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

Split Pow2 Montgomery GPU kernels by Montgomery conversion

Browse Source
This commit is contained in:
Marek Lemańczyk committed 2025-10-22 01:43:02 +02:00
1 parent b08175dadf
commit f12919e220
2 files changed
+43 -13
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

PerfectNumbers.Core/Gpu/Kernels/Pow2MontgomeryKernels.cs
+11 -3
View File
@@ -6,13 +6,21 @@ namespace PerfectNumbers.Core.Gpu;
internal static class Pow2MontgomeryKernels internal static class Pow2MontgomeryKernels
{ {
public static void Pow2MontgomeryKernel( public static void Pow2MontgomeryKernelKeepMontgomery(
Index1D index, Index1D index,
ArrayView1D<ulong, Stride1D.Dense> exponents, ArrayView1D<ulong, Stride1D.Dense> exponents,
MontgomeryDivisorData divisor, MontgomeryDivisorData divisor,
byte keepMontgomery,
ArrayView1D<ulong, Stride1D.Dense> results) ArrayView1D<ulong, Stride1D.Dense> results)
{ {
results[index] = ULongExtensions.Pow2MontgomeryModWindowedGpu(divisor, exponents[index], keepMontgomery != 0); results[index] = ULongExtensions.Pow2MontgomeryModWindowedGpu(divisor, exponents[index], keepMontgomery: true);
}
public static void Pow2MontgomeryKernelConvertToStandard(
Index1D index,
ArrayView1D<ulong, Stride1D.Dense> exponents,
MontgomeryDivisorData divisor,
ArrayView1D<ulong, Stride1D.Dense> results)
{
results[index] = ULongExtensions.Pow2MontgomeryModWindowedGpu(divisor, exponents[index], keepMontgomery: false);
} }
} }
PerfectNumbers.Core/ULongExtensions.Gpu.cs
+32 -10
View File
@@ -303,7 +303,7 @@ public static partial class ULongExtensions
private static class Pow2MontgomeryGpuExecutor private static class Pow2MontgomeryGpuExecutor
{ {
private static readonly ConcurrentDictionary<Accelerator, Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, byte, ArrayView1D<ulong, Stride1D.Dense>>> KernelCache = new(); private static readonly ConcurrentDictionary<Accelerator, Pow2MontgomeryKernelGroup> KernelCache = new();
public static ulong Execute(ulong exponent, in MontgomeryDivisorData divisor, bool keepMontgomery) public static ulong Execute(ulong exponent, in MontgomeryDivisorData divisor, bool keepMontgomery)
{ {
@@ -319,12 +319,7 @@ public static partial class ULongExtensions
// return 0UL; // return 0UL;
// } // }
var kernel = KernelCache.GetOrAdd(accelerator, static accel => var kernelGroup = KernelCache.GetOrAdd(accelerator, static accel => Pow2MontgomeryKernelGroup.Create(accel));
{
var loaded = accel.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, byte, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernel);
var launcher = KernelUtil.GetKernel(loaded);
return launcher.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, byte, ArrayView1D<ulong, Stride1D.Dense>>>();
});
var exponentBuffer = accelerator.Allocate1D<ulong>(1); var exponentBuffer = accelerator.Allocate1D<ulong>(1);
var resultBuffer = accelerator.Allocate1D<ulong>(1); var resultBuffer = accelerator.Allocate1D<ulong>(1);
@@ -333,10 +328,9 @@ public static partial class ULongExtensions
// We don't need to worry about any left-overs here. // We don't need to worry about any left-overs here.
// resultBuffer.MemSetToZero(); // resultBuffer.MemSetToZero();
// TODO: Have 2 kernels for both cases to have branch-less solution. var kernel = keepMontgomery ? kernelGroup.KeepMontgomery : kernelGroup.ConvertToStandard;
byte keepFlag = keepMontgomery ? (byte)1 : (byte)0;
AcceleratorStream stream = lease.Stream; AcceleratorStream stream = lease.Stream;
kernel(stream, 1, exponentBuffer.View, divisor, keepFlag, resultBuffer.View); kernel(stream, 1, exponentBuffer.View, divisor, resultBuffer.View);
stream.Synchronize(); stream.Synchronize();
resultBuffer.View.CopyToCPU(ref result, 1); resultBuffer.View.CopyToCPU(ref result, 1);
@@ -357,5 +351,33 @@ public static partial class ULongExtensions
// Intentionally avoid exception handling here; any accelerator failure should crash the scanner. // Intentionally avoid exception handling here; any accelerator failure should crash the scanner.
return result; return result;
} }
private sealed class Pow2MontgomeryKernelGroup
{
internal Pow2MontgomeryKernelGroup(
Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> keepMontgomery,
Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> convertToStandard)
{
KeepMontgomery = keepMontgomery;
ConvertToStandard = convertToStandard;
}
internal Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> KeepMontgomery { get; }
internal Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>> ConvertToStandard { get; }
internal static Pow2MontgomeryKernelGroup Create(Accelerator accelerator)
{
var keepKernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelKeepMontgomery);
var keepLauncher = KernelUtil.GetKernel(keepKernel);
var keepDelegate = keepLauncher.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
var convertKernel = accelerator.LoadAutoGroupedStreamKernel<Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>(Pow2MontgomeryKernels.Pow2MontgomeryKernelConvertToStandard);
var convertLauncher = KernelUtil.GetKernel(convertKernel);
var convertDelegate = convertLauncher.CreateLauncherDelegate<Action<AcceleratorStream, Index1D, ArrayView1D<ulong, Stride1D.Dense>, MontgomeryDivisorData, ArrayView1D<ulong, Stride1D.Dense>>>();
return new Pow2MontgomeryKernelGroup(keepDelegate, convertDelegate);
}
}
} }
} }