Split GPU heuristic tables by scope

This commit is contained in:
Marek Lemańczyk committed 2025-11-01 09:22:44 +01:00
1 parent 936ce239db
commit 6027a73839
6 files changed
+574 -392

No files matched your search

@@ -8,83 +8,21 @@ internal readonly struct HeuristicGpuDivisorTables
{
public static readonly HeuristicGpuDivisorTables Shared = default;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding1;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding3;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding7;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding9;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding1;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding3;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding7;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorSquaresEnding9;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisors;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupADivisorSquares;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding1;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding1;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding7;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding7;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorsEnding9;
public readonly ArrayView1D<ulong, Stride1D.Dense> GroupBDivisorSquaresEnding9;
public HeuristicGpuDivisorTables(
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding1,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding3,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding7,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding9,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding1,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding3,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding7,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorSquaresEnding9,
ArrayView1D<ulong, Stride1D.Dense> groupADivisors,
ArrayView1D<ulong, Stride1D.Dense> groupADivisorSquares,
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding1,
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding1,
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding7,
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding7,
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding9,
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorSquaresEnding9)
ArrayView1D<ulong, Stride1D.Dense> groupBDivisorsEnding9)
{
CombinedDivisorsEnding1 = combinedDivisorsEnding1;
CombinedDivisorsEnding3 = combinedDivisorsEnding3;
CombinedDivisorsEnding7 = combinedDivisorsEnding7;
CombinedDivisorsEnding9 = combinedDivisorsEnding9;
CombinedDivisorSquaresEnding1 = combinedDivisorSquaresEnding1;
CombinedDivisorSquaresEnding3 = combinedDivisorSquaresEnding3;
CombinedDivisorSquaresEnding7 = combinedDivisorSquaresEnding7;
CombinedDivisorSquaresEnding9 = combinedDivisorSquaresEnding9;
GroupADivisors = groupADivisors;
GroupADivisorSquares = groupADivisorSquares;
GroupBDivisorsEnding1 = groupBDivisorsEnding1;
GroupBDivisorSquaresEnding1 = groupBDivisorSquaresEnding1;
GroupBDivisorsEnding7 = groupBDivisorsEnding7;
GroupBDivisorSquaresEnding7 = groupBDivisorSquaresEnding7;
GroupBDivisorsEnding9 = groupBDivisorsEnding9;
GroupBDivisorSquaresEnding9 = groupBDivisorSquaresEnding9;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ArrayView1D<ulong, Stride1D.Dense> SelectCombined(byte ending)
{
return ending switch
{
1 => CombinedDivisorsEnding1,
3 => CombinedDivisorsEnding3,
7 => CombinedDivisorsEnding7,
9 => CombinedDivisorsEnding9,
_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
};
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ArrayView1D<ulong, Stride1D.Dense> SelectCombinedSquares(byte ending)
{
return ending switch
{
1 => CombinedDivisorSquaresEnding1,
3 => CombinedDivisorSquaresEnding3,
7 => CombinedDivisorSquaresEnding7,
9 => CombinedDivisorSquaresEnding9,
_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
};
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
@@ -99,18 +37,6 @@ internal readonly struct HeuristicGpuDivisorTables
};
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ArrayView1D<ulong, Stride1D.Dense> SelectGroupBSquares(byte ending)
{
return ending switch
{
1 => GroupBDivisorSquaresEnding1,
7 => GroupBDivisorSquaresEnding7,
9 => GroupBDivisorSquaresEnding9,
_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
};
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ArrayView1D<ulong, Stride1D.Dense> SelectDivisors(HeuristicGpuDivisorTableKind kind, byte ending)
{
@@ -120,21 +46,6 @@ internal readonly struct HeuristicGpuDivisorTables
HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorsEnding1,
HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorsEnding7,
HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorsEnding9,
HeuristicGpuDivisorTableKind.Combined => SelectCombined(ending),
_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
};
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ArrayView1D<ulong, Stride1D.Dense> SelectDivisorSquares(HeuristicGpuDivisorTableKind kind, byte ending)
{
return kind switch
{
HeuristicGpuDivisorTableKind.GroupA => GroupADivisorSquares,
HeuristicGpuDivisorTableKind.GroupBEnding1 => GroupBDivisorSquaresEnding1,
HeuristicGpuDivisorTableKind.GroupBEnding7 => GroupBDivisorSquaresEnding7,
HeuristicGpuDivisorTableKind.GroupBEnding9 => GroupBDivisorSquaresEnding9,
HeuristicGpuDivisorTableKind.Combined => SelectCombinedSquares(ending),
_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
};
}
@@ -144,3 +55,43 @@ internal readonly struct HeuristicGpuDivisorTables
Unsafe.AsRef(in Shared) = tables;
}
}
internal readonly struct HeuristicGpuCombinedDivisorTables
{
public static readonly HeuristicGpuCombinedDivisorTables Shared = default;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding1;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding3;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding7;
public readonly ArrayView1D<ulong, Stride1D.Dense> CombinedDivisorsEnding9;
public HeuristicGpuCombinedDivisorTables(
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding1,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding3,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding7,
ArrayView1D<ulong, Stride1D.Dense> combinedDivisorsEnding9)
{
CombinedDivisorsEnding1 = combinedDivisorsEnding1;
CombinedDivisorsEnding3 = combinedDivisorsEnding3;
CombinedDivisorsEnding7 = combinedDivisorsEnding7;
CombinedDivisorsEnding9 = combinedDivisorsEnding9;
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public ArrayView1D<ulong, Stride1D.Dense> SelectCombined(byte ending)
{
return ending switch
{
1 => CombinedDivisorsEnding1,
3 => CombinedDivisorsEnding3,
7 => CombinedDivisorsEnding7,
9 => CombinedDivisorsEnding9,
_ => ArrayView1D<ulong, Stride1D.Dense>.Empty,
};
}
internal static void InitializeShared(in HeuristicGpuCombinedDivisorTables tables)
{
Unsafe.AsRef(in Shared) = tables;
}
}
@@ -82,21 +82,22 @@ internal static class PrimeTesterKernels
Index1D index,
ArrayView<int> resultFlag,
ulong n,
ulong maxDivisorSquare,
ulong maxDivisor,
HeuristicGpuDivisorTableKind tableKind,
HeuristicGpuDivisorTables tables)
HeuristicGpuDivisorTables tables,
HeuristicGpuCombinedDivisorTables combinedTables)
{
byte nMod10 = (byte)(n % 10UL);
ArrayView1D<ulong, Stride1D.Dense> divisors = tables.SelectDivisors(tableKind, nMod10);
ArrayView1D<ulong, Stride1D.Dense> divisorSquares = tables.SelectDivisorSquares(tableKind, nMod10);
ArrayView1D<ulong, Stride1D.Dense> divisors = tableKind == HeuristicGpuDivisorTableKind.Combined
? combinedTables.SelectCombined(nMod10)
: tables.SelectDivisors(tableKind, nMod10);
int divisorLength = (int)divisorSquares.Length;
int divisorLength = (int)divisors.Length;
int threadIndex = index;
bool skipThread = threadIndex >= divisorLength;
ulong divisorSquare = skipThread ? 0UL : divisorSquares[threadIndex];
skipThread = skipThread | (divisorSquare > maxDivisorSquare);
ulong divisor = skipThread ? 1UL : divisors[threadIndex];
skipThread = skipThread | (divisor > maxDivisor);
ulong remainder = skipThread ? 1UL : n % divisor;
bool hasFactor = !skipThread & (remainder == 0UL);
if (hasFactor)