Document GPU kernel audit progress

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

No files matched your search

@@ -28,23 +28,23 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
private readonly ConcurrentBag<DivisorScanSession> _sessionPool = new();
private Action<Index1D, ArrayView<MontgomeryDivisorData>, ArrayView<ulong>, ArrayView<byte>> GetKernel(Accelerator accelerator) =>
_kernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<MontgomeryDivisorData>, ArrayView<ulong>, ArrayView<byte>>(CheckKernel));
_kernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<MontgomeryDivisorData>, ArrayView<ulong>, ArrayView<byte>>(DivisorByDivisorKernels.CheckKernel));
private Action<Index1D, MontgomeryDivisorData, ArrayView<ulong>, ArrayView<ulong>> GetKernelByPrimeExponent(Accelerator accelerator) =>
_kernelExponentCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, MontgomeryDivisorData, ArrayView<ulong>, ArrayView<ulong>>(ComputeMontgomeryExponentKernel));
_kernelExponentCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, MontgomeryDivisorData, ArrayView<ulong>, ArrayView<ulong>>(DivisorByDivisorKernels.ComputeMontgomeryExponentKernel));
private Action<Index1D, ArrayView<ulong>, ArrayView<byte>, byte> GetRemainderDeltaKernel(Accelerator accelerator) =>
_remainderDeltaKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>, byte>(ComputeRemainderDeltasKernel));
_remainderDeltaKernelCache.GetOrAdd(accelerator, acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>, byte>(DivisorByDivisorKernels.ComputeRemainderDeltasKernel));
private Action<KernelConfig, Index1D, ArrayView<byte>, ArrayView<byte>, byte, byte> GetRemainderScanKernel(Accelerator accelerator) =>
_remainderScanKernelCache.GetOrAdd(
accelerator,
acc => acc.LoadStreamKernel<Index1D, ArrayView<byte>, ArrayView<byte>, byte, byte>(AccumulateRemaindersKernel));
acc => acc.LoadStreamKernel<Index1D, ArrayView<byte>, ArrayView<byte>, byte, byte>(DivisorByDivisorKernels.AccumulateRemaindersKernel));
private Action<Index1D, ArrayView<byte>, ArrayView<byte>, ArrayView<byte>, ArrayView<byte>, byte, ArrayView<byte>> GetCandidateMaskKernel(Accelerator accelerator) =>
_candidateMaskKernelCache.GetOrAdd(
accelerator,
acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<byte>, ArrayView<byte>, ArrayView<byte>, ArrayView<byte>, byte, ArrayView<byte>>(EvaluateCandidateMaskKernel));
acc => acc.LoadAutoGroupedStreamKernel<Index1D, ArrayView<byte>, ArrayView<byte>, ArrayView<byte>, ArrayView<byte>, byte, ArrayView<byte>>(DivisorByDivisorKernels.EvaluateCandidateMaskKernel));
public int GpuBatchSize
{
@@ -535,20 +535,6 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
return Math.Min((1UL << (int)(prime - 1UL)) - 1UL, divisorLimit);
}
private static void CheckKernel(Index1D index, ArrayView<MontgomeryDivisorData> divisors, ArrayView<ulong> exponents, ArrayView<byte> hits)
{
MontgomeryDivisorData divisor = divisors[index];
ulong modulus = divisor.Modulus;
if (modulus <= 1UL || (modulus & 1UL) == 0UL)
{
hits[index] = 0;
return;
}
ulong exponent = exponents[index];
hits[index] = ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisor, keepMontgomery: false) == 1UL ? (byte)1 : (byte)0;
}
public sealed class DivisorScanSession : IMersenneNumberDivisorByDivisorTester.IDivisorScanSession
{
private readonly MersenneNumberDivisorByDivisorGpuTester _owner;
@@ -636,137 +622,6 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
_owner._sessionPool.Add(this);
}
}
private static void ComputeRemainderDeltasKernel(Index1D index, ArrayView<ulong> gaps, ArrayView<byte> deltas, byte modulus)
{
deltas[index] = (byte)(gaps[index] % modulus);
}
private static void AccumulateRemaindersKernel(
Index1D index,
ArrayView<byte> deltas,
ArrayView<byte> remainders,
byte baseRemainder,
byte modulus)
{
int globalIndex = index;
int length = (int)remainders.Length;
if (globalIndex >= length)
{
return;
}
int localIndex = Group.IdxX;
int groupSize = Group.Dimension.X;
var shared = SharedMemory.GetDynamic<int>();
shared[localIndex] = deltas[globalIndex];
Group.Barrier();
int offset = 1;
int mod = modulus;
while (offset < groupSize)
{
int addend = 0;
if (localIndex >= offset)
{
addend = shared[localIndex - offset];
}
Group.Barrier();
if (localIndex >= offset)
{
int sum = shared[localIndex] + addend;
if (sum >= mod)
{
sum %= mod;
}
shared[localIndex] = sum;
}
Group.Barrier();
offset <<= 1;
}
if (globalIndex == 0)
{
remainders[0] = baseRemainder;
return;
}
int remainder = baseRemainder + shared[localIndex];
if (remainder >= mod)
{
remainder %= mod;
}
remainders[globalIndex] = (byte)remainder;
}
private static void EvaluateCandidateMaskKernel(
Index1D index,
ArrayView<byte> remainder10,
ArrayView<byte> remainder8,
ArrayView<byte> remainder5,
ArrayView<byte> remainder3,
byte lastIsSevenFlag,
ArrayView<byte> mask)
{
int globalIndex = index;
int length = (int)mask.Length;
if (globalIndex >= length)
{
return;
}
byte value10 = remainder10[globalIndex];
bool accept10;
if (lastIsSevenFlag != 0)
{
accept10 = value10 == 3 || value10 == 7 || value10 == 9;
}
else
{
accept10 = value10 == 1 || value10 == 3 || value10 == 7 || value10 == 9;
}
if (!accept10)
{
mask[globalIndex] = 0;
return;
}
byte value8 = remainder8[globalIndex];
if (value8 != 1 && value8 != 7)
{
mask[globalIndex] = 0;
return;
}
if (remainder3[globalIndex] == 0 || remainder5[globalIndex] == 0)
{
mask[globalIndex] = 0;
return;
}
mask[globalIndex] = 1;
}
private static void ComputeMontgomeryExponentKernel(Index1D index, MontgomeryDivisorData divisor, ArrayView<ulong> exponents, ArrayView<ulong> results)
{
ulong modulus = divisor.Modulus;
if (modulus <= 1UL || (modulus & 1UL) == 0UL)
{
results[index] = 0UL;
return;
}
ulong exponent = exponents[index];
results[index] = ULongExtensions.Pow2MontgomeryModWindowedKernel(exponent, divisor, keepMontgomery: true);
}
public ulong DivisorLimit
{