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

GPU memory allocations protected with lock to prevent shared memory access issues. GPU usage limited for the best performance.

Browse Source
This commit is contained in:
mlemanczyk committed 2025-11-19 16:49:01 +01:00
1 parent 02d16bf05c
commit 5cac51479e
12 files changed
+236 -148
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/Accelerators/PrimeOrderCalculatorAccelerator.cs
+101 -65
View File
@@ -117,7 +117,11 @@ public sealed class PrimeOrderCalculatorAccelerator
public readonly Kernel CheckFactorsKernel;
public readonly Kernel ConvertToStandardKernel;
public readonly Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandardKernelLauncher;
public readonly Kernel HeuristicCombinedTrialDivisionKernel;
public readonly Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews> HeuristicCombinedTrialDivisionKernelLauncher;
public readonly Kernel KeepMontgomeryKernel;
public readonly Kernel PollardRhoKernel;
public readonly Kernel Pow2ModWideKernel;
@@ -133,29 +137,41 @@ public sealed class PrimeOrderCalculatorAccelerator
MemoryBuffer1D<KeyValuePair<ulong, int>, Stride1D.Dense> pow2ModEntriesToTestOnDevice = Pow2ModEntriesToTestOnDevice;
if (pow2ModEntriesToTestOnDevice.Length < factorsCount)
{
pow2ModEntriesToTestOnDevice.Dispose();
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
lock (accelerator)
{
pow2ModEntriesToTestOnDevice.Dispose();
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
}
}
MemoryBuffer1D<ulong, Stride1D.Dense> input = Input;
if (input.Length < primeTesterCapacity)
{
input.Dispose();
Input = accelerator.Allocate1D<ulong>(primeTesterCapacity);
lock (accelerator)
{
input.Dispose();
Input = accelerator.Allocate1D<ulong>(primeTesterCapacity);
}
}
MemoryBuffer1D<byte, Stride1D.Dense> outputByte = OutputByte;
if (outputByte.Length < primeTesterCapacity)
{
outputByte.Dispose();
OutputByte = accelerator.Allocate1D<byte>(primeTesterCapacity);
lock (accelerator)
{
outputByte.Dispose();
OutputByte = accelerator.Allocate1D<byte>(primeTesterCapacity);
}
}
MemoryBuffer1D<int, Stride1D.Dense> outputInt = OutputInt;
if (outputInt.Length < primeTesterCapacity)
{
outputInt.Dispose();
OutputInt = accelerator.Allocate1D<int>(primeTesterCapacity);
lock (accelerator)
{
outputInt.Dispose();
OutputInt = accelerator.Allocate1D<int>(primeTesterCapacity);
}
}
}
@@ -166,15 +182,21 @@ public sealed class PrimeOrderCalculatorAccelerator
MemoryBuffer1D<ulong, Stride1D.Dense> outputUlong = OutputUlong;
if (outputUlong.Length < newSize)
{
outputUlong.Dispose();
OutputUlong = accelerator.Allocate1D<ulong>(newSize);
lock (accelerator)
{
outputUlong.Dispose();
OutputUlong = accelerator.Allocate1D<ulong>(newSize);
}
}
MemoryBuffer1D<int, Stride1D.Dense> outputInt = OutputInt;
if (outputInt.Length < newSize)
{
outputInt.Dispose();
OutputInt = accelerator.Allocate1D<int>(newSize);
lock (accelerator)
{
outputInt.Dispose();
OutputInt = accelerator.Allocate1D<int>(newSize);
}
}
}
@@ -184,15 +206,21 @@ public sealed class PrimeOrderCalculatorAccelerator
MemoryBuffer1D<ulong, Stride1D.Dense> input = Input;
if (input.Length < newSize)
{
input.Dispose();
Input = accelerator.Allocate1D<ulong>(newSize);
lock (accelerator)
{
input.Dispose();
Input = accelerator.Allocate1D<ulong>(newSize);
}
}
MemoryBuffer1D<ulong, Stride1D.Dense> outputUlong = OutputUlong;
if (outputUlong.Length < newSize)
{
outputUlong.Dispose();
OutputUlong = accelerator.Allocate1D<ulong>(newSize);
lock (accelerator)
{
outputUlong.Dispose();
OutputUlong = accelerator.Allocate1D<ulong>(newSize);
}
}
}
@@ -201,57 +229,65 @@ public sealed class PrimeOrderCalculatorAccelerator
var accelerator = _accelerators[acceleratorIndex];
Accelerator = accelerator;
AcceleratorIndex = acceleratorIndex;
Input = accelerator.Allocate1D<ulong>(Math.Max(specialMaxFactorCapacity, primeTesterCapacity));
OutputByte = Accelerator.Allocate1D<byte>(primeTesterCapacity);
OutputInt = Accelerator.Allocate1D<int>(Math.Max(primeTesterCapacity, smallPrimeFactorSlotCount));
OutputUlong = accelerator.Allocate1D<ulong>(Math.Max(specialMaxFactorCapacity, smallPrimeFactorSlotCount));
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
SmallPrimeFactorCountSlot = accelerator.Allocate1D<int>(1);
SpecialMaxResult = accelerator.Allocate1D<ulong>(1);
LastDigitGpuTables lastDigitSharedTables = LastDigitGpuTables.GetStaticTables(acceleratorIndex);
DevicePrimesLastOne = lastDigitSharedTables.DevicePrimesLastOne.View;
DevicePrimesLastSeven = lastDigitSharedTables.DevicePrimesLastSeven.View;
DevicePrimesLastThree = lastDigitSharedTables.DevicePrimesLastThree.View;
DevicePrimesLastNine = lastDigitSharedTables.DevicePrimesLastNine.View;
DevicePrimesPow2LastOne = lastDigitSharedTables.DevicePrimesPow2LastOne.View;
DevicePrimesPow2LastSeven = lastDigitSharedTables.DevicePrimesPow2LastSeven.View;
DevicePrimesPow2LastThree = lastDigitSharedTables.DevicePrimesPow2LastThree.View;
DevicePrimesPow2LastNine = lastDigitSharedTables.DevicePrimesPow2LastNine.View;
DivisorTables = HeuristicCombinedGpuTables
.GetStaticTables(acceleratorIndex)
.CreateViews();
var smallPrimeFactorTables = SmallPrimeFactorGpuTables.GetStaticTables(acceleratorIndex);
if (smallPrimeFactorTables.HasValue)
lock (accelerator)
{
SmallPrimeFactorGpuTables tables = smallPrimeFactorTables.Value;
SmallPrimeFactorPrimes = tables.Primes.View;
SmallPrimeFactorSquares = tables.Squares.View;
Input = accelerator.Allocate1D<ulong>(Math.Max(specialMaxFactorCapacity, primeTesterCapacity));
OutputByte = Accelerator.Allocate1D<byte>(primeTesterCapacity);
OutputInt = Accelerator.Allocate1D<int>(Math.Max(primeTesterCapacity, smallPrimeFactorSlotCount));
OutputUlong = accelerator.Allocate1D<ulong>(Math.Max(specialMaxFactorCapacity, smallPrimeFactorSlotCount));
Pow2ModEntriesToTestOnDevice = accelerator.Allocate1D<KeyValuePair<ulong, int>>(factorsCount);
SmallPrimeFactorCountSlot = accelerator.Allocate1D<int>(1);
SpecialMaxResult = accelerator.Allocate1D<ulong>(1);
LastDigitGpuTables lastDigitSharedTables = LastDigitGpuTables.GetStaticTables(acceleratorIndex);
DevicePrimesLastOne = lastDigitSharedTables.DevicePrimesLastOne.View;
DevicePrimesLastSeven = lastDigitSharedTables.DevicePrimesLastSeven.View;
DevicePrimesLastThree = lastDigitSharedTables.DevicePrimesLastThree.View;
DevicePrimesLastNine = lastDigitSharedTables.DevicePrimesLastNine.View;
DevicePrimesPow2LastOne = lastDigitSharedTables.DevicePrimesPow2LastOne.View;
DevicePrimesPow2LastSeven = lastDigitSharedTables.DevicePrimesPow2LastSeven.View;
DevicePrimesPow2LastThree = lastDigitSharedTables.DevicePrimesPow2LastThree.View;
DevicePrimesPow2LastNine = lastDigitSharedTables.DevicePrimesPow2LastNine.View;
DivisorTables = HeuristicCombinedGpuTables
.GetStaticTables(acceleratorIndex)
.CreateViews();
var smallPrimeFactorTables = SmallPrimeFactorGpuTables.GetStaticTables(acceleratorIndex);
if (smallPrimeFactorTables.HasValue)
{
SmallPrimeFactorGpuTables tables = smallPrimeFactorTables.Value;
SmallPrimeFactorPrimes = tables.Primes.View;
SmallPrimeFactorSquares = tables.Squares.View;
}
CheckFactorsKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel));
Kernel convertToStandardKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelConvertToStandard));
ConvertToStandardKernel = convertToStandardKernel;
ConvertToStandardKernelLauncher = convertToStandardKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
var heuristicCombinedTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>(PrimeTesterKernels.HeuristicTrialCombinedDivisionKernel));
HeuristicCombinedTrialDivisionKernel = heuristicCombinedTrialDivisionKernel;
HeuristicCombinedTrialDivisionKernelLauncher = heuristicCombinedTrialDivisionKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>>();
KeepMontgomeryKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery));
PollardRhoKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.TryPollardRhoKernel));
Pow2ModWideKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(PrimeOrderGpuHeuristics.Pow2ModKernelWide));
SharesFactorKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel));
SmallPrimeFactorKernel = 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));
SmallPrimeSieveKernel = 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));
SpecialMaxKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel));
}
CheckFactorsKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<int, ulong, ArrayView1D<KeyValuePair<ulong, int>, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.CheckFactorsKernel));
ConvertToStandardKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelConvertToStandard));
HeuristicCombinedTrialDivisionKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, byte, ArrayView<int>, ulong, ulong, HeuristicCombinedGpuViews>(PrimeTesterKernels.HeuristicTrialCombinedDivisionKernel));
KeepMontgomeryKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery));
PollardRhoKernel = KernelUtil.GetKernel(accelerator.LoadStreamKernel<ulong, int, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.TryPollardRhoKernel));
Pow2ModWideKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>(PrimeOrderGpuHeuristics.Pow2ModKernelWide));
SharesFactorKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView<ulong>, ArrayView<byte>>(PrimeTesterKernels.SharesFactorKernel));
SmallPrimeFactorKernel = 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));
SmallPrimeSieveKernel = 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));
SpecialMaxKernel = KernelUtil.GetKernel(accelerator.LoadAutoGroupedStreamKernel<Index1D, ulong, ArrayView1D<ulong, Stride1D.Dense>, int, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>>(PrimeOrderGpuHeuristics.EvaluateSpecialMaxCandidatesKernel));
}
public void Dispose()
PerfectNumbers.Core/Gpu/GpuKernelPool.cs
+9 -9
View File
@@ -26,7 +26,7 @@ public class GpuKernelPool
internal static KernelContainer GetOrAddKernels(int acceleratorIndex, AcceleratorStream stream, KernelType kernelType)
{
var pool = _pool ??= [];
if (pool[acceleratorIndex] is not {} kernels)
if (pool[acceleratorIndex] is not { } kernels)
{
kernels = new();
}
@@ -151,7 +151,7 @@ public class GpuKernelPool
var host = MersenneDivisorCycles.Shared.ExportSmallCyclesSnapshot();
MemoryBuffer1D<ulong, Stride1D.Dense>? device;
// lock (accelerator)
lock (accelerator)
{
device = accelerator.Allocate1D<ulong>(host.Length);
}
@@ -195,13 +195,13 @@ public class GpuKernelPool
MemoryBuffer1D<ulong, Stride1D.Dense>? deviceLastOnePow2;
MemoryBuffer1D<ulong, Stride1D.Dense>? deviceLastSevenPow2;
// lock (accelerator)
// {
deviceLastOne = accelerator.Allocate1D<uint>(hostLastOne.Length);
deviceLastSeven = accelerator.Allocate1D<uint>(hostLastSeven.Length);
deviceLastOnePow2 = accelerator.Allocate1D<ulong>(hostLastOnePow2.Length);
deviceLastSevenPow2 = accelerator.Allocate1D<ulong>(hostLastSevenPow2.Length);
// }
lock (accelerator)
{
deviceLastOne = accelerator.Allocate1D<uint>(hostLastOne.Length);
deviceLastSeven = accelerator.Allocate1D<uint>(hostLastSeven.Length);
deviceLastOnePow2 = accelerator.Allocate1D<ulong>(hostLastOnePow2.Length);
deviceLastSevenPow2 = accelerator.Allocate1D<ulong>(hostLastSevenPow2.Length);
}
deviceLastOne.View.CopyFromCPU(stream, hostLastOne);
deviceLastSeven.View.CopyFromCPU(stream, hostLastSeven);
PerfectNumbers.Core/Gpu/MersenneNumberDivisorByDivisorGpuTester.cs
+3 -3
View File
@@ -540,7 +540,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
if (_divisorBuffer == null)
{
// lock(accelerator)
lock(accelerator)
{
_divisorBuffer = accelerator.Allocate1D<GpuDivisorPartialData>(1);
_offsetBuffer = accelerator.Allocate1D<int>(1);
@@ -570,7 +570,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
{
int desiredCapacity = requiredCapacity > _capacity ? requiredCapacity : _capacity;
_capacity = desiredCapacity;
// lock (accelerator)
lock (accelerator)
{
_exponentsBuffer = accelerator.Allocate1D<ulong>(desiredCapacity);
_hitBuffer = accelerator.Allocate1D<byte>(desiredCapacity);
@@ -798,7 +798,7 @@ public sealed class MersenneNumberDivisorByDivisorGpuTester : IMersenneNumberDiv
private void AllocateDeviceBuffers(Accelerator accelerator, int capacity)
{
// lock (accelerator)
lock (accelerator)
{
_divisorDataBuffer = accelerator.Allocate1D<GpuDivisorPartialData>(capacity);
_offsetBuffer = accelerator.Allocate1D<int>(capacity);
PerfectNumbers.Core/Gpu/Pow2MontgomeryGpuCalculator.cs
+9 -5
View File
@@ -55,8 +55,10 @@ public static class Pow2MontgomeryGpuCalculator
// We don't need to worry about any left-overs here.
// resultBuffer.MemSetToZero(stream);
var kernel = gpu.ConvertToStandardKernel;
kernel.Launch(stream, 1, 1, exponentBuffer.View, divisor, resultBuffer.View);
var kernelLauncher = gpu.ConvertToStandardKernelLauncher!;
kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
resultBuffer.View.CopyToCPU(stream, ref result, 1);
stream.Synchronize();
@@ -73,7 +75,6 @@ public static class Pow2MontgomeryGpuCalculator
ulong result = 0UL;
MemoryBuffer1D<ulong, Stride1D.Dense> exponentBuffer = gpu.Input;
MemoryBuffer1D<ulong, Stride1D.Dense> resultBuffer = gpu.OutputUlong;
var kernel = gpu.ConvertToStandardKernel;
int acceleratorIndex = gpu.AcceleratorIndex;
AcceleratorStream stream = AcceleratorStreamPool.Rent(acceleratorIndex);
@@ -81,8 +82,11 @@ public static class Pow2MontgomeryGpuCalculator
// We don't need to worry about any left-overs here.
// resultBuffer.MemSetToZero(stream);
var kernelLauncher = kernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
kernelLauncher(stream, 1,exponentBuffer.View, divisor, resultBuffer.View);
var accelerator = gpu.Accelerator;
var kernelLauncher = gpu.ConvertToStandardKernelLauncher;
kernelLauncher(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
resultBuffer.View.CopyToCPU(stream, ref result, 1);
stream.Synchronize();
PerfectNumbers.Core/Gpu/PrimeOrderGpuHeuristics.cs
+24 -14
View File
@@ -190,14 +190,14 @@ internal static partial class PrimeOrderGpuHeuristics
// fullyFactoredBuffer.MemSetToZero(stream);
var partialFactorKernel = GetPartialFactorKernel(acceleratorIndex);
var kernelLauncher = partialFactorKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, int, ulong, uint, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<int, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, ArrayView1D<byte, Stride1D.Dense>>>();
partialFactorKernel.Launch(
kernelLauncher(
stream,
1,
1,
smallPrimesView,
smallSquaresView,
smallPrimesView.Length,
(int)smallPrimesView.Length,
primeTargets.Length,
value,
limit,
@@ -354,7 +354,7 @@ internal static partial class PrimeOrderGpuHeuristics
MemoryBuffer1D<ulong, Stride1D.Dense>? resultBuffer;
MemoryBuffer1D<byte, Stride1D.Dense>? statusBuffer;
// lock(accelerator)
lock(accelerator)
{
phiFactorBuffer = accelerator.Allocate1D<ulong>(GpuSmallPrimeFactorSlots);
phiExponentBuffer = accelerator.Allocate1D<int>(GpuSmallPrimeFactorSlots);
@@ -398,16 +398,17 @@ internal static partial class PrimeOrderGpuHeuristics
statusBuffer.View);
var orderKernel = GetOrderKernel(acceleratorIndex);
orderKernel.Launch(
var kernelLauncher = orderKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ulong, OrderKernelConfig, MontgomeryDivisorData, ArrayView1D<uint, Stride1D.Dense>, ArrayView1D<ulong, Stride1D.Dense>, int, OrderKernelBuffers>>();
kernelLauncher(
stream,
1,
1,
prime,
kernelConfig,
divisorData,
smallPrimesView,
smallSquaresView,
smallPrimesView.Length,
(int)smallPrimesView.Length,
buffers);
@@ -517,7 +518,7 @@ internal static partial class PrimeOrderGpuHeuristics
// with out preallocated buffers. Share the pool with Pow2ModWide kernel.
MemoryBuffer1D<ulong, Stride1D.Dense>? exponentBuffer;
MemoryBuffer1D<ulong, Stride1D.Dense>? remainderBuffer;
// lock(accelerator)
lock(accelerator)
{
exponentBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
remainderBuffer = accelerator.Allocate1D<ulong>(exponents.Length);
@@ -529,9 +530,11 @@ 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>>>();
try
{
pow2ModKernel.Launch(stream, 1, exponents.Length, exponentBuffer.View, divisorData, remainderBuffer.View);
kernelLauncher(stream, exponents.Length, exponentBuffer.View, divisorData, remainderBuffer.View);
remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(results), exponents.Length);
stream.Synchronize();
@@ -590,13 +593,20 @@ internal static partial class PrimeOrderGpuHeuristics
int acceleratorIndex = gpu.AcceleratorIndex;
Accelerator accelerator = gpu.Accelerator;
MemoryBuffer1D<GpuUInt128, Stride1D.Dense> exponentBuffer = accelerator.Allocate1D<GpuUInt128>(length);
MemoryBuffer1D<GpuUInt128, Stride1D.Dense> remainderBuffer = accelerator.Allocate1D<GpuUInt128>(length);
AcceleratorStream? stream = AcceleratorStreamPool.Rent(acceleratorIndex);
exponentBuffer.View.CopyFromCPU(stream, exponentSpan);
MemoryBuffer1D<GpuUInt128, Stride1D.Dense> exponentBuffer;
MemoryBuffer1D<GpuUInt128, Stride1D.Dense> remainderBuffer;
gpu.Pow2ModWideKernel.Launch(stream, length, exponentBuffer.View, modulus, remainderBuffer.View);
lock(accelerator)
{
exponentBuffer = accelerator.Allocate1D<GpuUInt128>(length);
remainderBuffer = accelerator.Allocate1D<GpuUInt128>(length);
}
exponentBuffer.View.CopyFromCPU(stream, exponentSpan);
var kernelLauncher = gpu.Pow2ModWideKernel.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<GpuUInt128, Stride1D.Dense>, GpuUInt128, ArrayView1D<GpuUInt128, Stride1D.Dense>>>();
kernelLauncher(stream, length, exponentBuffer.View, modulus, remainderBuffer.View);
remainderBuffer.View.CopyToCPU(stream, ref MemoryMarshal.GetReference(resultSpan), length);
stream.Synchronize();