MontgomeryDivisorData converted to a pooled class to limit GC pressure. MontgomeryDivisorDataGpu added for GPU operations. PrimeOrderCalculator kernels precompiled during warmup.
This commit is contained in:
mlemanczyk committed
1 parent
07efd5d2a1
commit
070e2a14fd
32 files changed
+310
-192
No files matched your search
@@ -69,6 +69,7 @@ public sealed class PrimeOrderCalculatorAccelerator
|
||||
AcceleratorStream stream = accelerator.CreateStream();
|
||||
// TODO: Review which tables are needed when final execution path is defined.
|
||||
LastDigitGpuTables.WarmUp(i, stream);
|
||||
_kernels[i] = new CalculatorKernels(accelerator);
|
||||
// HeuristicCombinedGpuTables.WarmUp(i, stream);
|
||||
// SmallPrimeFactorGpuTables.WarmUp(i, stream);
|
||||
// SharedHeuristicGpuTables.EnsureStaticTables(accelerator, stream);
|
||||
@@ -115,18 +116,41 @@ public sealed class PrimeOrderCalculatorAccelerator
|
||||
|
||||
public Dictionary<ulong, int> Pow2ModEntriesToTestOnHost = [];
|
||||
|
||||
private static readonly CalculatorKernels[] _kernels =new CalculatorKernels[AcceleratorPool.Shared.Accelerators.Length];
|
||||
|
||||
private sealed class CalculatorKernels(Accelerator accelerator)
|
||||
{
|
||||
public readonly Kernel CheckFactorsKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel));
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelConvertToStandard)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews> HeuristicCombinedTrialDivisionKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>(PrimeTesterKernels.HeuristicTrialCombinedDivisionKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>>();
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>> KeepMontgomeryKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
public readonly Kernel PollardRhoKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.TryPollardRhoKernel));
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>> Pow2ModWideKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(PrimeOrderGpuHeuristics.Pow2ModKernelWide)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<byte>> SharesFactorKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<byte>>>();
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SmallPrimeFactorKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
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>> SmallPrimeSieveKernelLauncher = 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>>>();
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, ulong, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SpecialMaxKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, ulong, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, ulong, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
}
|
||||
|
||||
public readonly Kernel CheckFactorsKernel;
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernelLauncher;
|
||||
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews> HeuristicCombinedTrialDivisionKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> KeepMontgomeryKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>> KeepMontgomeryKernelLauncher;
|
||||
public readonly Kernel PollardRhoKernel;
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>> Pow2ModWideKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<byte>> SharesFactorKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SmallPrimeFactorKernelLauncher;
|
||||
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>> SmallPrimeSieveKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SpecialMaxKernelLauncher;
|
||||
public readonly Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, ulong, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>> SpecialMaxKernelLauncher;
|
||||
|
||||
public void EnsureCapacity(int factorsCount, int primeTesterCapacity)
|
||||
{
|
||||
@@ -260,25 +284,18 @@ public sealed class PrimeOrderCalculatorAccelerator
|
||||
SmallPrimeFactorSquares = tables.Squares.View;
|
||||
}
|
||||
|
||||
CheckFactorsKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel));
|
||||
var kernels = _kernels[acceleratorIndex];
|
||||
|
||||
ConvertToStandardKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelConvertToStandard)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
HeuristicCombinedTrialDivisionKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>(PrimeTesterKernels.HeuristicTrialCombinedDivisionKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>>();
|
||||
|
||||
KeepMontgomeryKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
PollardRhoKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.TryPollardRhoKernel));
|
||||
|
||||
Pow2ModWideKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(PrimeOrderGpuHeuristics.Pow2ModKernelWide)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
|
||||
|
||||
SharesFactorKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView<ulong>, ArrayView<byte>>>();
|
||||
|
||||
SmallPrimeFactorKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(SmallPrimeFactorKernels.SmallPrimeFactorKernelScan)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, uint, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
SmallPrimeSieveKernelLauncher = 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>>>();
|
||||
|
||||
SpecialMaxKernelLauncher = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel)).CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
CheckFactorsKernel = kernels.CheckFactorsKernel;
|
||||
ConvertToStandardKernelLauncher = kernels.ConvertToStandardKernelLauncher;
|
||||
HeuristicCombinedTrialDivisionKernelLauncher = kernels.HeuristicCombinedTrialDivisionKernelLauncher;
|
||||
KeepMontgomeryKernelLauncher = kernels.KeepMontgomeryKernelLauncher;
|
||||
PollardRhoKernel = kernels.PollardRhoKernel;
|
||||
Pow2ModWideKernelLauncher = kernels.Pow2ModWideKernelLauncher;
|
||||
SharesFactorKernelLauncher = kernels.SharesFactorKernelLauncher;
|
||||
SmallPrimeFactorKernelLauncher = kernels.SmallPrimeFactorKernelLauncher;
|
||||
SmallPrimeSieveKernelLauncher = kernels.SmallPrimeSieveKernelLauncher;
|
||||
SpecialMaxKernelLauncher = kernels.SpecialMaxKernelLauncher;
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
@@ -17,18 +17,28 @@ internal static class Pow2MontgomeryKernels
|
||||
public static void Pow2MontgomeryKernelKeepMontgomery(
|
||||
Index1D index,
|
||||
ArrayView1D<ulong, Stride1D.Dense> exponents,
|
||||
MontgomeryDivisorData divisor,
|
||||
ulong divisorModulus,
|
||||
ulong divisorNPrime,
|
||||
ulong divisorMontgomeryOne,
|
||||
ulong divisorMontgomeryTwo,
|
||||
ulong divisorMontgomeryTwoSquared,
|
||||
ArrayView1D<ulong, Stride1D.Dense> results)
|
||||
{
|
||||
MontgomeryDivisorDataGpu divisor = new(divisorModulus, divisorNPrime, divisorMontgomeryOne, divisorMontgomeryTwo, divisorMontgomeryTwoSquared);
|
||||
results[0] = ULongExtensions.Pow2MontgomeryModWindowedGpuKeepMontgomery(divisor, exponents[0]);
|
||||
}
|
||||
|
||||
public static void Pow2MontgomeryKernelConvertToStandard(
|
||||
Index1D index,
|
||||
ArrayView1D<ulong, Stride1D.Dense> exponents,
|
||||
MontgomeryDivisorData divisor,
|
||||
ulong divisorModulus,
|
||||
ulong divisorNPrime,
|
||||
ulong divisorMontgomeryOne,
|
||||
ulong divisorMontgomeryTwo,
|
||||
ulong divisorMontgomeryTwoSquared,
|
||||
ArrayView1D<ulong, Stride1D.Dense> results)
|
||||
{
|
||||
MontgomeryDivisorDataGpu divisor = new(divisorModulus, divisorNPrime, divisorMontgomeryOne, divisorMontgomeryTwo, divisorMontgomeryTwoSquared);
|
||||
results[0] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponents[0]);
|
||||
}
|
||||
|
||||
|
||||
@@ -11,12 +11,17 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
int count,
|
||||
ulong phi,
|
||||
ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense> toTest,
|
||||
MontgomeryDivisorData divisorData,
|
||||
ulong divisorModulus,
|
||||
ulong divisorNPrime,
|
||||
ulong divisorMontgomeryOne,
|
||||
ulong divisorMontgomeryTwo,
|
||||
ulong divisorMontgomeryTwoSquared,
|
||||
ArrayView1D<ulong, Stride1D.Dense> result
|
||||
)
|
||||
{
|
||||
ulong order = phi;
|
||||
int entryCount = count;
|
||||
MontgomeryDivisorDataGpu divisorData = new(divisorModulus, divisorNPrime, divisorMontgomeryOne, divisorMontgomeryTwo, divisorMontgomeryTwoSquared);
|
||||
|
||||
for (int i = 0; i < entryCount; i++)
|
||||
{
|
||||
@@ -155,7 +160,11 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
Index1D index,
|
||||
ulong prime,
|
||||
OrderKernelConfig config,
|
||||
MontgomeryDivisorData divisor,
|
||||
ulong divisorModulus,
|
||||
ulong divisorNPrime,
|
||||
ulong divisorMontgomeryOne,
|
||||
ulong divisorMontgomeryTwo,
|
||||
ulong divisorMontgomeryTwoSquared,
|
||||
ArrayView1D<uint, Stride1D.Dense> primes,
|
||||
ArrayView1D<ulong, Stride1D.Dense> squares,
|
||||
int primeCount,
|
||||
@@ -227,6 +236,8 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
|
||||
SortFactors(phiFactors, phiExponents, phiFactorCount);
|
||||
|
||||
MontgomeryDivisorDataGpu divisor = new(divisorModulus, divisorNPrime, divisorMontgomeryOne, divisorMontgomeryTwo, divisorMontgomeryTwoSquared);
|
||||
|
||||
if (TrySpecialMaxKernel(phi, prime, phiFactors, phiFactorCount, divisor))
|
||||
{
|
||||
resultOut[0] = phi;
|
||||
@@ -381,7 +392,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
ulong prime,
|
||||
ArrayView1D<ulong, Stride1D.Dense> factors,
|
||||
int factorCount,
|
||||
in MontgomeryDivisorData divisor)
|
||||
in MontgomeryDivisorDataGpu divisor)
|
||||
{
|
||||
for (int i = 0; i < factorCount; i++)
|
||||
{
|
||||
@@ -401,7 +412,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
return true;
|
||||
}
|
||||
|
||||
private static ulong InitializeStartingOrderKernel(ulong prime, ulong phi, in MontgomeryDivisorData divisor)
|
||||
private static ulong InitializeStartingOrderKernel(ulong prime, ulong phi, in MontgomeryDivisorDataGpu divisor)
|
||||
{
|
||||
ulong order = phi;
|
||||
ulong residue = prime & 7UL;
|
||||
@@ -423,7 +434,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
ArrayView1D<ulong, Stride1D.Dense> factors,
|
||||
ArrayView1D<int, Stride1D.Dense> exponents,
|
||||
int factorCount,
|
||||
in MontgomeryDivisorData divisor)
|
||||
in MontgomeryDivisorDataGpu divisor)
|
||||
{
|
||||
for (int i = 0; i < factorCount; i++)
|
||||
{
|
||||
@@ -458,7 +469,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
private static bool TryConfirmOrderKernel(
|
||||
ulong prime,
|
||||
ulong order,
|
||||
in MontgomeryDivisorData divisor,
|
||||
in MontgomeryDivisorDataGpu divisor,
|
||||
uint limit,
|
||||
ArrayView1D<uint, Stride1D.Dense> primes,
|
||||
ArrayView1D<ulong, Stride1D.Dense> squares,
|
||||
@@ -532,7 +543,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
ulong order,
|
||||
ulong previousOrder,
|
||||
byte hasPreviousOrder,
|
||||
in MontgomeryDivisorData divisor,
|
||||
in MontgomeryDivisorDataGpu divisor,
|
||||
uint limit,
|
||||
int maxPowChecks,
|
||||
ArrayView1D<uint, Stride1D.Dense> primes,
|
||||
@@ -851,7 +862,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
private static bool TryConfirmCandidateKernel(
|
||||
ulong prime,
|
||||
ulong candidate,
|
||||
in MontgomeryDivisorData divisor,
|
||||
in MontgomeryDivisorDataGpu divisor,
|
||||
uint limit,
|
||||
ArrayView1D<uint, Stride1D.Dense> primes,
|
||||
ArrayView1D<ulong, Stride1D.Dense> squares,
|
||||
@@ -1174,7 +1185,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
}
|
||||
}
|
||||
|
||||
private static bool Pow2EqualsOneKernel(ulong exponent, in MontgomeryDivisorData divisor)
|
||||
private static bool Pow2EqualsOneKernel(ulong exponent, in MontgomeryDivisorDataGpu divisor)
|
||||
{
|
||||
return ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponent) == 1UL;
|
||||
}
|
||||
@@ -1200,9 +1211,10 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
|
||||
/// This kernel always sets the result of the corresponding element. Callers don't need to clear the output buffers.
|
||||
|
||||
internal static void Pow2ModKernel(Index1D index, ArrayView1D<ulong, Stride1D.Dense> exponents, MontgomeryDivisorData divisor, ArrayView1D<ulong, Stride1D.Dense> remainders)
|
||||
internal static void Pow2ModKernel(Index1D index, ArrayView1D<ulong, Stride1D.Dense> exponents, ulong divisorModulus, ulong divisorNPrime, ulong divisorMontgomeryOne, ulong divisorMontgomeryTwo, ulong divisorMontgomeryTwoSquared, ArrayView1D<ulong, Stride1D.Dense> remainders)
|
||||
{
|
||||
ulong exponent = exponents[index];
|
||||
MontgomeryDivisorDataGpu divisor = new(divisorModulus, divisorNPrime, divisorMontgomeryOne, divisorMontgomeryTwo, divisorMontgomeryTwoSquared);
|
||||
remainders[index] = ULongExtensions.Pow2MontgomeryModWindowedGpuConvertToStandard(divisor, exponent);
|
||||
}
|
||||
|
||||
|
||||
@@ -10,7 +10,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
ulong phi,
|
||||
ArrayView1D<ulong, Stride1D.Dense> factors,
|
||||
int factorCount,
|
||||
MontgomeryDivisorData divisor,
|
||||
ulong divisorModulus,
|
||||
ArrayView1D<ulong, Stride1D.Dense> candidates,
|
||||
ArrayView1D<ulong, Stride1D.Dense> resultOut)
|
||||
{
|
||||
@@ -57,7 +57,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
candidates[insert + 1] = current;
|
||||
}
|
||||
|
||||
var divisorPartial = new GpuDivisorPartialData(divisor.Modulus);
|
||||
var divisorPartial = new GpuDivisorPartialData(divisorModulus);
|
||||
var stepper = new ExponentRemainderStepperGpu(divisorPartial);
|
||||
if (stepper.InitializeIsUnityGpu(candidates[0]))
|
||||
{
|
||||
|
||||
@@ -260,7 +260,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
|
||||
ArrayView1D<ulong, Stride1D.Dense> exponentViewDevice = exponentBuffer.View;
|
||||
ArrayView1D<byte, Stride1D.Dense> hitsView = hitsBuffer.View;
|
||||
ArrayView1D<int, Stride1D.Dense> hitIndexView = hitIndexBuffer.View;
|
||||
|
||||
Queue<MontgomeryDivisorData> divisorPool = MontgomeryDivisorDataPool.Shared;
|
||||
while (remainingCount > 0UL && !composite)
|
||||
{
|
||||
int chunkCount = chunkCountBaseline;
|
||||
@@ -284,7 +284,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
|
||||
if (passesSmallModuli && (localRemainder8 == 1 || localRemainder8 == 7) && ((decimalMask >> localRemainder10) & 1) != 0)
|
||||
{
|
||||
ulong divisorValue = (ulong)nextDivisor128;
|
||||
MontgomeryDivisorData montgomeryData = MontgomeryDivisorData.FromModulus(divisorValue);
|
||||
MontgomeryDivisorData montgomeryData = divisorPool.FromModulus(divisorValue);
|
||||
ulong divisorCycle = ResolveDivisorCycle(gpu, divisorValue, prime, in montgomeryData);
|
||||
if (divisorCycle == prime)
|
||||
{
|
||||
@@ -294,6 +294,7 @@ public sealed partial class MersenneNumberDivisorByDivisorGpuTester : IMersenneN
|
||||
return true;
|
||||
}
|
||||
|
||||
divisorPool.Return(montgomeryData);
|
||||
divisorSpan[admissibleCount] = divisorValue;
|
||||
divisorDataSpan[admissibleCount] = new GpuDivisorPartialData(divisorValue);
|
||||
offsetSpan[admissibleCount] = admissibleCount;
|
||||
|
||||
@@ -0,0 +1,10 @@
|
||||
namespace PerfectNumbers.Core.Gpu;
|
||||
|
||||
public readonly struct MontgomeryDivisorDataGpu(ulong modulus, ulong nPrime, ulong montgomeryOne, ulong montgomeryTwo, ulong montgomeryTwoSquared)
|
||||
{
|
||||
public readonly ulong Modulus = modulus;
|
||||
public readonly ulong NPrime = nPrime;
|
||||
public readonly ulong MontgomeryOne = montgomeryOne;
|
||||
public readonly ulong MontgomeryTwo = montgomeryTwo;
|
||||
public readonly ulong MontgomeryTwoSquared = montgomeryTwoSquared;
|
||||
}
|
||||
@@ -25,7 +25,7 @@ public static class Pow2MontgomeryGpuCalculator
|
||||
// resultBuffer.MemSetToZero(stream);
|
||||
|
||||
var kernelLauncher = gpu.KeepMontgomeryKernelLauncher;
|
||||
kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
|
||||
kernelLauncher(stream, 1, exponentBuffer.View, divisor.Modulus, divisor.NPrime, divisor.MontgomeryOne, divisor.MontgomeryTwo, divisor.MontgomeryTwoSquared, resultBuffer.View);
|
||||
|
||||
resultBuffer.View.CopyToCPU(stream, ref result, 1);
|
||||
stream.Synchronize();
|
||||
@@ -54,7 +54,7 @@ public static class Pow2MontgomeryGpuCalculator
|
||||
|
||||
var kernelLauncher = gpu.ConvertToStandardKernelLauncher!;
|
||||
|
||||
kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
|
||||
kernelLauncher(stream, 1, exponentBuffer.View, divisor.Modulus, divisor.NPrime, divisor.MontgomeryOne, divisor.MontgomeryTwo, divisor.MontgomeryTwoSquared, resultBuffer.View);
|
||||
|
||||
resultBuffer.View.CopyToCPU(stream, ref result, 1);
|
||||
stream.Synchronize();
|
||||
@@ -82,7 +82,7 @@ public static class Pow2MontgomeryGpuCalculator
|
||||
|
||||
var accelerator = gpu.Accelerator;
|
||||
var kernelLauncher = gpu.ConvertToStandardKernelLauncher;
|
||||
kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
|
||||
kernelLauncher(stream, 1, exponentBuffer.View, divisor.Modulus, divisor.NPrime, divisor.MontgomeryOne, divisor.MontgomeryTwo, divisor.MontgomeryTwoSquared, resultBuffer.View);
|
||||
|
||||
resultBuffer.View.CopyToCPU(stream, ref result, 1);
|
||||
stream.Synchronize();
|
||||
|
||||
@@ -398,14 +398,18 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
statusBuffer.View);
|
||||
|
||||
var orderKernel = GetOrderKernel(acceleratorIndex);
|
||||
var kernelLauncher = orderKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, OrderKernelConfig, MontgomeryDivisorData, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>>();
|
||||
var kernelLauncher = orderKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, OrderKernelConfig, ulong, ulong, ulong, ulong, ulong, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>>();
|
||||
|
||||
kernelLauncher(
|
||||
stream,
|
||||
1,
|
||||
prime,
|
||||
kernelConfig,
|
||||
divisorData,
|
||||
divisorData.Modulus,
|
||||
divisorData.NPrime,
|
||||
divisorData.MontgomeryOne,
|
||||
divisorData.MontgomeryTwo,
|
||||
divisorData.MontgomeryTwoSquared,
|
||||
smallPrimesView,
|
||||
smallSquaresView,
|
||||
(int)smallPrimesView.Length,
|
||||
@@ -417,7 +421,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
resultBuffer.View.CopyToCPU(stream, ref order, 1);
|
||||
stream.Synchronize();
|
||||
|
||||
stream.Dispose();
|
||||
AcceleratorStreamPool.Return(acceleratorIndex, stream);
|
||||
phiFactorBuffer.Dispose();
|
||||
phiExponentBuffer.Dispose();
|
||||
workFactorBuffer.Dispose();
|
||||
@@ -458,7 +462,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
}
|
||||
|
||||
var accelerator = _accelerators[acceleratorIndex];
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, OrderKernelConfig, MontgomeryDivisorData, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>(CalculateOrderKernel);
|
||||
var loaded = accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, OrderKernelConfig, ulong, ulong, ulong, ulong, ulong, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>(CalculateOrderKernel);
|
||||
|
||||
var kernel = KernelUtil.GetKernel(loaded);
|
||||
pool[acceleratorIndex] = kernel;
|
||||
@@ -530,11 +534,12 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
// remainderBuffer.MemSetToZero(stream);
|
||||
|
||||
var pow2ModKernel = GetPow2ModKernel(acceleratorIndex);
|
||||
var kernelLauncher = pow2ModKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
var kernelLauncher = pow2ModKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>>();
|
||||
|
||||
try
|
||||
{
|
||||
kernelLauncher(stream, exponents.Length, exponentBuffer.View, divisorData, remainderBuffer.View);
|
||||
kernelLauncher(stream, exponents.Length, exponentBuffer.View, divisorData.Modulus, divisorData.NPrime, divisorData.MontgomeryOne, divisorData.MontgomeryTwo, divisorData.MontgomeryTwoSquared, remainderBuffer.View);
|
||||
|
||||
remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(results), exponents.Length);
|
||||
stream.Synchronize();
|
||||
|
||||
@@ -557,7 +562,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
return true;
|
||||
}
|
||||
|
||||
private static void ComputePow2ModCpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorData divisorData, Span<ulong> results)
|
||||
private static void ComputePow2ModCpu(ReadOnlySpan<ulong> exponents, ulong prime, in MontgomeryDivisorDataGpu divisorData, Span<ulong> results)
|
||||
{
|
||||
int length = exponents.Length;
|
||||
for (int i = 0; i < length; i++)
|
||||
@@ -636,7 +641,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
}
|
||||
}
|
||||
|
||||
private static ulong Pow2ModCpu(ulong exponent, ulong modulus, in MontgomeryDivisorData divisorData)
|
||||
private static ulong Pow2ModCpu(ulong exponent, ulong modulus, in MontgomeryDivisorDataGpu divisorData)
|
||||
{
|
||||
if (modulus <= 1UL)
|
||||
{
|
||||
@@ -655,7 +660,7 @@ internal static partial class PrimeOrderGpuHeuristics
|
||||
}
|
||||
|
||||
var accelerator = _accelerators[acceleratorIndex];
|
||||
var loaded = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernel));
|
||||
var loaded = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, ulong, ulong, ulong, ulong, ulong, ArrayView1D<ulong, Stride1D.Dense>>(Pow2ModKernel));
|
||||
|
||||
pool[acceleratorIndex] = loaded;
|
||||
return loaded;
|
||||
|
||||
Reference in new issue
Block a user