Hybrid device mode implemented as the current --mersenne=bydivisor CPU execution path. CPU path is now limited to CPU calculations, only.
This commit is contained in:
mlemanczyk committed
1 parent
353a121d03
commit
941733905b
49 files changed
+4965
-1726
No files matched your search
@@ -1,3 +1,4 @@
|
||||
using System.Buffers;
|
||||
using ILGPU;
|
||||
using ILGPU.Runtime;
|
||||
using PerfectNumbers.Core.Gpu;
|
||||
@@ -7,17 +8,408 @@ namespace PerfectNumbers.Core;
|
||||
|
||||
internal static partial class PrimeOrderCalculator
|
||||
{
|
||||
private static bool IsGpuPow2Allowed => s_pow2ModeInitialized && s_allowGpuPow2;
|
||||
|
||||
private const int GpuSmallPrimeFactorSlots = 64;
|
||||
|
||||
public static ulong CalculateGpu(
|
||||
PrimeOrderCalculatorAccelerator gpu,
|
||||
ulong prime,
|
||||
ulong? previousOrder,
|
||||
in MontgomeryDivisorData divisorData,
|
||||
in PrimeOrderSearchConfig config)
|
||||
{
|
||||
// TODO: Is this condition ever met on EvenPerfectBitScanner's execution path? If not, we can add a clarification comment and comment out the entire block. We want to support p candidates at least greater or equal to 31.
|
||||
if (prime <= 3UL)
|
||||
{
|
||||
return prime == 3UL ? 2UL : 1UL;
|
||||
}
|
||||
|
||||
ulong phi = prime - 1UL;
|
||||
|
||||
if (PrimeOrderGpuHeuristics.TryCalculateOrder(gpu, prime, previousOrder, config, divisorData, out ulong gpuOrder))
|
||||
{
|
||||
return gpuOrder;
|
||||
}
|
||||
|
||||
PartialFactorResult phiFactors = PartialFactorGpu(gpu, phi, config);
|
||||
|
||||
ulong result;
|
||||
if (phiFactors.Factors is null)
|
||||
{
|
||||
result = CalculateByFactorizationCpu(prime, divisorData);
|
||||
|
||||
phiFactors.Dispose();
|
||||
return result;
|
||||
}
|
||||
|
||||
result = RunHeuristicPipelineCpu(prime, previousOrder, config, divisorData, phi, phiFactors);
|
||||
phiFactors.Dispose();
|
||||
return result;
|
||||
}
|
||||
|
||||
public static UInt128 CalculateGpu(
|
||||
PrimeOrderCalculatorAccelerator gpu,
|
||||
in UInt128 prime,
|
||||
in UInt128? previousOrder,
|
||||
in PrimeOrderSearchConfig config)
|
||||
{
|
||||
MontgomeryDivisorData divisorData;
|
||||
UInt128 result;
|
||||
if (prime <= ulong.MaxValue)
|
||||
{
|
||||
ulong? previous = null;
|
||||
if (previousOrder.HasValue)
|
||||
{
|
||||
UInt128 previousValue = previousOrder.Value;
|
||||
if (previousValue <= ulong.MaxValue)
|
||||
{
|
||||
previous = (ulong)previousValue;
|
||||
}
|
||||
else
|
||||
{
|
||||
previous = ulong.MaxValue;
|
||||
}
|
||||
}
|
||||
|
||||
ulong prime64 = (ulong)prime;
|
||||
Queue<MontgomeryDivisorData> divisorPool = MontgomeryDivisorDataPool.Shared;
|
||||
divisorData = divisorPool.FromModulus(prime64);
|
||||
ulong order64 = CalculateGpu(gpu, prime64, previous, divisorData, config);
|
||||
divisorPool.Return(divisorData);
|
||||
result = order64 == 0UL ? UInt128.Zero : (UInt128)order64;
|
||||
}
|
||||
else
|
||||
{
|
||||
divisorData = MontgomeryDivisorData.Empty;
|
||||
result = CalculateWideInternalGpu(gpu, prime, previousOrder, divisorData, config);
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
private static UInt128 CalculateWideInternalGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 prime, in UInt128? previousOrder, in MontgomeryDivisorData? divisorData, in PrimeOrderSearchConfig config)
|
||||
{
|
||||
if (prime <= UInt128.One)
|
||||
{
|
||||
return UInt128.One;
|
||||
}
|
||||
|
||||
if (prime == (UInt128)3UL)
|
||||
{
|
||||
return (UInt128)2UL;
|
||||
}
|
||||
|
||||
UInt128 phi = prime - UInt128.One;
|
||||
PartialFactorResult128 phiFactors = PartialFactorWide(phi, config);
|
||||
UInt128 result;
|
||||
if (phiFactors.Factors is null)
|
||||
{
|
||||
result = FinishStrictlyWideGpu(gpu, prime, divisorData);
|
||||
}
|
||||
else
|
||||
{
|
||||
result = RunHeuristicPipelineWideGpu(gpu, prime, previousOrder, divisorData, config, phi, phiFactors);
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
private static UInt128 RunHeuristicPipelineWideGpu(
|
||||
PrimeOrderCalculatorAccelerator gpu,
|
||||
in UInt128 prime,
|
||||
in UInt128? previousOrder,
|
||||
in MontgomeryDivisorData? divisorData,
|
||||
in PrimeOrderSearchConfig config,
|
||||
in UInt128 phi,
|
||||
in PartialFactorResult128 phiFactors)
|
||||
{
|
||||
if (phiFactors.FullyFactored && TrySpecialMaxWideGpu(gpu, phi, prime, divisorData, phiFactors))
|
||||
{
|
||||
return phi;
|
||||
}
|
||||
|
||||
UInt128 candidateOrder = InitializeStartingOrderWideGpu(gpu, prime, phi, divisorData);
|
||||
candidateOrder = ExponentLoweringWideGpu(gpu, candidateOrder, prime, divisorData, phiFactors);
|
||||
|
||||
if (TryConfirmOrderWideGpu(gpu, prime, candidateOrder, divisorData, config))
|
||||
{
|
||||
return candidateOrder;
|
||||
}
|
||||
|
||||
if (config.Mode == PrimeOrderMode.Strict)
|
||||
{
|
||||
return FinishStrictlyWideGpu(gpu, prime, divisorData);
|
||||
}
|
||||
|
||||
if (TryHeuristicFinishWideGpu(gpu, prime, candidateOrder, previousOrder, divisorData, config, out UInt128 order))
|
||||
{
|
||||
return order;
|
||||
}
|
||||
|
||||
return candidateOrder;
|
||||
}
|
||||
|
||||
private static UInt128 FinishStrictlyWideGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 prime, in MontgomeryDivisorData? divisorData)
|
||||
{
|
||||
UInt128 phi = prime - UInt128.One;
|
||||
Dictionary<UInt128, int> counts = new(capacity: 8);
|
||||
FactorCompletelyWide(phi, counts);
|
||||
if (counts.Count == 0)
|
||||
{
|
||||
return phi;
|
||||
}
|
||||
|
||||
List<KeyValuePair<UInt128, int>> entries = new(counts);
|
||||
entries.Sort(static (a, b) => a.Key.CompareTo(b.Key));
|
||||
|
||||
UInt128 order = phi;
|
||||
int entryCount = entries.Count;
|
||||
for (int i = 0; i < entryCount; i++)
|
||||
{
|
||||
UInt128 primeFactor = entries[i].Key;
|
||||
int exponent = entries[i].Value;
|
||||
for (int iteration = 0; iteration < exponent; iteration++)
|
||||
{
|
||||
if ((order % primeFactor) != UInt128.Zero)
|
||||
{
|
||||
UInt128 candidate = order / primeFactor;
|
||||
if (Pow2ModWideGpu(gpu, candidate, prime, divisorData) == UInt128.One)
|
||||
{
|
||||
order = candidate;
|
||||
}
|
||||
else
|
||||
{
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return order;
|
||||
}
|
||||
|
||||
private static bool TrySpecialMaxWideGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 phi, in UInt128 prime, in MontgomeryDivisorData? divisorData, in PartialFactorResult128 factors)
|
||||
{
|
||||
ReadOnlySpan<FactorEntry128> factorSpan = factors.Factors;
|
||||
int length = factors.Count;
|
||||
for (int i = 0; i < length; i++)
|
||||
{
|
||||
UInt128 factor = factorSpan[i].Value;
|
||||
UInt128 reduced = phi / factor;
|
||||
if (Pow2ModWideGpu(gpu, reduced, prime, divisorData) == UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
private static UInt128 InitializeStartingOrderWideGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 prime, in UInt128 phi, in MontgomeryDivisorData? divisorData)
|
||||
{
|
||||
UInt128 order = phi;
|
||||
UInt128 mod8 = prime & (UInt128)7UL;
|
||||
if (mod8 == UInt128.One || mod8 == (UInt128)7UL)
|
||||
{
|
||||
UInt128 half = phi >> 1;
|
||||
if (Pow2ModWideGpu(gpu, half, prime, divisorData) == UInt128.One)
|
||||
{
|
||||
order = half;
|
||||
}
|
||||
}
|
||||
|
||||
return order;
|
||||
}
|
||||
|
||||
private static UInt128 ExponentLoweringWideGpu(PrimeOrderCalculatorAccelerator gpu, UInt128 order, in UInt128 prime, in MontgomeryDivisorData? divisorData, in PartialFactorResult128 factors)
|
||||
{
|
||||
ArrayPool<FactorEntry128> pool = ThreadStaticPools.FactorEntry128Pool;
|
||||
ReadOnlySpan<FactorEntry128> factorSpan = factors.Factors;
|
||||
int length = factors.Count;
|
||||
// TODO: This should never trigger from production code - check
|
||||
// if (length == 0)
|
||||
// {
|
||||
// return order;
|
||||
// }
|
||||
FactorEntry128[] tempArray = pool.Rent(length + 1);
|
||||
|
||||
Span<FactorEntry128> buffer = tempArray.AsSpan(0, length);
|
||||
factorSpan.CopyTo(buffer);
|
||||
|
||||
if (!factors.FullyFactored && factors.Cofactor > UInt128.One && IsPrimeWide(factors.Cofactor))
|
||||
{
|
||||
buffer[length] = new FactorEntry128(factors.Cofactor, 1);
|
||||
length++;
|
||||
}
|
||||
|
||||
buffer[..length].Sort(static (a, b) => a.Value.CompareTo(b.Value));
|
||||
|
||||
// TODO: Investigating using stepper to replace modulo operation
|
||||
for (int i = 0; i < length; i++)
|
||||
{
|
||||
UInt128 primeFactor = buffer[i].Value;
|
||||
int exponent = buffer[i].Exponent;
|
||||
for (int iteration = 0; iteration < exponent; iteration++)
|
||||
{
|
||||
if ((order % primeFactor) == UInt128.Zero)
|
||||
{
|
||||
UInt128 reduced = order / primeFactor;
|
||||
if (Pow2ModWideGpu(gpu, reduced, prime, divisorData) == UInt128.One)
|
||||
{
|
||||
order = reduced;
|
||||
continue;
|
||||
}
|
||||
}
|
||||
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
pool.Return(tempArray, clearArray: false);
|
||||
return order;
|
||||
}
|
||||
|
||||
private static bool TryConfirmOrderWideGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 prime, in UInt128 order, in MontgomeryDivisorData? divisorData, in PrimeOrderSearchConfig config)
|
||||
{
|
||||
if (order == UInt128.Zero)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (Pow2ModWideGpu(gpu, order, prime, divisorData) != UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
PartialFactorResult128 factorization = PartialFactorWide(order, config);
|
||||
if (factorization.Factors is null)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!factorization.FullyFactored)
|
||||
{
|
||||
if (factorization.Cofactor <= UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!IsPrimeWide(factorization.Cofactor))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
factorization = factorization.WithAdditionalPrime(factorization.Cofactor);
|
||||
}
|
||||
|
||||
ReadOnlySpan<FactorEntry128> span = factorization.Factors;
|
||||
int length = factorization.Count;
|
||||
for (int i = 0; i < length; i++)
|
||||
{
|
||||
UInt128 primeFactor = span[i].Value;
|
||||
UInt128 reduced = order;
|
||||
for (int iteration = 0; iteration < span[i].Exponent; iteration++)
|
||||
{
|
||||
if ((reduced % primeFactor) != UInt128.Zero)
|
||||
{
|
||||
break;
|
||||
}
|
||||
|
||||
reduced /= primeFactor;
|
||||
if (Pow2ModWideGpu(gpu, reduced, prime, divisorData) == UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
private static bool TryHeuristicFinishWideGpu(
|
||||
PrimeOrderCalculatorAccelerator gpu,
|
||||
in UInt128 prime,
|
||||
in UInt128 order,
|
||||
in UInt128? previousOrder,
|
||||
in MontgomeryDivisorData? divisorData,
|
||||
in PrimeOrderSearchConfig config,
|
||||
out UInt128 result)
|
||||
{
|
||||
result = UInt128.Zero;
|
||||
if (order <= UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
PartialFactorResult128 orderFactors = PartialFactorWide(order, config);
|
||||
if (orderFactors.Factors is null)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!orderFactors.FullyFactored)
|
||||
{
|
||||
if (orderFactors.Cofactor <= UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!IsPrimeWide(orderFactors.Cofactor))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
orderFactors = orderFactors.WithAdditionalPrime(orderFactors.Cofactor);
|
||||
}
|
||||
|
||||
int capacity = config.MaxPowChecks <= 0 ? 64 : config.MaxPowChecks * 4;
|
||||
List<UInt128> candidates = new(capacity);
|
||||
FactorEntry128[] factorArray = orderFactors.Factors!;
|
||||
BuildCandidatesWide(order, factorArray, orderFactors.Count, candidates, capacity);
|
||||
if (candidates.Count == 0)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
SortCandidatesWide(prime, previousOrder, candidates);
|
||||
|
||||
int powBudget = config.MaxPowChecks <= 0 ? candidates.Count : config.MaxPowChecks;
|
||||
int powUsed = 0;
|
||||
int candidateCount = candidates.Count;
|
||||
|
||||
for (int i = 0; i < candidateCount; i++)
|
||||
{
|
||||
if (powUsed >= powBudget && powBudget > 0)
|
||||
{
|
||||
break;
|
||||
}
|
||||
|
||||
UInt128 candidate = candidates[i];
|
||||
powUsed++;
|
||||
|
||||
if (Pow2ModWideGpu(gpu, candidate, prime, divisorData) != UInt128.One)
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
if (!TryConfirmCandidateWideGpu(gpu, prime, candidate, config, ref powUsed, powBudget, divisorData))
|
||||
{
|
||||
continue;
|
||||
}
|
||||
|
||||
result = candidate;
|
||||
return true;
|
||||
}
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
private static ulong CalculateByFactorizationGpu(PrimeOrderCalculatorAccelerator gpu, ulong prime, in MontgomeryDivisorData divisorData)
|
||||
{
|
||||
ulong phi = prime - 1UL;
|
||||
Dictionary<ulong, int> counts = gpu.Pow2ModEntriesToTestOnHost;
|
||||
counts.Clear();
|
||||
|
||||
FactorCompletelyCpu(gpu, phi, counts);
|
||||
FactorCompletelyCpu(phi, counts, false);
|
||||
if (counts.Count == 0)
|
||||
{
|
||||
return phi;
|
||||
@@ -156,6 +548,86 @@ internal static partial class PrimeOrderCalculator
|
||||
return result != 0;
|
||||
}
|
||||
|
||||
private static UInt128 Pow2ModWideGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 exponent, in UInt128 modulus, in MontgomeryDivisorData? divisorData)
|
||||
{
|
||||
if (modulus == UInt128.One)
|
||||
{
|
||||
return UInt128.Zero;
|
||||
}
|
||||
|
||||
if (modulus <= (UInt128)ulong.MaxValue && exponent <= (UInt128)ulong.MaxValue)
|
||||
{
|
||||
ulong prime64 = (ulong)modulus;
|
||||
ulong exponent64 = (ulong)exponent;
|
||||
GpuPow2ModStatus status = PrimeOrderGpuHeuristics.TryPow2Mod(gpu, exponent64, prime64, out ulong remainder, divisorData);
|
||||
if (status == GpuPow2ModStatus.Success)
|
||||
{
|
||||
return remainder;
|
||||
}
|
||||
}
|
||||
|
||||
GpuPow2ModStatus wideStatus = PrimeOrderGpuHeuristics.TryPow2Mod(gpu, exponent, modulus, out UInt128 wideRemainder);
|
||||
if (wideStatus == GpuPow2ModStatus.Success)
|
||||
{
|
||||
return wideRemainder;
|
||||
}
|
||||
|
||||
return exponent.Pow2MontgomeryModWindowed(modulus);
|
||||
}
|
||||
|
||||
private static bool TryConfirmCandidateWideGpu(PrimeOrderCalculatorAccelerator gpu, in UInt128 prime, in UInt128 candidate, in PrimeOrderSearchConfig config, ref int powUsed, int powBudget, in MontgomeryDivisorData? divisorData)
|
||||
{
|
||||
PartialFactorResult128 factorization = PartialFactorWide(candidate, config);
|
||||
if (factorization.Factors is null)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!factorization.FullyFactored)
|
||||
{
|
||||
if (factorization.Cofactor <= UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
if (!IsPrimeWide(factorization.Cofactor))
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
factorization = factorization.WithAdditionalPrime(factorization.Cofactor);
|
||||
}
|
||||
|
||||
ReadOnlySpan<FactorEntry128> span = factorization.Factors;
|
||||
int length = factorization.Count;
|
||||
for (int i = 0; i < length; i++)
|
||||
{
|
||||
UInt128 primeFactor = span[i].Value;
|
||||
UInt128 reduced = candidate;
|
||||
for (int iteration = 0; iteration < span[i].Exponent; iteration++)
|
||||
{
|
||||
if ((reduced % primeFactor) != UInt128.Zero)
|
||||
{
|
||||
break;
|
||||
}
|
||||
|
||||
reduced /= primeFactor;
|
||||
if (powUsed >= powBudget && powBudget > 0)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
|
||||
powUsed++;
|
||||
if (Pow2ModWideGpu(gpu, reduced, prime, divisorData) == UInt128.One)
|
||||
{
|
||||
return false;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
private static bool TryPollardRhoGpu(PrimeOrderCalculatorAccelerator gpu, ulong n, out ulong factor)
|
||||
{
|
||||
Span<ulong> randomStateSpan = stackalloc ulong[1];
|
||||
@@ -184,6 +656,5 @@ internal static partial class PrimeOrderCalculator
|
||||
factor = factored ? factorSpan[0] : 0UL;
|
||||
return factored;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
Reference in new issue
Block a user