youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Daily update on 20260111-005714 by Nitload Youmu. Additional Message : Use optimized tile operation to accelerate matrix mul, still has problem

Browse Source 
Signed-off-by: Nitload <YoumuLiu_NSI@whv0t.onmicrosoft.com>
This commit is contained in:
Nitload
2026-01-11 00:57:14 +08:00
parent 6827259891
commit 4be5c1c7c9
9 changed files with 141 additions and 219 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixCompute.cs
+27 -28
View File
@@ -1,9 +1,9 @@
//-----------------------------------------------------------------------
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixCompute"
// Project: StgSharp
// AuthorGroup: Nitload
// Copyright (c) Nitload. All rights reserved.
// AuthorGroup: Nitload Space
// Copyright (c) Nitload Space. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
@@ -40,9 +40,7 @@ namespace StgSharp.Mathematics.Numeric
public partial class MatrixCompute
{
internal static unsafe void ComputeFmaPanelF32(
MatrixParallelTaskPackage* p
)
internal static unsafe void ComputeFmaPanelF32(MatrixParallelTaskPackage* p)
{
int f32 = (int)MatrixElementType.F32;
MatrixKernel* leftBuffer = (MatrixKernel*)p->Left;
@@ -64,19 +62,14 @@ namespace StgSharp.Mathematics.Numeric
for (int k = 0; k < commonK; k++) {
NativeIntrinsic.Context
.mat[f32].kernel_fma(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<float>*)ansBuffer,
ansHeight, i, j));
GetKernelAddressUnsafe((MatrixKernel<float>*)leftBuffer, ansHeight, k, j),
GetKernelAddressUnsafe((MatrixKernel<float>*)rightBuffer, commonK, j, k),
GetKernelAddressUnsafe((MatrixKernel<float>*)ansBuffer, ansHeight, i, j));
}
}
}
internal static unsafe void ComputeFmaPanelF64(
MatrixParallelTaskPackage* p
)
internal static unsafe void ComputeFmaPanelF64(MatrixParallelTaskPackage* p)
{
int f64 = (int)MatrixElementType.F64;
@@ -90,33 +83,39 @@ namespace StgSharp.Mathematics.Numeric
int ansHeight = p->AnsSecOffset;
int commonK = p->LeftPrimOffset;
M128* enumeration = stackalloc M128[1];
// in this occasion size must be 1
for (long c = offsetRef; c < offsetRef + lengthRef; c++)
{
int i = (int)(c / ansWidth);
int j = (int)(c % ansWidth);
enumeration->Member<int>(0) = i;
enumeration->Member<int>(1) = j;
enumeration->Member<int>(2) = commonK;
enumeration->Member<int>(3) = ansHeight;
NativeIntrinsic.Context.mat[f64].kernel_tile_fma(leftBuffer, rightBuffer, ansBuffer, enumeration);
/*
for (int k = 0; k < commonK; k += 1) {
NativeIntrinsic.Context
.mat[f64].kernel_fma(
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)leftBuffer,
ansHeight, k, j),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)rightBuffer,
commonK, j, k),
(MatrixPanel*)GetKernelAddressUnsafe((MatrixKernel<double>*)ansBuffer,
ansHeight, i, j));
GetKernelAddressUnsafe((MatrixKernel<double>*)leftBuffer, ansHeight, k, j),
GetKernelAddressUnsafe((MatrixKernel<double>*)rightBuffer, commonK, j, k),
GetKernelAddressUnsafe((MatrixKernel<double>*)ansBuffer, ansHeight, i, j));
}
/**/
}
}
[MethodImpl(MethodImplOptions.AggressiveInlining)]
internal static unsafe MatrixKernel<T>* GetKernelAddressUnsafe<T>(
MatrixKernel<T>* buffer,
int colLength,
int x,
int y
) where T : unmanaged, INumber<T>
internal static unsafe MatrixKernel* GetKernelAddressUnsafe<T>(
MatrixKernel<T>* buffer,
int colLength,
int x,
int y)
where T: unmanaged, INumber<T>
{
return buffer + (colLength * x + y);
return (MatrixKernel*)(buffer + (colLength * x + y));
}
}