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

Document GPU kernel audit progress

Browse Source
This commit is contained in:
Marek Lemańczyk committed 2025-10-18 17:42:33 +02:00
1 parent 1e25ea0ccf
commit 6217ea7888
31 files changed
+3496 -3246
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/PrimeTester.cs
+2 -92
View File
@@ -169,7 +169,7 @@ public sealed class PrimeTester(bool useInternal = false)
{
_accel = accelerator;
// Compile once per accelerator and upload primes once.
Kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<byte>>(SmallPrimeSieveKernel);
Kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<uint>, ArrayView<byte>>(PrimeTesterKernels.SmallPrimeSieveKernel);
var primes = PrimesGenerator.SmallPrimes;
DevicePrimes = accelerator.Allocate1D<uint>(primes.Length);
@@ -266,56 +266,6 @@ public sealed class PrimeTester(bool useInternal = false)
GpuKernelState.Clear(accelerator);
}
// GPU kernel: small-prime sieve only. Returns 1 if passes sieve (probable prime), 0 otherwise.
private static void SmallPrimeSieveKernel(Index1D index, ArrayView<ulong> numbers, ArrayView<uint> smallPrimes, ArrayView<byte> results)
{
ulong n = numbers[index];
if (n <= 3UL)
{
results[index] = (byte)(n >= 2UL ? 1 : 0);
return;
}
if ((n & 1UL) == 0UL || (n > 5UL && (n % 5UL) == 0UL))
{
// TODO: Replace the `% 5` branch with the GPU Mod5 helper once the sieve kernel can
// reuse the benchmarked bitmask reduction instead of modulo on every candidate.
results[index] = 0;
return;
}
if (n.Mod10() == 1UL)
{
// Early reject special GCD heuristic with floor(log2 n)
ulong m = 63UL - (ulong)ILGPU.Algorithms.XMath.LeadingZeroCount(n);
if (BinaryGcdGpu(n, m) != 1UL)
{
results[index] = 0;
return;
}
}
long len = smallPrimes.Length;
for (int i = 0; i < len; i++)
{
ulong p = smallPrimes[i];
if (p * p > n)
{
break;
}
if ((n % p) == 0UL)
{
// TODO: Route this modulo through the shared divisor-cycle cache once exposed to GPU kernels
// so batched sieves avoid per-prime `%` operations that profiling showed expensive.
results[index] = 0;
return;
}
}
results[index] = 1;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static bool SharesFactorWithMaxExponent(ulong n)
{
@@ -333,7 +283,7 @@ public sealed class PrimeTester(bool useInternal = false)
// and divisor-cycle staging instead of allocating new device buffers per call.
var gpu = GpuContextPool.Rent();
var accelerator = gpu.Accelerator;
var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(SharesFactorKernel);
var kernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel);
int length = values.Length;
var inputBuffer = accelerator.Allocate1D<ulong>(length);
@@ -384,45 +334,5 @@ public sealed class PrimeTester(bool useInternal = false)
return u << shift;
}
private static void SharesFactorKernel(Index1D index, ArrayView<ulong> numbers, ArrayView<byte> results)
{
ulong n = numbers[index];
ulong m = 63UL - (ulong)XMath.LeadingZeroCount(n);
ulong gcd = BinaryGcdGpu(n, m);
results[index] = gcd == 1UL ? (byte)0 : (byte)1;
}
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.
if (u == 0UL)
{
return v;
}
if (v == 0UL)
{
return u;
}
int shift = XMath.TrailingZeroCount(u | v);
u >>= XMath.TrailingZeroCount(u);
do
{
v >>= XMath.TrailingZeroCount(v);
if (u > v)
{
(u, v) = (v, u);
}
v -= u;
}
while (v != 0UL);
return u << shift;
}
}