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.double.cs
+35 -38
View File
@@ -1,9 +1,9 @@
//-----------------------------------------------------------------------
//-----------------------------------------------------------------------
// -----------------------------------------------------------------------
// file="MatrixCompute.double"
// 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
@@ -38,11 +38,7 @@ namespace StgSharp.Mathematics.Numeric
public partial class MatrixCompute
{
public static unsafe void Add(
Matrix<double> left,
Matrix<double> right,
Matrix<double> ans
)
public static unsafe void Add(Matrix<double> left, Matrix<double> right, Matrix<double> ans)
{
int f64 = (int)MatrixElementType.F64;
if (left.ColumnLength != right.ColumnLength ||
@@ -88,10 +84,7 @@ namespace StgSharp.Mathematics.Numeric
return;
}
public static unsafe void Fill(
Matrix<double> ans,
double value
)
public static unsafe void Fill(Matrix<double> ans, double value)
{
int f64 = (int)MatrixElementType.F64;
long count = (long)ans.KernelColumnLength * ans.KernelRowLength;
@@ -122,11 +115,7 @@ namespace StgSharp.Mathematics.Numeric
taskGroup.Dispose();
}
public static unsafe void Mul(
Matrix<double> left,
Matrix<double> right,
Matrix<double> ans
)
public static unsafe void Mul(Matrix<double> left, Matrix<double> right, Matrix<double> ans)
{
int f64 = (int)MatrixElementType.F64;
@@ -139,25 +128,42 @@ namespace StgSharp.Mathematics.Numeric
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
int size = MatrixPanel.KernelSideCount<double>();
// int size = MatrixPanel.KernelSideCount<double>();
if (count <= 1024)
{
MatrixParallelThread[] pool = null;
if (count > 256) {
pool = MatrixParallel.LeadParallel(1);
}
for (int i = 0; i < ans.KernelRowLength; i += size)
M128* enumeration = stackalloc M128[1];
for (int i = 0; i < ans.KernelRowLength; i++)
{
for (int j = 0; j < ans.KernelColumnLength; j += size)
for (int j = 0; j < ans.KernelColumnLength; j += 1)
{
enumeration->Member<int>(0) = i;
enumeration->Member<int>(1) = j;
enumeration->Member<int>(2) = ans.KernelRowLength;
enumeration->Member<int>(3) = left.KernelColumnLength;
NativeIntrinsic.Context
.mat[f64].kernel_tile_fma((MatrixKernel*)left.Buffer,
(MatrixKernel*)right.Buffer,
(MatrixKernel*)ans.Buffer, enumeration);
/*
// Console.WriteLine($"{i}{j}");
for (int k = 0; k < left.KernelColumnLength; k += size)
for (int k = 0; k < left.KernelColumnLength; k++)
{
NativeIntrinsic.Context
.mat[f64].kernel_fma((MatrixPanel*)left.GetKernelAddressUnsafe(i, k),
(MatrixPanel*)right.GetKernelAddressUnsafe(k, j),
(MatrixPanel*)ans.GetKernelAddressUnsafe(i, j));
.mat[f64].kernel_fma(GetKernelAddressUnsafe(left.Buffer,
left.KernelColumnLength, i,
k),
GetKernelAddressUnsafe(right.Buffer,
right.KernelColumnLength, k, j),
GetKernelAddressUnsafe(ans.Buffer,
ans.KernelColumnLength, i, j));
}
/**/
}
}
if (count > 256) {
@@ -172,8 +178,8 @@ namespace StgSharp.Mathematics.Numeric
package->AnsPrimOffset = ans.KernelRowLength;
package->AnsSecOffset = ans.KernelColumnLength;
package->LeftPrimOffset = left.KernelColumnLength;
package->PrimCount = ans.KernelColumnLength / size;
package->SecCount = ans.KernelRowLength / size;
package->PrimCount = ans.KernelColumnLength;
package->SecCount = ans.KernelRowLength;
package->ElementSize = (int)MatrixElementType.F64;
package->ComputeMode = new MatrixOpMode(MatrixIndexStyle.BUFFER_INDEX, MatrixOperationSpecial.SEQ_FMA);
MatrixParallelHandle taskGroup = MatrixParallel.ScheduleTask(package);
@@ -183,18 +189,13 @@ namespace StgSharp.Mathematics.Numeric
return;
}
public static unsafe void Mul(
Matrix<double> left,
double right,
Matrix<double> ans
)
public static unsafe void Mul(Matrix<double> left, double right, Matrix<double> ans)
{
int f64 = (int)MatrixElementType.F64;
if (left.ColumnLength != ans.ColumnLength || left.RowLength != ans.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
if (left.Layout != MatrixLayout.DenseRectangle ||
ans.Layout != MatrixLayout.DenseRectangle) {
if (left.Layout != MatrixLayout.DenseRectangle || ans.Layout != MatrixLayout.DenseRectangle) {
throw new NotSupportedException("Only dense rectangle matrices are supported.");
}
long count = (long)left.KernelColumnLength * left.KernelRowLength;
@@ -235,11 +236,7 @@ namespace StgSharp.Mathematics.Numeric
return;
}
public static unsafe void Sub(
Matrix<double> left,
Matrix<double> right,
Matrix<double> ans
)
public static unsafe void Sub(Matrix<double> left, Matrix<double> right, Matrix<double> ans)
{
int f64 = (int)MatrixElementType.F64;
if (left.ColumnLength != right.ColumnLength ||