daily update on 20251922

This commit is contained in:
Youmu Liu
2025-09-22 21:33:43 +08:00
parent b8a51c13dd
commit 8abc82ae08
5 changed files with 35 additions and 41 deletions
+20 -28
View File
@@ -41,20 +41,26 @@ using System.Threading.Tasks;
namespace StgSharp.HighPerformance
{
[StructLayout(LayoutKind.Explicit, Size = 64)]
[StructLayout(LayoutKind.Sequential)]
public unsafe struct MatrixKernel<T> where T: unmanaged, INumber<T>
{
[FieldOffset(0)] internal fixed byte Buffer[64];
[FieldOffset(0)] internal fixed double FP64Buffer[8];
[FieldOffset(0)] internal fixed float FP32Buffer[16];
[FieldOffset(0)] public M128 XMM0;
[FieldOffset(16)] public M128 XMM1;
[FieldOffset(32)] public M128 XMM2;
[FieldOffset(48)] public M128 XMM3;
[FieldOffset(0)] public M256 YMM0;
[FieldOffset(32)] public M256 YMM1;
[FieldOffset(64)] public M512 ZMM0;
private T c0r0;
private T c0r1;
private T c0r2;
private T c0r3;
private T c1r0;
private T c1r1;
private T c1r2;
private T c1r3;
private T c2r0;
private T c2r1;
private T c2r2;
private T c2r3;
private T c3r0;
private T c3r1;
private T c3r2;
private T c3r3;
public MatrixKernel()
{
@@ -68,24 +74,10 @@ namespace StgSharp.HighPerformance
}
}
public static MatrixKernel<T> Unit
public ref T this[int column, int row]
{
get
{
MatrixKernel<T> kernel = new(true);
kernel.FP32Buffer[0] = 1f;
kernel.FP32Buffer[3] = 1f;
kernel.FP32Buffer[8] = 1f;
kernel.FP32Buffer[15] = 1f;
return kernel;
}
}
public ref T UnsafeIndex(int column, int row)
{
ref M128 col = ref Unsafe.As<byte, M128>(ref Buffer[0]);
col = ref Unsafe.Add(ref col, column);
return ref col.Member<T>(row);
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get => ref Unsafe.Add(ref c0r0, (column * 4) + row);
}
}
@@ -49,8 +49,11 @@ namespace StgSharp.Mathematics
delegate*<MatrixKernel<T>*, MatrixKernel<T>*, MatrixKernel<T>*, void> Operation)
where T: unmanaged, INumber<T>
{
int columnCount = left.PrimSize;
int rowCount = right.PrimSize;
int length = _threads.Length,
columnCount = left.PrimSize,
rowCount = right.PrimSize,
cAverage = left.PrimSize / length, cRemain = left.PrimSize % length,
rAverage = right.PrimSize / length, rRemain = right.PrimSize % length;
MatrixParallelTaskPackage<T> package = new()
{