ulong, Stride1D.Dense> DevicePrimesPow2LastNine { get; }
public bool IsDisposed { get; private set; }
private readonly System.Collections.Concurrent.ConcurrentBag<ScratchBuffers> _scratchPool = [];
// TODO: Replace this ConcurrentBag with the lock-free ring buffer variant validated in
@@ -292,7 +302,7 @@ public sealed class PrimeTester
{
IsDisposed = false;
// Compile once per accelerator and upload primes once.
Kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel);
Kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<uint>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel);
HeuristicTrialDivisionKernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ulong, ArrayView<byte>>(PrimeTesterKernels.HeuristicTrialDivisionKernel);
var primesDefault = DivisorGenerator.SmallPrimes;
@@ -314,6 +324,26 @@ public sealed class PrimeTester
var primesLastNine = DivisorGenerator.SmallPrimesLastNine;
DevicePrimesLastNine = accelerator.Allocate1D<uint>(primesLastNine.Length);
DevicePrimesLastNine.View.CopyFromCPU(primesLastNine);
var primesPow2Default = DivisorGenerator.SmallPrimesPow2;
DevicePrimesPow2Default = accelerator.Allocate1D<ulong>(primesPow2Default.Length);
DevicePrimesPow2Default.View.CopyFromCPU(primesPow2Default);
var primesPow2LastOne = DivisorGenerator.SmallPrimesPow2LastOne;
DevicePrimesPow2LastOne = accelerator.Allocate1D<ulong>(primesPow2LastOne.Length);
DevicePrimesPow2LastOne.View.CopyFromCPU(primesPow2LastOne);
var primesPow2LastSeven = DivisorGenerator.SmallPrimesPow2LastSeven;
DevicePrimesPow2LastSeven = accelerator.Allocate1D<ulong>(primesPow2LastSeven.Length);
DevicePrimesPow2LastSeven.View.CopyFromCPU(primesPow2LastSeven);
var primesPow2LastThree = DivisorGenerator.SmallPrimesPow2LastThree;
DevicePrimesPow2LastThree = accelerator.Allocate1D<ulong>(primesPow2LastThree.Length);
DevicePrimesPow2LastThree.View.CopyFromCPU(primesPow2LastThree);
var primesPow2LastNine = DivisorGenerator.SmallPrimesPow2LastNine;
DevicePrimesPow2LastNine = accelerator.Allocate1D<ulong>(primesPow2LastNine.Length);
DevicePrimesPow2LastNine.View.CopyFromCPU(primesPow2LastNine);
}
internal sealed class ScratchBuffers
@@ -376,6 +406,11 @@ public sealed class PrimeTester
DevicePrimesLastSeven.Dispose();
DevicePrimesLastThree.Dispose();
DevicePrimesLastNine.Dispose();
DevicePrimesPow2Default.Dispose();
DevicePrimesPow2LastOne.Dispose();
DevicePrimesPow2LastSeven.Dispose();
DevicePrimesPow2LastThree.Dispose();
DevicePrimesPow2LastNine.Dispose();
IsDisposed = true;
}
}