This commit is contained in:
Youmu Liu
2025-10-28 20:44:38 +08:00
parent e1ec761e83
commit daf9076e54
3 changed files with 49 additions and 18 deletions
-6
View File
@@ -25,15 +25,9 @@
//
// -----------------------------------------------------------------------
// -----------------------------------------------------------------------
using System;
using System.Collections.Generic;
using System.Drawing;
using System.Linq;
using System.Numerics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Text;
using System.Threading.Tasks;
namespace StgSharp.HighPerformance
{
+2 -2
View File
@@ -167,7 +167,7 @@ namespace StgSharp.Mathematics.Numeric
{
private readonly MatrixKernel<T>* _source = enumeration._source;
private int _indexCount = 0;
private int _indexCount = -1;
private readonly int _stride = enumeration.PrimStride;
private readonly long _bound = (long)enumeration.PrimSize * enumeration.PrimStride;
private long _index = (long)enumeration.PrimOffset * enumeration.PrimStride;
@@ -230,7 +230,7 @@ namespace StgSharp.Mathematics.Numeric
private readonly int _end = segment._size;
private int _indexCount = 0;
private int _indexCount = -1;
private readonly long _bound = (long)segment._size * segment._stride;
private long _index = (long)segment._offset * segment._stride;
private readonly MatrixSegment<T> _segment = segment;
@@ -45,24 +45,61 @@ namespace StgSharp.Mathematics.Numeric
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
Matrix<float> ans = Matrix<float>.FromDefault(left.ColumnLength, left.RowLength);
if (left.KernelColumnLength * left.KernelRowLength <= 1024)
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 1024)
{
MatrixParallelPool pool = MatrixParallel.LeadParallel();
if (!pool.TryRentIndividualHandle(out MatrixParallelWrap.Handle? handle)) { }
MatrixSegmentEnumeration<float>.Enumerator leftEnum = left.GetColumnEnumeration().GetEnumerator();
MatrixSegmentEnumeration<float>.Enumerator rightEnum = right.GetColumnEnumeration().GetEnumerator();
MatrixSegmentEnumeration<float>.Enumerator ansEnum = ans.GetColumnEnumeration().GetEnumerator();
for (int i = 0; i < left.KernelColumnLength; i++)
if (!pool.TryRentIndividualHandle(out MatrixParallelWrap.Handle? handle))
{
MatrixSegment<float>.Enumerator leftColumEnum = leftEnum.Current.GetEnumerator();
MatrixSegment<float>.Enumerator rightColumEnum = rightEnum.Current.GetEnumerator();
MatrixSegment<float>.Enumerator ansColumEnum = ansEnum.Current.GetEnumerator();
/*
* wait
*/
}
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
for (int i = 0; i < count; i++) {
NativeIntrinsic.Intrinsic.f32_add(leftPtr + i, rightPtr + i, ansPtr + i);
}
return ans;
}
Stack<MatrixParallelWrap> taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
(delegate*<MatrixKernel<float>*, MatrixKernel<float>*, MatrixKernel<float>*, void>)NativeIntrinsic.Intrinsic.f32_add);
/*
* run parallel
*/
return ans;
}
public static unsafe Matrix<float> Sub(Matrix<float> left, Matrix<float> right)
{
if (left.ColumnLength != right.ColumnLength || left.RowLength != right.RowLength) {
throw new IndexOutOfRangeException("Matrix dimensions must agree.");
}
Matrix<float> ans = Matrix<float>.FromDefault(left.ColumnLength, left.RowLength);
long count = (long)left.KernelColumnLength * left.KernelRowLength;
if (count <= 1024)
{
MatrixParallelPool pool = MatrixParallel.LeadParallel();
if (!pool.TryRentIndividualHandle(out MatrixParallelWrap.Handle? handle))
{
/*
* wait
*/
}
MatrixKernel<float>* leftPtr = left.Buffer;
MatrixKernel<float>* rightPtr = right.Buffer;
MatrixKernel<float>* ansPtr = ans.Buffer;
for (int i = 0; i < count; i++) {
NativeIntrinsic.Intrinsic.f32_sub(leftPtr + i, rightPtr + i, ansPtr + i);
}
return ans;
}
Stack<MatrixParallelWrap> taskGroup = MatrixParallel.PublicTask(left.GetColumnEnumeration(), right.GetColumnEnumeration(), ans.GetColumnEnumeration(),
(delegate*<MatrixKernel<float>*, MatrixKernel<float>*, MatrixKernel<float>*, void>)NativeIntrinsic.Intrinsic.f32_sub);
/*
* run parallel
*/
return ans;
}