Better L4 and array storage design

This commit is contained in:
Youmu Liu
2026-03-04 18:12:58 +08:00
parent 7e970352ab
commit dfa6f4ba92
103 changed files with 853 additions and 6837 deletions
@@ -31,33 +31,16 @@ using System.Runtime.InteropServices;
namespace StgSharp.Mathematics.Internal
{
[InlineArray(10)]
internal unsafe struct MatrixIntrinsic
{
private nint handle; //f32_panel_mul
public readonly nint this[
MatrixIntrinsicHandle handle
]
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get => this[(int)handle];
}
}
internal enum MatrixIntrinsicHandle : int
{
BufferAdd = 0,
BufferFill = 1,
BufferScalarMul = 2,
BufferSub = 3,
BufferTranspose = 4,
Add = 0,
Fill = 1,
ScalarMul = 2,
Sub = 3,
Transpose = 4,
ClearPanel = 5,
KerFma = 6,
TileFma = 7,
Fma = 6,
Pivot = 8,
}
@@ -66,18 +49,18 @@ namespace StgSharp.Mathematics.Internal
public struct MatrixElementType
{
public const int IntrinsicCount = 0xa;
public const int IntrinsicCount = 0x0a;
public static readonly MatrixElementType F32 = new(0x00010004);
public static readonly MatrixElementType F64 = new(0x00020008);
public static readonly MatrixElementType I08 = new(0x00030001);
public static readonly MatrixElementType I16 = new(0x00040002);
public static readonly MatrixElementType I32 = new(0x00050004);
public static readonly MatrixElementType I64 = new(0x00060008);
public static readonly MatrixElementType U08 = new(0x00070001);
public static readonly MatrixElementType U16 = new(0x00080002);
public static readonly MatrixElementType U32 = new(0x00090004);
public static readonly MatrixElementType U64 = new(0x000a0008);
public static readonly MatrixElementType F32 = new(0x00000004);
public static readonly MatrixElementType F64 = new(0x00010008);
public static readonly MatrixElementType I08 = new(0x00020001);
public static readonly MatrixElementType I16 = new(0x00030002);
public static readonly MatrixElementType I32 = new(0x00040004);
public static readonly MatrixElementType I64 = new(0x00050008);
public static readonly MatrixElementType U08 = new(0x00060001);
public static readonly MatrixElementType U16 = new(0x00070002);
public static readonly MatrixElementType U32 = new(0x00080004);
public static readonly MatrixElementType U64 = new(0x00090008);
[FieldOffset(0)] internal readonly int Mask;
[FieldOffset(2)] public readonly ushort IntrinsicNode;
@@ -96,14 +79,6 @@ namespace StgSharp.Mathematics.Internal
}
[InlineArray(MatrixElementType.IntrinsicCount)]
internal unsafe struct MatrixIntrinsicArray
{
private MatrixIntrinsic e0;
}
[StructLayout(LayoutKind.Sequential)]
internal unsafe struct IntrinsicContext
{
@@ -112,7 +87,8 @@ namespace StgSharp.Mathematics.Internal
public delegate* unmanaged[Cdecl]<int, ulong> factorial_simd; // factorial_simd
public delegate* unmanaged[Cdecl]<char*, int, int, int> index_pair; // index_pair
public delegate* unmanaged[Cdecl]<Vector4*, Vector4*, void> vec_f32_normalize_3; // f32_normalize_3
public MatrixIntrinsicArray mat; // mat[4] in native order
public MatrixMultiKernelIntrinsicArray mat_mk; // mat[4] in native order
public MatrixSingleKernelIntrinsicArray mat_sk; // mat[4] in native order
public IntrinsicContext()
{
@@ -120,4 +96,79 @@ namespace StgSharp.Mathematics.Internal
}
}
#region mk
[InlineArray(10)]
internal unsafe struct MatrixMultiKernelIntrinsic
{
private nint handle; //f32_panel_mul
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public readonly void Call(
MatrixIntrinsicHandle index,
MatrixParallelTask* task
)
{
nint ptr = this[(int)index];
((delegate* unmanaged[Cdecl]<MatrixParallelTask*, void>)ptr)(task);
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public readonly nint Get(
MatrixIntrinsicHandle index
)
{
return this[(int)index];
}
}
[InlineArray(MatrixElementType.IntrinsicCount)]
internal unsafe struct MatrixMultiKernelIntrinsicArray
{
private MatrixMultiKernelIntrinsic e0;
}
#endregion
#region sk
[InlineArray(10)]
internal unsafe struct MatrixSingleKernelIntrinsic
{
private nint handle; //f32_panel_mul
[MethodImpl(MethodImplOptions.AggressiveInlining)]
public readonly void Call(
MatrixIntrinsicHandle index,
MatrixKernel* left,
MatrixKernel* right,
MatrixKernel* ans,
ulong scalar
)
{
nint ptr = this[(int)index];
((delegate* unmanaged[Cdecl]<MatrixKernel*, MatrixKernel*, MatrixKernel*, ulong, void>)ptr)(left,
right,
ans,
scalar);
}
}
[InlineArray(MatrixElementType.IntrinsicCount)]
internal unsafe struct MatrixSingleKernelIntrinsicArray
{
private MatrixSingleKernelIntrinsic e0;
}
#endregion
}