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

BinaryGcd switched to the best performing variant according to benchmarks.

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-14 11:43:16 +01:00
1 parent be435b7760
commit 42447c4e43
9 files changed
+112 -166
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/Kernels/PrimeOrderKernels.cs
-18
View File
@@ -2,7 +2,6 @@ using System.Numerics;
using System.Runtime.CompilerServices;
using ILGPU;
using ILGPU.Runtime;
using PerfectNumbers.Core.Cpu;
namespace PerfectNumbers.Core.Gpu;
@@ -1207,23 +1206,6 @@ internal static partial class PrimeOrderGpuHeuristics
remainders[index] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponent);
}
private static Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>> GetPow2ModWideKernel(Accelerator accelerator)
{
var pool = _pow2ModWideKernel ??= [];
if (pool.TryGetValue(accelerator, out var cached))
{
return cached;
}
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(Pow2ModKernelWide);
var kernel = KernelUtil.GetKernel(loaded);
cached = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
pool[accelerator] = cached;
return cached;
}
/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.
internal static void Pow2ModKernelWide(Index1D index, ArrayView1D<GpuUInt128, Stride1D.Dense> exponents, GpuUInt128 modulus, ArrayView1D<GpuUInt128, Stride1D.Dense> remainders)
{
PerfectNumbers.Core/Gpu/Kernels/PrimeTesterKernels.cs
+21 -19
View File
@@ -121,35 +121,37 @@ internal static class PrimeTesterKernels
private static ulong BinaryGcdGpu(ulong u, ulong v)
{
// TODO: Replace this inline GPU binary GCD with the kernel extracted from
// GpuUInt128BinaryGcdBenchmarks via GpuKernelPool so device callers reuse the
// fully unrolled ladder instead of this branchy fallback.
bool uIsZero = u == 0UL;
bool vIsZero = v == 0UL;
bool eitherZero = uIsZero | vIsZero;
ulong zeroResult = uIsZero ? v : u;
ulong combined = u | v | 1UL;
ulong combined = (u | v) | 1UL;
int shift = XMath.TrailingZeroCount(combined);
ulong currentU = u >> XMath.TrailingZeroCount(u | 1UL);
ulong currentV = v >> XMath.TrailingZeroCount(v | 1UL);
ulong normalizedU = u >> XMath.TrailingZeroCount(u | 1UL);
ulong normalizedV = v >> XMath.TrailingZeroCount(v | 1UL);
ulong currentU = normalizedU;
ulong currentV = normalizedV;
bool loopCondition = currentV != 0UL;
while (loopCondition)
while (true)
{
currentV >>= XMath.TrailingZeroCount(currentV);
ulong minValue = XMath.Min(currentU, currentV);
ulong maxValue = XMath.Max(currentU, currentV);
currentU = minValue;
bool swap = currentU > currentV;
ulong minValue = ConditionalSelect(currentU, currentV, swap);
ulong maxValue = ConditionalSelect(currentV, currentU, swap);
currentV = maxValue - minValue;
loopCondition = currentV != 0UL;
currentU = minValue;
if (currentV == 0UL)
{
ulong gcd = currentU << shift;
return eitherZero ? zeroResult : gcd;
}
}
ulong gcd = currentU << shift;
return eitherZero ? zeroResult : gcd;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
private static ulong ConditionalSelect(ulong left, ulong right, bool useRight)
{
ulong mask = useRight ? ulong.MaxValue : 0UL;
return left ^ ((left ^ right) & mask);
}
}